texonom/texonom-md · Datasets at Hugging Face

title stringlengths 1 544 ⌀	parent stringlengths 0 57 ⌀	created stringlengths 11 12 ⌀	editor stringclasses 1 value	creator stringclasses 4 values	edited stringlengths 11 12 ⌀	refs stringlengths 0 536 ⌀	text stringlengths 1 26k	id stringlengths 32 32
PPO	Reinforcement Learnings	Jul 15, 2023	Alan Jo	Alan Jo	Jul 15, 2023		### Proximal Policy Optimization balance between ease of implementation, sample complexity, and tuning 2017년 OpenAI에서 개발된 모델 없는 강화 학습 알고리즘 태-동작 쌍에 값을 할당하는 대신 정책 공간을 검색 data를 만들어낸 old policy와 train 대상이 되는 현재의 network의 new policy를 구분 [PPO2](https://texonom.com/ppo2-378944bb3b8848a185c20fc529c390d2) > [Proximal Policy Optimization Algorithms](https://arxiv.org/abs/1707.06347) > [Proximal Policy Optimization](https://openai.com/research/openai-baselines-ppo)	87ce8ebe81f84ac1a7617b1f6def9e26
Q-Learning	Reinforcement Learnings	Nov 5, 2019	Alan Jo	Seong-lae Cho	Aug 31, 2023	[SARSA](https://texonom.com/sarsa-e5d847a4fb6e41cdad5b9ffb6e974a10)	### Approximate Q-Learning Off police 시간차제어 현재 행동하는 정책과는 독립적으로 학습 ### Reinforce > [강화학습 기초(Q-learning)](https://bluediary8.tistory.com/18)	6fb81e4a53ab4e3097784cde99c8c038
RLHF	Reinforcement Learnings	Apr 30, 2023	Alan Jo	Alan Jo	Jul 15, 2023	[AI Alignment](https://texonom.com/ai-alignment-f676f1a29ffd45e19b3d170afa4f2244) [Active Learning](https://texonom.com/active-learning-85d42ac892e84e5ba4fd2727f1791f65)	## Reinforcement learning from human feedback ### Limitation LM의 근본적인 문제인 Size, hallucination을 아직까지는 개선할 수는 없는 한계점 Scaling 이슈, 너무 복잡 > [RLHF란?](https://velog.io/@nellcome/RLHF란) > [Reinforcement learning from human feedback](https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback)	4b184f9c9e8b4c7a8861fb6374e91aa6
RRHF	Reinforcement Learnings	Jul 15, 2023	Alan Jo	Alan Jo	Jul 15, 2023	[RLHF](https://texonom.com/rlhf-4b184f9c9e8b4c7a8861fb6374e91aa6)	### *Rank Response to align Human F*eedback efficiently align language model output probabilities with human preferences as robust as fine-tuning and it only needs 1 to 2 models during tuning	073c6bdf48444a1b9acad9e65057f3d0
SARSA	Reinforcement Learnings	Jul 18, 2023	Alan Jo	Alan Jo	Aug 31, 2023		## state-action-reward-state-action 현재 가지고 있는 큐함수 를 토대로 샘플을 탐욕 정책으로 모으고, 그 샘플로 방문한 큐함수를 업데이트하는 과정을 반복 Almost start of Reinforcement learning after [Policy Iteration](https://texonom.com/policy-iteration-55cbe9219f8a48fbb850f64b677e847a) GPI에서는 벨만 방정식에 따라 정책을 평가 Temporal-Difference 방법에서는 가치 이터레이션의 방법을 도입 현재 상태의 큐함수를 보고 판단한다면 환경의 모델을 몰라도 된다 시간차 제어에서는 큐함수를 사용한 탐욕 정책 을 통해 행동을 선택 초기의 에이전트에게 탐욕정책은 잘못된 학습으로 가게할 가능성이 크기 때문에 epsilon-탐욕 정책 을 사용 ### Limitation 특정 state에 갇혀버리는 현상 자신이 행동한 대로 학습하는 것을 On-Policy 시간차 제어 그래서 [Q-Learning](https://texonom.com/q-learning-6fb81e4a53ab4e3097784cde99c8c038) > [(7) 살사(SARSA)와 큐러닝(Q-Learning)](https://jang-inspiration.com/sarsa-qlearning#132c2000aab74666a2273d8b2d71cdac)	e5d847a4fb6e41cdad5b9ffb6e974a10
TRPO	Reinforcement Learnings	Jul 15, 2023	Alan Jo	Alan Jo	Jul 15, 2023		### **Trust Region Policy Optimization**	23c0f917124d419cbfe478d20b804276
PPO2	PPO	null	null	null	null	null	> [Untitled](https://github.com/openai/baselines/blob/master/baselines/ppo2/ppo2.py)	378944bb3b8848a185c20fc529c390d2
ML Compiler Optimization	Machine Learning Techniques	Jul 9, 2022	Alan Jo	Alan Jo	Apr 25, 2023	[Compiler Optimization](https://texonom.com/compiler-optimization-3eee7067387b45d7875df50f4473ad18) [Parallel Training](https://texonom.com/parallel-training-4a8896bc837b4dddb47c3700b715cdc8)	[relax](https://github.com/mlc-ai/relax) ### ML Compiler Optimization Tools \|Title\| \|:-:\| \|[XLA](https://texonom.com/xla-4f75dff43dc0451aa5ca92e9218a3028)\| \|[MLGO](https://texonom.com/mlgo-60d88727f0b944ec9ce62d254c2e8a76)\| \|[Hidet](https://texonom.com/hidet-7b706024cb3040269f71e47b5b87d2b2)\|	011e7bd0ba8f417bb111ec5ea2171c8e
Parallel Training	Machine Learning Techniques	Mar 15, 2022	Alan Jo	Alan Jo	Apr 25, 2023	[ML Compiler Optimization](https://texonom.com/ml-compiler-optimization-011e7bd0ba8f417bb111ec5ea2171c8e)	### data parallelism or model parallelism - In data parallelism, the data is split into multiple parts - in model parallelism, different parts of the model are processed by separate processors ### Parallel Training Notion \|Title\| \|:-:\| \|[Model Parallelism](https://texonom.com/model-parallelism-76dd813ada7b4e50b645af2f05821d48)\| \|[Data Parallelism](https://texonom.com/data-parallelism-8e90f1c595a84dad9f4e921e74f86ba6)\| ### Parallel Training Usages \|Title\| \|:-:\| \|[Parallel Learning Tool](https://texonom.com/parallel-learning-tool-2a9741aa76c14c16a1240f3422f11421)\| \|[Parallel Training Example](https://texonom.com/parallel-training-example-3c16ddf97fbe43359e7da3dbd3ce96ee)\| ![https://xiandong79.github.io](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F74479c70-d665-4f9e-a0df-dda36042a7ee%2FUntitled.png?table=block&id=cf76f928-10e0-4fb6-bb03-ec8afb321aa7&cache=v2) > [딥러닝 모델의 분산학습이란? (Data parallelism과 Model parallelism)](https://lifeisenjoyable.tistory.com/21)	4a8896bc837b4dddb47c3700b715cdc8
Quantum Machine Learning	Machine Learning Techniques	Mar 9, 2022	Alan Jo	Alan Jo	Mar 5, 2023		### Quantum Machine Learnings \|Title\| \|:-:\| > [Spooky Action Could Help Boost Quantum Machine Learning](https://spectrum.ieee.org/quantum-machine-learning)	8f5276045f4d43b8b96f3b4ec6646f66
Weight Initialization	Machine Learning Techniques	Jun 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023		- weight to small random numbers - bias (zero or small nonzero) ### Weight Initialization Usages \|Title\| \|:-:\| \|[He initialization](https://texonom.com/he-initialization-a255fdaeec8e485faf3215a28ed5fdb9)\| \|[Xavier Initialization](https://texonom.com/xavier-initialization-40045bfdf72343aea3a234214145f9dd)\| > [0025 Initialization - Deepest Documentation](https://deepestdocs.readthedocs.io/en/latest/002_deep_learning_part_1/0025/)	6cfc10eb06f948528aa76a9814a9ac85
Hidet	ML Compiler Optimization Tools	May 1, 2023	Alan Jo	Alan Jo	May 1, 2023	[Pytorch](https://texonom.com/pytorch-2dd232d99b3a46d5b7d1e4e686070686)	> [PyTorch](https://pytorch.org/blog/introducing-hidet)	7b706024cb3040269f71e47b5b87d2b2
MLGO	ML Compiler Optimization Tools	Jul 9, 2022	Alan Jo	Alan Jo	Mar 11, 2023	[LLVM](https://texonom.com/llvm-5dc6acb10b5244a2af349319ef87c797) [ml-compiler-opt](https://github.com/google/ml-compiler-opt)	### Infrastructure for Machine Learning Guided Optimization > [MLGO: A Machine Learning Framework for Compiler Optimization](https://ai.googleblog.com/2022/07/mlgo-machine-learning-framework-for.html)	60d88727f0b944ec9ce62d254c2e8a76
XLA	ML Compiler Optimization Tools	Mar 11, 2023	Alan Jo	Alan Jo	Mar 11, 2023	[xla](https://github.com/openxla/xla)	- pytorch - tensorflow - jax	4f75dff43dc0451aa5ca92e9218a3028
Data Parallelism	Parallel Training Notion	Apr 25, 2023	Alan Jo	Alan Jo	Apr 25, 2023		학습 데이터를 여러 GPU에 나눠 학습 ### Data Parallelism Usages \|Title\| \|:-:\|	8e90f1c595a84dad9f4e921e74f86ba6
Model Parallelism	Parallel Training Notion	Apr 25, 2023	Alan Jo	Alan Jo	Apr 25, 2023		모델을 여러 GPU에 나누는 ### Model Parallelism Usages \|Title\| \|:-:\|	76dd813ada7b4e50b645af2f05821d48
Parallel Learning Tool	Parallel Training Usages	Apr 25, 2023	Alan Jo	Alan Jo	Apr 25, 2023		### Parallel Training System \|Title\| \|:-:\| \|[Colossal AI](https://texonom.com/colossal-ai-031c5480b24249ce903fea4e0f8d435c)\| \|[Megatrom LM](https://texonom.com/megatrom-lm-87be23e9b623465395a0d6a4e94470ae)\| \|[DeepSpeed](https://texonom.com/deepspeed-3866b23c00eb4d529de6e33dc48ffae7)\|	2a9741aa76c14c16a1240f3422f11421
Parallel Training Example	Parallel Training Usages	Apr 25, 2023	Alan Jo	Alan Jo	Apr 25, 2023		### Parallel Learning Examples \|Title\| \|:-:\| \|[](https://texonom.com/48f191fcffa04c068978381a78b4ca8d)\|	3c16ddf97fbe43359e7da3dbd3ce96ee
Colossal AI	Parallel Training System	Mar 15, 2022	Alan Jo	Alan Jo	Apr 25, 2023	[ColossalAI](https://github.com/hpcaitech/ColossalAI)	> [Colossal-AI](https://colossalai.org/)	031c5480b24249ce903fea4e0f8d435c
DeepSpeed	Parallel Training System	Feb 19, 2021	Alan Jo	Alan Jo	Apr 25, 2023	[DeepSpeed](https://github.com/microsoft/DeepSpeed)	### Pipeline Parallelism > [DeepSpeed Pipeline Parallelism](https://velog.io/@nawnoes/DeepSpeed-Pipeline-Parallelism) > [PyTorch Lightning DeepSpeed](https://velog.io/@nawnoes/PyTorch-Lightning-DeepSpeed)	3866b23c00eb4d529de6e33dc48ffae7
Megatrom LM	Parallel Training System	Apr 25, 2023	Alan Jo	Alan Jo	Apr 25, 2023	[Megatron-LM](https://github.com/NVIDIA/Megatron-LM)		87be23e9b623465395a0d6a4e94470ae
[gpt-neox](https://github.com/EleutherAI/gpt-neox)	Parallel Learning Examples	Apr 25, 2023	Alan Jo	Alan Jo	Apr 25, 2023	[Megatrom LM](https://texonom.com/megatrom-lm-87be23e9b623465395a0d6a4e94470ae) [DeepSpeed](https://texonom.com/deepspeed-3866b23c00eb4d529de6e33dc48ffae7)		48f191fcffa04c068978381a78b4ca8d
He initialization	Weight Initialization Usages	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023	[ReLU](https://texonom.com/relu-e582549804da48b893758895e446ffb9)	ReLU + He 초기화 방법이 보편적 > [Untitled](https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/He_Delving_Deep_into_ICCV_2015_paper.pdf)	a255fdaeec8e485faf3215a28ed5fdb9
Xavier Initialization	Weight Initialization Usages	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023		## Glorot Initialization 이전 층의 뉴런 개수와 다음 층의 뉴런 개수 이용 여러 층의 기울기 분산 사이에 균형을 맞춘다 S자 형태인 활성화 함수와 함께 사용할 경우에는 좋은 성능을 보이지만, ReLU와 함께 사용할 경우에는 성능이 좋지 않다 ### Uniform Distribution ### Normal distribution > [07-07 기울기 소실(Gradient Vanishing)과 폭주(Exploding)](https://wikidocs.net/61375)	40045bfdf72343aea3a234214145f9dd
Function Transformers	Machine Learning Tools	Jun 13, 2021	Alan Jo	Alan Jo	Jan 9, 2023		### Composable transformations > [Transformers](https://huggingface.co/transformers/)	c4396d81a01a4425ba0c6702501c911a
ML Accelerator	Machine Learning Tools	Jun 1, 2022	Alan Jo	Alan Jo	Jun 1, 2022		### ML Accelerators \|Title\| \|:-:\|	0df3a271237b4e63a342aa7ce704870d
ML Analyze Tool	Machine Learning Tools	Aug 5, 2021	Alan Jo	Alan Jo	Apr 19, 2022		### ML Analyze Tools \|Title\| \|:-:\| \|[Uptrain](https://texonom.com/uptrain-a12b6a1c3410434ea318cb76a2f99a98)\| \|[Evidently](https://texonom.com/evidently-0500b048a7c34d8e91e18207e43cefd1)\|	5aeb48a9850245eb97e20ec56448a15f
ML Container Tool	Machine Learning Tools	May 11, 2022	Alan Jo	Alan Jo	May 11, 2022		### ML Container Tools \|Title\| \|:-:\| \|[Cog](https://texonom.com/cog-1b9175a7b63c43b5b6cd69a223f9d99d)\|	14223d2021c748268fba092dee1fa357
ML Feature Store	Machine Learning Tools	Apr 19, 2022	Alan Jo	Alan Jo	Apr 19, 2022		### ML Feature Stores \|Title\| \|:-:\| \|[Feathr](https://texonom.com/feathr-5ff843203b4745df949b956391bf9423)\|	f53069a7083b4719b1e5fab18a5a9bbd
ML Platform	Machine Learning Tools	Sep 8, 2021	Alan Jo	Alan Jo	Aug 4, 2022	[](https://texonom.com/fee31eaaf53a45c28b0b305c3874b856)	### ML Platforms \|Title\| \|:-:\| \|[diffgram](https://texonom.com/diffgram-99e0ffaf73f84615a556b6bbe71c4572)\| \|[Wandb](https://texonom.com/wandb-2219d228212940068aa5a604af7d5dbc)\|	9d4142db8db042ed9e4a79085348cc55
Evidently	ML Analyze Tools	Aug 5, 2021	null	null	null	null	> [GitHub - evidentlyai/evidently: Interactive reports to analyze machine learning models during validation or production monitoring.](https://github.com/evidentlyai/evidently?ref=producthunt?utm_source=tldrnewsletter)	0500b048a7c34d8e91e18207e43cefd1
Uptrain	ML Analyze Tools	Mar 9, 2023	null	null	null	null	[uptrain](https://github.com/uptrain-ai/uptrain)	a12b6a1c3410434ea318cb76a2f99a98
Cog	ML Container Tools	May 11, 2022	Alan Jo	Alan Jo	May 11, 2022	[cog](https://github.com/replicate/cog)		1b9175a7b63c43b5b6cd69a223f9d99d
Feathr	ML Feature Stores	Apr 19, 2022	Alan Jo	Alan Jo	Apr 19, 2022	[LinkedIn](https://texonom.com/linkedin-1c0eb8ae1ca346a388e79c15b34355dc) [feathr](https://github.com/linkedin/feathr)	> [Open sourcing Feathr - LinkedIn's feature store for productive machine learning](https://engineering.linkedin.com/blog/2022/open-sourcing-feathr linkedin-s-feature-store-for-productive-m) ### Template Gallery \|Title\| \|:-:\| \|[Template Page](https://texonom.com/template-page-b6dd128730be402fbf47e98d1a81c5f2)\|	5ff843203b4745df949b956391bf9423
Template Page	Template Gallery	Apr 19, 2022	Alan Jo	Alan Jo	Apr 19, 2022			b6dd128730be402fbf47e98d1a81c5f2
diffgram	ML Platforms	Sep 8, 2021	null	null	null		> [GitHub - diffgram/diffgram: Complete training data platform for machine learning delivered as a single application.](https://github.com/diffgram/diffgram)	99e0ffaf73f84615a556b6bbe71c4572
Wandb	ML Platforms	Aug 4, 2022	null	null	null		[wandb](https://github.com/wandb/wandb)	2219d228212940068aa5a604af7d5dbc
AdaBoost	ML Meta Algorithms	Oct 6, 2021	Alan Jo	Alan Jo	Oct 6, 2021		성능을 향상시키기 위하여 다른 많은 형태의 학습 알고리즘과 결합하여 사용	79feac6204384e7299e08e3cfa40d05e
Deep Learning	Neural Network Notion	Nov 5, 2019	Alan Jo	Seong-lae Cho	Jul 5, 2023	[Neuroscience ](https://texonom.com/neuroscience-b45d9f638a2b4330906556c402307925)	### Neural Network based Machine Learning method composition of differentiable functions Brain Algorithm + [Neural Network](https://texonom.com/neural-network-86f54f9f1de848c1a29c56c24f7d5094) + [Big Data](https://texonom.com/big-data-236ec9f0ed844f4d8a5ca3236dfa442c) limit - explainability, fairness, generalizability, causality Main feature of neural network It can find non-heuristic feature representation ### Deep Learning Notion \|Title\| \|:-:\| \|[Deep Learning Math](https://texonom.com/deep-learning-math-57c204edef3042568bbd0d5268b877fd)\| \|[Deep Learning Network](https://texonom.com/deep-learning-network-f368faf8fa634699aeda503d01c193f0)\| \|[End-to-end Deep Learning](https://texonom.com/end-to-end-deep-learning-16bd62f447a144f1b251acf25f1f8789)\| ### Deep Learning Usages \|Title\| \|:-:\| \|[Deep Learning Tool](https://texonom.com/deep-learning-tool-a14ea6f4574342ef974443634e27c6ce)\| \|[Deep Learning Compiler](https://texonom.com/deep-learning-compiler-7d79af9683764b6b983793c1856578c6)\| \|[Sentiment Neuron](https://texonom.com/sentiment-neuron-2e44e9754b894534af6c121b6d6074d6)\| \|[Learn Deep Learning](https://texonom.com/learn-deep-learning-4382083fb54d4705984e7f45a9af2d86)\| ### Interview from popular people > [제프리 힌턴이 말하는 AI의 영향력과 잠재력](https://www.youtube.com/watch?v=IvUw9um4Bv8) > [OpenAI의 핵심, Ilya Sutskever 인터뷰](https://www.youtube.com/watch?v=SGCFeIbpGlU&t=722s)	7d3c8b9ce05b49cf9eed92dbcdc80cfd
Neural Network History	Neural Network Notion	May 11, 2023	Alan Jo	Alan Jo	Jun 7, 2023		> [[미라클레터] AI시대 필수품 아날로그 컴퓨터!!](https://stibee.com/api/v1.0/emails/share/KeeaUnUVO5o8muoWpT9bjEtpsL5hny0=)	2127bdb56da54a659e570f47704e40b1
Neural Network Structure	Neural Network Notion	May 11, 2023	Alan Jo	Alan Jo	Jul 4, 2023		### Neural Network Components \|Title\| \|:-:\| \|[Activation Function](https://texonom.com/activation-function-8e52ee5f83a244d88abeeee3fb9497a8)\| \|[Neural Network Layer](https://texonom.com/neural-network-layer-e10ef2afe1954cf6b909f8aa40077393)\| \|[Forward Forward Algorithm](https://texonom.com/forward-forward-algorithm-6c989313e382466bba02d15c412fb17f)\| \|[skip connections](https://texonom.com/skip-connections-d7ad187f3487468db6eea278f0236a22)\| ### Neural Networks \|Title\| \|:-:\| \|[Perceptron](https://texonom.com/perceptron-1deb66f486d54c93bb928d8afba1864c)\| \|[FFNN](https://texonom.com/ffnn-89ecee87d8b7482e86995950db90eb31)\| \|[CNN](https://texonom.com/cnn-002bf81a77bc40d1858740d26b61d97b)\| \|[RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16)\| \|[ANN](https://texonom.com/ann-d4232205ecf9463c95a911d179c87a84)\| \|[SNN](https://texonom.com/snn-6cf69239a87e4df9a32b3494862374e4)\| \|[GNN](https://texonom.com/gnn-58adb81cb2b649d9af19019182960bb2)\|	400bbea8029c4eb1a97c0dd063735551
Deep Learning Math	Deep Learning Notion	Nov 5, 2019	Alan Jo	Seong-lae Cho	Mar 26, 2023		### Sub-field of ML: learning representations of data Existing ML uses manually designed features - often over-specified and incomplete - take a long time to design and validate DP - Learned Features are easy to adapt, fast - Deep learning provides a very flexible, (almost?) universal - Effective end-to-end joint system learning > speach recog not good > visual perception good > Question Answering good > 4 + 2 = 6 neurons (not counting inputs) - biases has same number (for each resulting node) > [3 x 4] + [4 x 2] = 20 weights > > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F5986fe3d-198a-4608-959f-6bd5bc738065%2FUntitled.png?table=block&id=8f6601d1-344a-4488-b982-926c53650f7f&cache=v2) > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F4110b2cb-bf0f-4221-b2c9-cc8c6f2c8786%2FUntitled.png?table=block&id=0e436d3f-24b4-4328-bc11-1e91e5aa5747&cache=v2) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Ff7503f3f-0f31-446e-b91b-32842c7f119d%2FUntitled.png?table=block&id=4e369bcf-c0fd-4d82-9523-8156a1c29aac&cache=v2) Optimize (min. or max.) objective/cost function 𝐽(𝜃) Generate error signal that measures difference between predictions and target values word representation is 74 72 65 65 in x0 but it is pool so we need new presentation of word ## wordnet WordNet: contains the list of synonyms/hypernyms ← using human resources Word’s meaning is given by words that frequently appear close-by ### Context: set of words that appears nearby in the fixed-size window (ex: before/after 5 words) Vector dimension = Number of words * Number of words memroy O(n^2) poor so we use dimensionality reduction (pca, svd) ← almost all value is 0 so l;we can → word embedding 이제 벡터곱으로 similar판단가능 (become dense vector) - from one hot vector > We can make embedding for between other languagex > vector similarity (cosine) > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F30c1772d-b6df-4eed-ad3f-2777aebcef9a%2FUntitled.png?table=block&id=e97e701f-21dd-4ab7-8e3e-d039ad4d6d08&cache=v2) # Language Modeling: Models ofP(text) - sentance score sentance > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F3c48b3c2-5971-4a4d-8c78-896f6b721be1%2FUntitled.png?table=block&id=70af901f-dd8c-4e16-a2b5-408adf172390&cache=v2) > add all is exampel (linearfeatures) > usually softmax the sum > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F66fd7479-7a77-416a-9684-a30972d659e4%2FUntitled.png?table=block&id=4f96ddb9-9fa4-48df-b29f-2012e73945f1&cache=v2) > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F7bcd216a-04a9-411b-9c95-c2c8728c44e7%2FUntitled.png?table=block&id=2b6a582b-5666-47cd-8fa5-0185c4b045bb&cache=v2) soft max is for before output ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F52fd4795-45f9-4e20-8f4a-b9a18df4d4d7%2FUntitled.png?table=block&id=5a5b7c77-f67f-486b-b148-0265844a4461&cache=v2) # 1. P(text) - linear model ## CBOW Predict word based on sum of surroundingembeddings ## Skip-gram Predict each word in the context given theword [very good, good, neutral, bad, verybad] Linear Models can’tLearn FeatureCombinations # 2. Models of P(label \|text) ## BOW each word has its own 5 elements corresponding to ## DeepCBOW CombinationFeatures - Each vector has “features” (e.g. is this an animate object? is this a positive word, etc.) # ConvolutionalNetworks ### x/ CNN - pooling weak for long distance feature extractor don't have holistic view of # x/ RNN good for long disatnce fraeture extractor weakness - Indirect passing of information, credit assignment moredifficult Can be slow, due to incrementalprocessing ## ModelingSentences w/n-grams # P(text \|text) > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F547f55b2-840c-4e50-80a0-8b936f4a64c3%2FUntitled.png?table=block&id=7fa3ac7f-fd55-4575-b6b1-afbae72438fe&cache=v2) > Conditional LanguageModels > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fba8eaa2d-8c36-4f5b-b222-e60b091aa063%2FUntitled.png?table=block&id=6f689ac1-dfc7-45ac-9f72-5704d81f6dcf&cache=v2) > Calculating the Probability of aSentence > RNN is frequently used in language modeling since RNN can capture long-distance dependencies	57c204edef3042568bbd0d5268b877fd
Deep Learning Network	Deep Learning Notion	Oct 6, 2021	Alan Jo	Alan Jo	May 29, 2023		### Deep Learning Models \|Title\| \|:-:\| \|[Seq2Seq](https://texonom.com/seq2seq-01a9854dffa6417c87d92c11a607250c)\| \|[GAN](https://texonom.com/gan-66482b5f518d47f6b337eba9a30ff792)\| \|[Capsule Network](https://texonom.com/capsule-network-a14fc7e569864154aae1ef44106e8991)\| \|[MANN](https://texonom.com/mann-84ed691391c8439ba8e6b297623c9c0e)\|	f368faf8fa634699aeda503d01c193f0
End-to-end Deep Learning	Deep Learning Notion	Apr 3, 2023	Alan Jo	Alan Jo	Apr 3, 2023		입력에서 바로 출력을 구할 수 있다 모든 매개변수가 하나의 손실함수에 대해 동시에 훈련되는 경로가 가능한 네트워크를 뜻한다 [Text Tokenizer](https://texonom.com/text-tokenizer-2bbc41eaa76c4674a4f4b9127fbe5da1) [Text Encoding](https://texonom.com/text-encoding-ab7377cfc5c648059de4860510ad9134) 하지않고 메모리 많이 사용 > [What is end-to-end deep learning?](https://velog.io/@jeewoo1025/What-is-end-to-end-deep-learning)	16bd62f447a144f1b251acf25f1f8789
Capsule Network	Deep Learning Models	Aug 21, 2021	Alan Jo	Alan Jo	Oct 6, 2021	[CNN](https://texonom.com/cnn-002bf81a77bc40d1858740d26b61d97b)	## CapsNet cnn에서 이미지 recognition 에서 생기는 문제가 없는 인공신경망 detects the rotation and leans it as one of the activation vector ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fb1ec5a35-c647-46dd-90c3-07dcc5e94b7c%2FUntitled.png?table=block&id=9eb6f960-e681-498c-8c63-3b6096168011&cache=v2) > [What is a CapsNet or Capsule Network?](https://medium.com/hackernoon/what-is-a-capsnet-or-capsule-network-2bfbe48769cc) > [Why Do Capsule Networks Work Better Than Convolutional Neural Networks?](https://medium.com/@ashukumar27/why-do-capsule-networks-work-better-than-convolutional-neural-networks-f4a105a53aff)	a14fc7e569864154aae1ef44106e8991
GAN	Deep Learning Models	Nov 18, 2019	Alan Jo	Alan Jo	Jun 1, 2023	[Unsupervised learning](https://texonom.com/unsupervised-learning-8cb6e253bfa845b5931d22963ea93019) [Generative Model](https://texonom.com/generative-model-6e5204d2982b4042847aa42e88eb8fb5) [Transfer Learning](https://texonom.com/transfer-learning-442feb66465944eebf144d4e9dd1dbf8)	## Generative Adversarial Network Learn how to generate samples, 위조지폐와 경찰같이 경쟁하며 발전 ### GAN Notion \|Title\| \|:-:\| \|[Generator Network](https://texonom.com/generator-network-3d3fb237d8f149979c7ed172aca65529)\| \|[Discriminator Network](https://texonom.com/discriminator-network-a02287522a264d779574871285eed3b6)\| \|[GAN Minmax Game](https://texonom.com/gan-minmax-game-d3b08fa32fc34cf5920bfc0e80c34b90)\| \|[GAN Issues](https://texonom.com/gan-issues-be6312c3a8184b55ba443167164101ba)\| ### GANs \|Title\| \|:-:\| \|[3D GAN](https://texonom.com/3d-gan-672af0e2639041e484ed1a26b56f84cb)\| \|[DCGAN](https://texonom.com/dcgan-ae93b511fed5402c926e818243d8a966)\| \|[DragGan](https://texonom.com/draggan-bd335de6767243e7b7ec48ff88dea800)\| > [Generative adversarial network](https://en.wikipedia.org/wiki/Generative_adversarial_network)	66482b5f518d47f6b337eba9a30ff792
**MANN**	Deep Learning Models	Apr 29, 2023	Alan Jo	Alan Jo	May 29, 2023		### Memory Augmented Neural Networks RNN, CNN 등의 기반 모델과 메모리 구조를 결합한 모델 ### **MANN Notion** \|Title\| \|:-:\| \|[Differentiable Neural Computer](https://texonom.com/differentiable-neural-computer-2c182410c5c34222b41605b63c37c777)\| \|[Neural Turing Machine](https://texonom.com/neural-turing-machine-c03efee39e1942f197f4b3d6553e4ac1)\|	84ed691391c8439ba8e6b297623c9c0e
Seq2Seq	Deep Learning Models	Mar 4, 2023	Alan Jo	Alan Jo	Jul 30, 2023		### Variable Length of inputs and outputs 인코더-디코더 구조는 주로 입력 문장과 출력 문장의 길이가 다를 경우에 사용 - Encoder takes the input sequence and converts it into a fixed length vector representation - Decoder uses this vector to generate the output sequence ### Seq2Seq Notion \|Title\| \|:-:\| \|[Attention Mechanism](https://texonom.com/attention-mechanism-762711860abb45f59904f1ac4e4af285)\| \|[Copy mechanism](https://texonom.com/copy-mechanism-b5dee9c80ca24bb993b5f152129b3577)\| ### Seq2Seq Models \|Title\| \|:-:\| \|[Decoder Model](https://texonom.com/decoder-model-36e78a40265c473d90197089aebfa83b)\| \|[Transformer Model](https://texonom.com/transformer-model-f3e8053cc5b447a2bc9c6b5d0874dafc)\| \|[Encoder Model](https://texonom.com/encoder-model-321d79943c8940fcaac0c9ccca0b5f6f)\| > [14-01 시퀀스-투-시퀀스(Sequence-to-Sequence, seq2seq)](https://wikidocs.net/24996)	01a9854dffa6417c87d92c11a607250c
Discriminator Network	GAN Notion	Nov 18, 2019	Alan Jo	Alan Jo	Jun 1, 2023		tries to distinguish between real and fake images Discriminator $\phi$ aims at maximizing the objective - $D(x)$ to be close to 1 for real - $D(G(z))$ to be close to 0 for fake	a02287522a264d779574871285eed3b6
GAN Issues	GAN Notion	Jun 1, 2023	Alan Jo	Alan Jo	Jun 14, 2023		### 1. Non-convergence minmax objective cycle generated so repeated. Even with a small learning rate, it will not converge example. $min_xmax_yV(x, y ) = xy$ ### 2. Mode-Collapse What if the generator keeps generating a single realistic image? The discriminator will be always fooled by the single sample ### Mini Batch Trick Compute the similarity of the image $x$ with other images in the same batch to avoid Mode-Collapse. This measures the diversity of the batch. Feed the similarity score along with the image to the discriminator as an input feature. This penalizes the generator and encourages it to generate less similar images Many more advanced techniques have also been proposed so far.	be6312c3a8184b55ba443167164101ba
GAN Minmax Game	GAN Notion	Jun 1, 2023	Alan Jo	Alan Jo	Jun 1, 2023		## Training GAN Aims to $D(G(z))$ to be close to 1, i.e. discriminator is fooled ### Gradient descent on generator ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F19a41272-13ca-4ad9-b94a-c3fe8dbd031f%2FUntitled.png?table=block&id=17e51d57-7290-4e22-8703-8a64cf6e1b95&cache=v2) If $G$ is very bad compared to $D$, then we have almost zero gradient. Hence, $-D$ term can be used ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F5c0d0cce-916a-44e9-a971-c7201830e7db%2FUntitled.png?table=block&id=5841e261-a075-413d-860c-0186cb106361&cache=v2) ### Gradient ascent on discriminator Generator $\theta$ and Discriminator $\phi$ has very different objective so hard to train stable	d3b08fa32fc34cf5920bfc0e80c34b90
Generator Network	GAN Notion	Nov 18, 2019	Alan Jo	Alan Jo	Jun 1, 2023		tries to fool the discriminator by generating realistic sample Generator $\theta $ aims at minimizing the objective	3d3fb237d8f149979c7ed172aca65529
3D GAN	GANs	Feb 24, 2022	Alan Jo	Alan Jo	Jun 1, 2023		> [3D Generative Adversarial Network](http://3dgan.csail.mit.edu/)	672af0e2639041e484ed1a26b56f84cb
DCGAN	GANs	Mar 5, 2023	Alan Jo	Alan Jo	Jun 1, 2023		[DCGAN-tensorflow](https://github.com/carpedm20/DCGAN-tensorflow)	ae93b511fed5402c926e818243d8a966
DragGan	GANs	May 29, 2023	Alan Jo	Alan Jo	Jun 1, 2023	[DragGAN](https://github.com/XingangPan/DragGAN)	### Unofficial Huggingface implementation > [DragGAN - a Hugging Face Space by fffiloni](https://huggingface.co/spaces/fffiloni/DragGAN) > [Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold](https://vcai.mpi-inf.mpg.de/projects/DragGAN/)	bd335de6767243e7b7ec48ff88dea800
Differentiable Neural Computer	MANN Notion	May 29, 2023	Alan Jo	Alan Jo	May 29, 2023			2c182410c5c34222b41605b63c37c777
Neural Turing Machine	MANN Notion	May 29, 2023	Alan Jo	Alan Jo	May 29, 2023			c03efee39e1942f197f4b3d6553e4ac1
Decoder Model	Seq2Seq Models	Mar 6, 2023	Alan Jo	Alan Jo	Jul 17, 2023	[Encoder Model](https://texonom.com/encoder-model-321d79943c8940fcaac0c9ccca0b5f6f) [Text Generation](https://texonom.com/text-generation-cb0e475216c043fdbb4f309d324c9d19)	## Autoregressive Model, Causal language model 이전 시점의 출력을 현재 시점의 입력으로 사용하여, 시퀀스 데이터를 생성하는 모델 단방향 모델을 통해 다음 단어를 예측, 추측하는 언어모델 [Decoder Input IDs](https://texonom.com/decoder-input-ids-2b1e1d3c7c274b0880191a54290bed27) ### Autoregressive 현재 값이 이전 값들의 선형 조합으로 나타나는 모델을 의미 Transformer의 Decoder만 사용하므로 이전 시점의 출력을 현재 시점의 입력으로 사용하여, 시퀀스 데이터를 생성하는 방식을 사용 이 방식은 문장 전체를 한 번에 처리하지 않고, 단어나 토큰 단위로 처리 이전 단어들을 기반으로 다음 단어를 예측 즉 문장 전체를 한 번에 처리하지 않기 때문에, 문장 전체를 볼 수 없다 > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/) > [Decoder models - Hugging Face NLP Course](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt)	36e78a40265c473d90197089aebfa83b
Encoder Model	Seq2Seq Models	Mar 6, 2023	Alan Jo	Alan Jo	Jul 30, 2023	[Decoder Model](https://texonom.com/decoder-model-36e78a40265c473d90197089aebfa83b)	## Auto-encoding model, Masked Language Model transform text or images into a condensed numerical representation called an embedding ### These models are often characterized as having bi-directional attention 입력 문장을 벡터화하고 다시 복원하는 모델 입력 문장의 의미를 보존하면서, 모델이 처리하기 쉬운 형태로 변환 The pretraining of these models usually revolves around somehow corrupting a given sentence and tasking the model with finding or reconstructing the initial sentence ### auto-encoding 입력 시퀀스를 인코딩하여 전부 [Latent Space](https://texonom.com/latent-space-d67b6bdef18b4058bfbc3d25f87ec087) 에 매핑한다 입력 시퀀스를 토큰 단위로 처리하지만 Self-Attention으로 입력 시퀀스의 각 위치마다 해당 위치와 다른 모든 위치 간의 유사도를 계산하기 때문에 문맥고려 가능하다. 이때 유사도가 높은 위치들은 해당 위치의 임베딩 벡터에 더 많은 가중치가 부여 > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/) > [Encoder models - Hugging Face NLP Course](https://huggingface.co/learn/nlp-course/chapter1/5?fw=pt)	321d79943c8940fcaac0c9ccca0b5f6f
Transformer Model	Seq2Seq Models	Aug 17, 2020	Alan Jo	Alan Jo	May 29, 2023	[Attention Mechanism](https://texonom.com/attention-mechanism-762711860abb45f59904f1ac4e4af285) [Attention is all you need](https://texonom.com/attention-is-all-you-need-f52e664318eb47c0aa6cbd27b9a4c491)	## Self Attention is the core feature the Transformer gains a wider perspective and can attend to multiple interaction levels within the input sentence. 어텐션을 Encoder 보정을 위한 용도로서 사용하는 것이 아니라 어텐션만으로 인코더와 디코더 생성 모든 토큰을 동시에 받아 연산하기 때문에 병렬연산이 가능 인코더 블럭과 디코더 블럭이 각각 6개씩 모여있는 구조 인코더 블록은 2개의 sub-layer(Multi-Head(self) Attention, Feed Forward)로 나눌 수 있고 디코더 블록은 3개의 sub-layer(Masked Multi-Head(self) Attention, Multi-Head (Encoder-Decoder) Attention, Feed Forward)로 나눌 수 있다 기존의 Attention과는 다르게 각 벡터가 모두 가중치 벡터라는 점 ### Transformer Model Notion \|Title\| \|:-:\| \|[Attention is all you need](https://texonom.com/attention-is-all-you-need-f52e664318eb47c0aa6cbd27b9a4c491)\| \|[NLP Transformer Encoder](https://texonom.com/nlp-transformer-encoder-d5bc8cebe43d4fc7896d8e0290683b18)\| \|[NLP Transformer Decoder](https://texonom.com/nlp-transformer-decoder-1ee79f4754d24aa8b64312399782cbec)\| \|[Transformer Attention](https://texonom.com/transformer-attention-84d340d69722409a96ec1d806970608a)\| \|[Transformer Model Tool ](https://texonom.com/transformer-model-tool-01c749c1e9254d8c85ec5ef2feb0566b)\| ### Transformer Models \|Title\| \|:-:\| \|[RETRO Transformer](https://texonom.com/retro-transformer-e3efdb06419948ebb1be444f4e124867)\| \|[Torchscale](https://texonom.com/torchscale-c4b82c1245d74c7092617e70b9ddab51)\| \|[BART](https://texonom.com/bart-48bd2258eee74af3b2d929b03bb9553b)\| \|[RMT](https://texonom.com/rmt-f076841449f2419a82b3ce09281b9bb9)\| \|[Transformer-XL](https://texonom.com/transformer-xl-eb2128e01dc744539b6825f3880da761)\| ### Architecture > [Transformer’s Encoder-Decoder: Let’s Understand The Model Architecture - KiKaBeN](https://kikaben.com/transformers-encoder-decoder/) ### Pseudo Source code > [Transformers for software engineers](https://blog.nelhage.com/post/transformers-for-software-engineers/) ### Korean > [점프 투 파이썬](https://wikidocs.net/31379) > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리)	f3e8053cc5b447a2bc9c6b5d0874dafc
Decoder Input IDs	Decoder Model	null	null	null	null	null	## token indices	2b1e1d3c7c274b0880191a54290bed27
Attention is all you need	Transformer Model Notion	Aug 23, 2020	Alan Jo	Alan Jo	May 30, 2023	[Transformer Model](https://texonom.com/transformer-model-f3e8053cc5b447a2bc9c6b5d0874dafc)	2017년 발표된 트랜스포머 구조를 처음 발표한 논문 해당 논문은 Non-recurrent sequence to sequence encoder-decoder model을 만드는 것이 목표 Replaced [RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16) Encoder Decoder Model ## Background ### 1. Sequential computation sequence to sequence한 문제를 푸는 과정에서, Encoder-Decoder 구조의 RNN 모델들이 좋은 성능을 냈다. ### 2. Long term dependency RNN의 경우, Long term dependency의 문제가 항상 따라다니고, CNN의 경우 kernel 안에서 O(1)이나, kernel 간 정보가 공유되지 않는다. # Model Architecture 6개의 stack - 하나의 인코더는 Self-Attention Layer와 Feed Forward Neural Network(2개의 Sub-layer) - Encoder - Multi-Head Attention - Positional Encoding - Relative Positioning - The Residuals - Decoder ### Author > [ashVaswani](https://twitter.com/ashVaswani) ### pdf > [Untitled](https://arxiv.org/pdf/1706.03762.pdf) > [Attention Is All You Need(transformer) paper 정리](https://medium.com/@omicro03/attention-is-all-you-need-transformer-paper-%EC%A0%95%EB%A6%AC-83066192d9ab) > [점프 투 파이썬](https://wikidocs.net/31379)	f52e664318eb47c0aa6cbd27b9a4c491
NLP Transformer Decoder	Transformer Model Notion	Mar 7, 2023	Alan Jo	Alan Jo	Mar 7, 2023		### 3 Sub Layer [Masked Self-Attention](https://texonom.com/masked-self-attention-cb0e29589c93423780cc0ca60260f3e4) + [Multi-head Attention](https://texonom.com/multi-head-attention-d9dfc39b27494123ae4c81f3b98e50b5) + [Position-wise FFNN](https://texonom.com/position-wise-ffnn-6fe30c96aa9245d898c73ec34625377d) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fa99c4277-03ed-4ec6-95ff-c4e91a06f844%2FUntitled.png?table=block&id=00eb4408-f746-4d0e-8313-6f5a8e6b94e1&cache=v2)	1ee79f4754d24aa8b64312399782cbec
NLP Transformer Encoder	Transformer Model Notion	Mar 7, 2023	Alan Jo	Alan Jo	Mar 7, 2023		### 2 Sub Layer [Multi-head Attention](https://texonom.com/multi-head-attention-d9dfc39b27494123ae4c81f3b98e50b5) + [Position-wise FFNN](https://texonom.com/position-wise-ffnn-6fe30c96aa9245d898c73ec34625377d) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fa99c4277-03ed-4ec6-95ff-c4e91a06f844%2FUntitled.png?table=block&id=d48db9af-a165-4db5-842f-80d8046d66c6&cache=v2)	d5bc8cebe43d4fc7896d8e0290683b18
Transformer Attention	Transformer Model Notion	Apr 3, 2023	Alan Jo	Alan Jo	Apr 3, 2023		### Transformer Attentions \|Title\| \|:-:\| \|[Encoder-Decoder Attention](https://texonom.com/encoder-decoder-attention-4b3bb5b4aeac4681bf6bdca66ea79e04)\| \|[Encoder Self-Attention](https://texonom.com/encoder-self-attention-c1e03ab0c57f44c395f2056e201f4326)\| \|[Masked Self-Attention](https://texonom.com/masked-self-attention-cb0e29589c93423780cc0ca60260f3e4)\|	84d340d69722409a96ec1d806970608a
Transformer Model Tool	Transformer Model Notion	May 22, 2023	Alan Jo	Alan Jo	May 22, 2023	[Transformers.js](https://texonom.com/transformersjs-3e7f8bb1d88940298936f13ee1ce7ed7)	### Transformer Model Tools \|Title\| \|:-:\| \|[ ](https://texonom.com/3ddce975d3aa4f6dbca6e3d4b3eb6e6e)\|	01c749c1e9254d8c85ec5ef2feb0566b
Encoder-Decoder Attention	Transformer Attentions	Mar 5, 2023	Alan Jo	Alan Jo	Apr 3, 2023		input과 output의 정보를 엮어주는 역할	4b3bb5b4aeac4681bf6bdca66ea79e04
Encoder Self-Attention	Transformer Attentions	Mar 5, 2023	Alan Jo	Alan Jo	Apr 3, 2023			c1e03ab0c57f44c395f2056e201f4326
Masked Self-Attention	Transformer Attentions	Mar 7, 2023	Alan Jo	Alan Jo	Apr 3, 2023		디코더 블럭에서 사용되는 특수한 Self-Attention 디코더는 [Autoregressive](https://texonom.com/autoregressive-93cf5710b4b54730a7e7efcc6e0fc642) 하기 때문에 이후단어 보지않고 예측해야 그래서 뒤에 보지 않도록 Masking한다 > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리)	cb0e29589c93423780cc0ca60260f3e4
[trl](https://github.com/lvwerra/trl)	Transformer Model Tools	May 22, 2023	Alan Jo	Alan Jo	May 22, 2023			3ddce975d3aa4f6dbca6e3d4b3eb6e6e
BART	Transformer Models	Mar 25, 2023	Alan Jo	Alan Jo	May 31, 2023	[AI Summarization](https://texonom.com/ai-summarization-a314a6fb3162447086b8d2526ae8ef16) [KoBART-summarization](https://github.com/seujung/KoBART-summarization) [BERT](https://texonom.com/bert-e282abe8a34543988a0e71f6c8701ad2)	**Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension** 2019 > [BART: Denoising Sequence-to-Sequence Pre-training for Natural...](https://arxiv.org/abs/1910.13461) > [BART Text Summarization vs. GPT-3 vs. BERT: An In-Depth Comparison \| Width.ai](https://www.width.ai/post/bart-text-summarization) > [BART 논문 리뷰](https://dladustn95.github.io/nlp/BART_paper_review/)	48bd2258eee74af3b2d929b03bb9553b
RETRO Transformer	Transformer Models	Jul 5, 2022	Alan Jo	Alan Jo	May 22, 2023	[Vector Database](https://texonom.com/vector-database-5dfdb6e2bc294fed8ae80eaea2ee5c26) [Deepmind](https://texonom.com/deepmind-5eb171c77b344d4786a9a5b23ae70eca)	### Retrieval-Enhanced ### Fast 밖에다 retrieval database를 두는 형태 ### Implementations [RETRO-pytorch](https://github.com/lucidrains/RETRO-pytorch) > [Improving language models by retrieving from trillions of tokens](https://www.deepmind.com/publications/improving-language-models-by-retrieving-from-trillions-of-tokens) > [The Illustrated Retrieval Transformer](https://jalammar.github.io/illustrated-retrieval-transformer/) > [RETRO Is Blazingly Fast](http://mitchgordon.me/ml/2022/07/01/retro-is-blazing.html) ### Korean > [RETRO: Improving language models by retrieving from trillions of tokens](https://velog.io/@nawnoes/RETRO-Improving-language-models-by-retrieving-from-trillions-of-tokens)	e3efdb06419948ebb1be444f4e124867
RMT	Transformer Models	May 3, 2023	Alan Jo	Alan Jo	May 3, 2023	[LM-RMT](https://github.com/booydar/LM-RMT) [RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16)	## Recurrent Memory Transformer GPT-4’s maximum input token for inference is 32000204 This model can 2 million > [Recurrent Memory Transformer](https://arxiv.org/abs/2207.06881) > [“모든 것 바꿀지도”... 기억력 GPT-4 63배 ‘RMT’ 기반 AI 등장](https://contents.premium.naver.com/themiilk/business/contents/230426095632265ym)	f076841449f2419a82b3ce09281b9bb9
Torchscale	Transformer Models	Dec 2, 2022	Alan Jo	Alan Jo	May 22, 2023	[Pytorch](https://texonom.com/pytorch-2dd232d99b3a46d5b7d1e4e686070686) [torchscale](https://github.com/microsoft/torchscale)		c4b82c1245d74c7092617e70b9ddab51
Transformer-XL	Transformer Models	May 3, 2023	Alan Jo	Alan Jo	May 3, 2023	[transformer-xl](https://github.com/kimiyoung/transformer-xl)	Attentive Language Models Beyond a Fixed-Length Context > [Recurrent Memory Transformer](https://arxiv.org/abs/2207.06881)	eb2128e01dc744539b6825f3880da761
Attention Mechanism	Seq2Seq Notion	Mar 5, 2023	Alan Jo	Alan Jo	Jul 28, 2023	[Encoder Model](https://texonom.com/encoder-model-321d79943c8940fcaac0c9ccca0b5f6f) [Decoder Model](https://texonom.com/decoder-model-36e78a40265c473d90197089aebfa83b) [RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16)	한 번에 전체 문장을 읽고 문장을 기반으로 각 단어의 표현을 병렬로 계산가능 Imagine yourself in a library. You have a specific question (query). Books on the shelves have titles on their spines (keys) that suggest their content. You compare your question to these titles to decide how relevant each book is, and how much attention to give each book. Then, you get the information (value) from the relevant books to answer your question. the query vector points to the current input word (aka context). The *keys* represent the words in the input sentence. The key vectors help the model understand how each word relates to the context word. Attention is how much weight the query word should give each word in the sentence. This is computed via a dot product between the query vector and all the key vectors. These dot products then go through a softmax which makes the attention scores (across all keys) sum to 1. Each word is also represented by a *value* which contains the information of that word. As a result, each context word is now represented by an attention-based weightage of all the words in the sentence. ### NLP Attention Notion \|Title\| \|:-:\| \|[Self-Attention](https://texonom.com/self-attention-d06bde44563f455b951d17955d820f77)\| \|[Multi-head Attention](https://texonom.com/multi-head-attention-d9dfc39b27494123ae4c81f3b98e50b5)\| \|[Cross-Attention](https://texonom.com/cross-attention-6365d422d2bf4be78dac038df3a19aae)\| ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F36626d64-ac93-430d-aa3b-11ead7cb27e3%2FUntitled.png?table=block&id=6ce231d7-c03f-4556-b56f-0d2d3032f7a3&cache=v2) ### NLP Attention Usages \|Title\| \|:-:\| \|[Flash Attention](https://texonom.com/flash-attention-ca8093deb6c648059aff1580b0ddbf68)\| \|[Dilated Attention](https://texonom.com/dilated-attention-d9000e3460f547bca3f15ac9b8d7e36c)\| \|[Multi Query Attention](https://texonom.com/multi-query-attention-5641aba38a8b47caa3a9f364c13789f1)\| \|[Group Query Attentiion](https://texonom.com/group-query-attentiion-d72d2da59fba4e9d937a4ec856dac90f)\| \|[PagedAttention](https://texonom.com/pagedattention-abf197357343437681fb878ae5700926)\| > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/) > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리) > [16-01 트랜스포머(Transformer)](https://wikidocs.net/31379)	762711860abb45f59904f1ac4e4af285
*Copy mechanism*	Seq2Seq Notion	Mar 8, 2023	Alan Jo	Alan Jo	May 29, 2023		디코딩 과정에서 문장을 생성할 때 필요한 어휘가 출력 사전(output vocabulary)에 없는 문제(Out-of-Vocabulary)와고유명사들의 출력 확률이 작아지는 문제를 해결하기 위해 고안된 방법으로, 출력에 필요한 어휘를 입력 열에서찾아 출력 없이 복사(copy)하는 사용하는 방법 > [Untitled](https://koreascience.kr/article/CFKO201612470014629.pdf)	b5dee9c80ca24bb993b5f152129b3577
Cross-Attention	NLP Attention Notion	Apr 9, 2023	Alan Jo	Alan Jo	Jul 28, 2023	[MLLM](https://texonom.com/mllm-98172e9446c04bcc9cf52b2bc5d0bd17)	## Encoder-Decoder Attention in [Decoder Model](https://texonom.com/decoder-model-36e78a40265c473d90197089aebfa83b) Query : 디코더 벡터 / Key = Value : 인코더 벡터 인코더의 출력과 디코더의 현재 상태를 이용하여 디코더가 다음 단어를 예측하는 데 사용 MLLM에서 매우 중요한 부분 이미지나 비디오 데이터와 텍스트 데이터 간의 상호작용을 모델링 $$h_i = softmax((x_i * K_v) / sqrt(d_k)) * V_v$$	6365d422d2bf4be78dac038df3a19aae
**Multi-head Attention**	NLP Attention Notion	Mar 5, 2023	Alan Jo	Alan Jo	Jul 28, 2023		### Multiple heads lets the model consider multiple words simultaneously. Because we use the softmax function in attention, it amplifies the highest value while squashing the lower ones. As a result, each head tends to focus on a single element. Multiple heads let us attend to several words. It also provides redundancy, where if any single head fails, we have the other attention heads to rely on. 입력 벡터를 여러 개의 헤드로 분할하여 각각 어텐션을 수행하고 결과를 결합하는 방식으로 동작 > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/)	d9dfc39b27494123ae4c81f3b98e50b5
Self-Attention	NLP Attention Notion	Mar 5, 2023	Alan Jo	Alan Jo	Jul 28, 2023		### 텍스트 데이터 내에서 단어 간의 상호작용을 모델링 Self-attention enables the encoder to weigh the importance of each word and capture both short and long-range dependencies. Self-attention enables the decoder to focus on different parts of the output generated so far. - 인코더의 셀프 어텐션 : Query = Key = Value - 디코더의 마스크드 셀프 어텐션 : Query = Key = Value self-attention 은 어텐션을 자기 자신에게 수행한다는 의미 문장 내부 요소의 관계를 잘 파악하기 위해서 문장 자신에 대해 어텐션 매커니즘을 적용 - 쿼리 : 분석하고 하는 단어에 대한 가중치 벡터 - 키 : 각 단어가 쿼리에 해당하는 단어와 얼마나 연관있는 지를 비교하기 위한 가중치 벡터 - 밸류 : 각 단어의 의미를 살려주기 위한 가중치 벡터 ### self Attention 과정 1. 특정 단어의 쿼리(q) 벡터와 모든 단어의 키(k) 벡터를 내적한다. 내적해서 나온 값은 Attention Score가 된다. 2. 보정으로 트랜스포머에서는 이 가중치를 q,k,v 벡터 차원 dk 의 제곱근인 dk로 나누어 준다. 3. Softmax로 쿼리에 해당하는 단어와 문장 내 다른 단어가 가지는 관계의 비율 계산 4. Value 각 단어의 벡터를 곱해준 후 모두 더한다. > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리)	d06bde44563f455b951d17955d820f77
Dilated Attention	NLP Attention Usages	Jul 13, 2023	Alan Jo	Alan Jo	Jul 28, 2023		![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fe84a0f0b-ccda-469a-978d-c0ef60558a4b%2FUntitled.png?table=block&id=6014ac5c-6ab5-4d96-affe-7f504b2c4355&cache=v2) > [Microsoft’s LongNet Scales Transformer to One Billion Tokens](https://medium.com/syncedreview/microsofts-longnet-scales-transformer-to-one-billion-tokens-af02ff657d87)	d9000e3460f547bca3f15ac9b8d7e36c
Flash Attention	NLP Attention Usages	Jun 29, 2023	Alan Jo	Alan Jo	Jul 28, 2023	[flash-attention](https://github.com/HazyResearch/flash-attention)	Tiling과 Recomputation을 사용하여 Attention을 가속화 softmax를 block단위를 쪼개서 계산 CUDA kernel를 사용 > [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness \| Wonbeom Jang](https://www.wonbeomjang.kr/blog/2023/fastattention/) > [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness \| Wonbeom Jang](https://www.wonbeomjang.kr/blog/2023/fastattention/)	ca8093deb6c648059aff1580b0ddbf68
Group Query Attentiion	NLP Attention Usages	Aug 2, 2023	Alan Jo	Alan Jo	Aug 2, 2023			d72d2da59fba4e9d937a4ec856dac90f
Multi Query Attention	NLP Attention Usages	Aug 2, 2023	Alan Jo	Alan Jo	Aug 2, 2023			5641aba38a8b47caa3a9f364c13789f1
PagedAttention	NLP Attention Usages	Aug 3, 2023	Alan Jo	Alan Jo	Aug 3, 2023		Efficient management of attention key and value memory	abf197357343437681fb878ae5700926
Deep Learning Compiler	Deep Learning Usages	Mar 22, 2022	Alan Jo	Alan Jo	Mar 22, 2022		### Deep Learning Compiler Tools \|Title\| \|:-:\| \|[Nebullvm](https://texonom.com/nebullvm-d967865085f84f5ab2b1a4a5ea661dfd)\|	7d79af9683764b6b983793c1856578c6
Deep Learning Tool	Deep Learning Usages	Oct 6, 2021	Alan Jo	Alan Jo	May 29, 2023	[tuning_playbook](https://github.com/google-research/tuning_playbook)	### Deep Learning Hubs \|Title\| \|:-:\| \|[HuggingFace](https://texonom.com/huggingface-eb25c513b432477e9da51ca19bb06833)\| \|[Pytorch Hub](https://texonom.com/pytorch-hub-d46f5b09e91241578a0c63b4847396fb)\| ### Deep Learning Tools \|Title\| \|:-:\| \|[fastbook](https://texonom.com/fastbook-d022baa785a14ce5af5cf1ef59995cb1)\| \|[Netron](https://texonom.com/netron-2e410ef29aaa4a7d9e2a2777ce4dd3ee)\|	a14ea6f4574342ef974443634e27c6ce
Learn Deep Learning	Deep Learning Usages	Jun 6, 2023	Alan Jo	Alan Jo	Jun 6, 2023		> [0020 DL Terms & Concepts - Deepest Documentation](https://deepestdocs.readthedocs.io/en/latest/002_deep_learning_part_1/0020/)	4382083fb54d4705984e7f45a9af2d86
Sentiment Neuron	Deep Learning Usages	May 20, 2023	Alan Jo	Alan Jo	May 20, 2023		### Proof of compressed data in deep network before gpt1 ### Why ImportantWhy Important > [일리야 수츠케버와 AGI의 미싱링크](https://www.youtube.com/watch?v=LQviQS24uQY&t=840) > [Unsupervised sentiment neuron](https://openai.com/research/unsupervised-sentiment-neuron) > [Sentiment Neuron](https://tensorflow.blog/2017/04/07/sentiment-neuron/)	2e44e9754b894534af6c121b6d6074d6
Nebullvm	Deep Learning Compiler Tools	Mar 22, 2022	Alan Jo	Alan Jo	Mar 22, 2022		[nebuly](https://github.com/nebuly-ai/nebullvm)	d967865085f84f5ab2b1a4a5ea661dfd
HuggingFace	Deep Learning Hubs	Jun 27, 2022	Alan Jo	Alan Jo	Jul 17, 2023	[AI Industry](https://texonom.com/ai-industry-d8709bd0498145e3a66af6da3f963fa7) [Pytorch](https://texonom.com/pytorch-2dd232d99b3a46d5b7d1e4e686070686)	### [Pytorch Hub](https://texonom.com/pytorch-hub-d46f5b09e91241578a0c63b4847396fb) + [Lightening AI](https://texonom.com/lightening-ai-af651491f5474031bc51e6d1a99d5f22) ### HuggingFace Usages \|Title\| \|:-:\| \|[HuggingFace Hub](https://texonom.com/huggingface-hub-9b16b4841f2d4bdd9827120c47677b29)\| \|[Huggingface Model](https://texonom.com/huggingface-model-f985be46b10e462f91e65576b81f452f)\| \|[Huggingface Dataset](https://texonom.com/huggingface-dataset-43582d2257474a83874ecec2d4c6ab44)\| \|[HuggingFace Space](https://texonom.com/huggingface-space-e37f59c218ca48cebe509d3ab2381b34)\| ### LLM Leaderboard > [Open LLM Leaderboard - a Hugging Face Space by HuggingFaceH4](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) > [[D] HuggingFace ecosystem vs. Pytorch Lightning for big research NLP project with many collaborators.](https://www.reddit.com/r/MachineLearning/comments/si08qt/d_huggingface_ecosystem_vs_pytorch_lightning_for/)	eb25c513b432477e9da51ca19bb06833
Pytorch Hub	Deep Learning Hubs	May 29, 2023	Alan Jo	Alan Jo	May 29, 2023	[HuggingFace](https://texonom.com/huggingface-eb25c513b432477e9da51ca19bb06833) [Pytorch](https://texonom.com/pytorch-2dd232d99b3a46d5b7d1e4e686070686)	### Pytorch Hub Usages \|Title\| \|:-:\|	d46f5b09e91241578a0c63b4847396fb
Huggingface Dataset	HuggingFace Usages	May 23, 2023	Alan Jo	Alan Jo	Jul 19, 2023		### Huggingface Dataset Usages \|Title\| \|:-:\| \|[ HuggingFace Datasets Jax](https://texonom.com/huggingface-datasets-jax-a92d62f342de4ddc83820623059ca02e)\| > [Create a dataset](https://huggingface.co/docs/datasets/create_dataset)	43582d2257474a83874ecec2d4c6ab44
HuggingFace Hub	HuggingFace Usages	Jun 19, 2023	Alan Jo	Alan Jo	Jul 17, 2023		[huggingface_hub](https://github.com/huggingface/huggingface_hub) > [Quickstart](https://huggingface.co/docs/huggingface_hub/quick-start)	9b16b4841f2d4bdd9827120c47677b29
Huggingface Model	HuggingFace Usages	May 23, 2023	Alan Jo	Alan Jo	Jun 29, 2023		### Huggingface Model Usages \|Title\| \|:-:\| \|[Huggingface Provider](https://texonom.com/huggingface-provider-8bd64d5951774d6f9ed623abbe471b4c)\| \|[Huggingface H4](https://texonom.com/huggingface-h4-9b33bfe491704142929a794edd95a7df)\| \|[Huggingface Model Card](https://texonom.com/huggingface-model-card-39d3f0d8805c4cb0939b623b8bac64ea)\| > [Models](https://huggingface.co/docs/hub/models)	f985be46b10e462f91e65576b81f452f
HuggingFace Space	HuggingFace Usages	Mar 26, 2023	Alan Jo	Alan Jo	Jul 13, 2023	[Gradio](https://texonom.com/gradio-7e1647e9bc174161b4d7e8f44dd53707) [Streamlit](https://texonom.com/streamlit-9e295c64d27e4999878a022b1c538964)	### HuggingFace Space SDK \|Title\| \|:-:\| \|[HuggingFace Docker Space](https://texonom.com/huggingface-docker-space-0f250609ad1f493bb91146813de8d8a6)\|	e37f59c218ca48cebe509d3ab2381b34
HuggingFace Datasets Jax	Huggingface Dataset Usages	May 30, 2023	Alan Jo	Alan Jo	Jul 17, 2023		> [Use with JAX](https://huggingface.co/docs/datasets/use_with_jax)	a92d62f342de4ddc83820623059ca02e
Huggingface H4	Huggingface Model Usages	Jul 9, 2023	Alan Jo	Alan Jo	Jul 9, 2023		### helpful, honest, harmless, and huggy [StarChat](https://texonom.com/starchat-ea316ec509564a51b56ad92b22831220)	9b33bfe491704142929a794edd95a7df
Huggingface Model Card	Huggingface Model Usages	Jul 19, 2023	Alan Jo	Alan Jo	Jul 30, 2023		### Disable API ```type inference: false ``` - tags - pipeline tags - etc ### Widgets > [Widgets](https://huggingface.co/docs/hub/models-widgets) > [Disable Hosted inference API](https://discuss.huggingface.co/t/disable-hosted-inference-api/10379)	39d3f0d8805c4cb0939b623b8bac64ea
Huggingface Provider	Huggingface Model Usages	Jun 29, 2023	Alan Jo	Alan Jo	Aug 5, 2023		### Companies > [amazon (Amazon Web Services)](https://huggingface.co/amazon) > [stabilityai (Stability AI)](https://huggingface.co/stabilityai) > [EleutherAI (EleutherAI)](https://huggingface.co/EleutherAI) > [allenai (Allen Institute for AI)](https://huggingface.co/allenai) ### Model Users > [ehartford (Eric Hartford)](https://huggingface.co/ehartford) > [psmathur (Pankaj Mathur)](https://huggingface.co/psmathur) > [TheBloke (Tom Jobbins)](https://huggingface.co/TheBloke) > [jncraton (Jon)](https://huggingface.co/jncraton) > [bhenrym14 (Brandon)](https://huggingface.co/bhenrym14) ### Organization > [decapoda-research (Decapoda Research)](https://huggingface.co/decapoda-research) > [openchat (OpenChat)](https://huggingface.co/openchat) > [MBZUAI (Mohamed Bin Zayed University of Artificial Intelligence)](https://huggingface.co/MBZUAI) > [openchat/openchat · Hugging Face](https://huggingface.co/openchat/openchat) > [OpenAssistant (OpenAssistant)](https://huggingface.co/OpenAssistant)	8bd64d5951774d6f9ed623abbe471b4c

PPO

Reinforcement Learnings

Jul 15, 2023

Alan Jo

Jul 15, 2023

### Proximal Policy Optimization balance between ease of implementation, sample complexity, and tuning 2017년 OpenAI에서 개발된 모델 없는 강화 학습 알고리즘 태-동작 쌍에 값을 할당하는 대신 정책 공간을 검색 data를 만들어낸 old policy와 train 대상이 되는 현재의 network의 new policy를 구분 [PPO2](https://texonom.com/ppo2-378944bb3b8848a185c20fc529c390d2) > [Proximal Policy Optimization Algorithms](https://arxiv.org/abs/1707.06347) > [Proximal Policy Optimization](https://openai.com/research/openai-baselines-ppo)

87ce8ebe81f84ac1a7617b1f6def9e26

Q-Learning

Reinforcement Learnings

Nov 5, 2019

Alan Jo

Seong-lae Cho

Aug 31, 2023

[SARSA](https://texonom.com/sarsa-e5d847a4fb6e41cdad5b9ffb6e974a10)

### Approximate Q-Learning Off police 시간차제어 현재 행동하는 정책과는 독립적으로 학습 ### Reinforce > [강화학습 기초(Q-learning)](https://bluediary8.tistory.com/18)

6fb81e4a53ab4e3097784cde99c8c038

RLHF

Reinforcement Learnings

Apr 30, 2023

Alan Jo

Jul 15, 2023

[AI Alignment](https://texonom.com/ai-alignment-f676f1a29ffd45e19b3d170afa4f2244) [Active Learning](https://texonom.com/active-learning-85d42ac892e84e5ba4fd2727f1791f65)

## Reinforcement learning from human feedback ### Limitation LM의 근본적인 문제인 Size, hallucination을 아직까지는 개선할 수는 없는 한계점 Scaling 이슈, 너무 복잡 > [RLHF란?](https://velog.io/@nellcome/RLHF란) > [Reinforcement learning from human feedback](https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback)

4b184f9c9e8b4c7a8861fb6374e91aa6

RRHF

Reinforcement Learnings

Jul 15, 2023

Alan Jo

Jul 15, 2023

[RLHF](https://texonom.com/rlhf-4b184f9c9e8b4c7a8861fb6374e91aa6)

### ***R****ank ****R****esponse to align ****H****uman ****F****eedback* efficiently align language model output probabilities with human preferences as robust as fine-tuning and it only needs 1 to 2 models during tuning

073c6bdf48444a1b9acad9e65057f3d0

SARSA

Reinforcement Learnings

Jul 18, 2023

Alan Jo

Aug 31, 2023

## state-action-reward-state-action 현재 가지고 있는 **큐함수** 를 토대로 샘플을 탐욕 정책으로 모으고, 그 샘플로 방문한 큐함수를 업데이트하는 과정을 반복 Almost start of Reinforcement learning after [Policy Iteration](https://texonom.com/policy-iteration-55cbe9219f8a48fbb850f64b677e847a) GPI에서는 벨만 방정식에 따라 정책을 평가 Temporal-Difference 방법에서는 가치 이터레이션의 방법을 도입 **현재 상태의 큐함수를 보고 판단한다면 환경의 모델을 몰라도 된다** 시간차 제어에서는 **큐함수를 사용한 탐욕 정책** 을 통해 행동을 선택 초기의 에이전트에게 탐욕정책은 **잘못된 학습으로 가게할 가능성이 크기 때문에** epsilon**-탐욕 정책** 을 사용 ### Limitation 특정 state에 **갇혀버리는 현상** 자신이 행동한 대로 학습하는 것을 **On-Policy 시간차 제어** 그래서 [Q-Learning](https://texonom.com/q-learning-6fb81e4a53ab4e3097784cde99c8c038) > [(7) 살사(SARSA)와 큐러닝(Q-Learning)](https://jang-inspiration.com/sarsa-qlearning#132c2000aab74666a2273d8b2d71cdac)

e5d847a4fb6e41cdad5b9ffb6e974a10

TRPO

Reinforcement Learnings

Jul 15, 2023

Alan Jo

Jul 15, 2023

### ****Trust Region Policy Optimization****

23c0f917124d419cbfe478d20b804276

PPO2

PPO

null

> [Untitled](https://github.com/openai/baselines/blob/master/baselines/ppo2/ppo2.py)

378944bb3b8848a185c20fc529c390d2

ML Compiler Optimization

Machine Learning Techniques

Jul 9, 2022

Alan Jo

Apr 25, 2023

[Compiler Optimization](https://texonom.com/compiler-optimization-3eee7067387b45d7875df50f4473ad18) [Parallel Training](https://texonom.com/parallel-training-4a8896bc837b4dddb47c3700b715cdc8)

[relax](https://github.com/mlc-ai/relax) ### ML Compiler Optimization Tools |Title| |:-:| |[XLA](https://texonom.com/xla-4f75dff43dc0451aa5ca92e9218a3028)| |[MLGO](https://texonom.com/mlgo-60d88727f0b944ec9ce62d254c2e8a76)| |[Hidet](https://texonom.com/hidet-7b706024cb3040269f71e47b5b87d2b2)|

011e7bd0ba8f417bb111ec5ea2171c8e

Parallel Training

Machine Learning Techniques

Mar 15, 2022

Alan Jo

Apr 25, 2023

[ML Compiler Optimization](https://texonom.com/ml-compiler-optimization-011e7bd0ba8f417bb111ec5ea2171c8e)

### data parallelism or model parallelism - In data parallelism, the data is split into multiple parts - in model parallelism, different parts of the model are processed by separate processors ### Parallel Training Notion |Title| |:-:| |[Model Parallelism](https://texonom.com/model-parallelism-76dd813ada7b4e50b645af2f05821d48)| |[Data Parallelism](https://texonom.com/data-parallelism-8e90f1c595a84dad9f4e921e74f86ba6)| ### Parallel Training Usages |Title| |:-:| |[Parallel Learning Tool](https://texonom.com/parallel-learning-tool-2a9741aa76c14c16a1240f3422f11421)| |[Parallel Training Example](https://texonom.com/parallel-training-example-3c16ddf97fbe43359e7da3dbd3ce96ee)| ![https://xiandong79.github.io](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F74479c70-d665-4f9e-a0df-dda36042a7ee%2FUntitled.png?table=block&id=cf76f928-10e0-4fb6-bb03-ec8afb321aa7&cache=v2) > [딥러닝 모델의 분산학습이란? (Data parallelism과 Model parallelism)](https://lifeisenjoyable.tistory.com/21)

4a8896bc837b4dddb47c3700b715cdc8

Quantum Machine Learning

Machine Learning Techniques

Mar 9, 2022

Alan Jo

Mar 5, 2023

### Quantum Machine Learnings |Title| |:-:| > [Spooky Action Could Help Boost Quantum Machine Learning](https://spectrum.ieee.org/quantum-machine-learning)

8f5276045f4d43b8b96f3b4ec6646f66

Weight Initialization

Machine Learning Techniques

Jun 6, 2023

Alan Jo

Jul 6, 2023

- weight to small random numbers - bias (zero or small nonzero) ### Weight Initialization Usages |Title| |:-:| |[He initialization](https://texonom.com/he-initialization-a255fdaeec8e485faf3215a28ed5fdb9)| |[Xavier Initialization](https://texonom.com/xavier-initialization-40045bfdf72343aea3a234214145f9dd)| > [0025 Initialization - Deepest Documentation](https://deepestdocs.readthedocs.io/en/latest/002_deep_learning_part_1/0025/)

6cfc10eb06f948528aa76a9814a9ac85

Hidet

ML Compiler Optimization Tools

May 1, 2023

Alan Jo

May 1, 2023

[Pytorch](https://texonom.com/pytorch-2dd232d99b3a46d5b7d1e4e686070686)

> [PyTorch](https://pytorch.org/blog/introducing-hidet)

7b706024cb3040269f71e47b5b87d2b2

MLGO

ML Compiler Optimization Tools

Jul 9, 2022

Alan Jo

Mar 11, 2023

[LLVM](https://texonom.com/llvm-5dc6acb10b5244a2af349319ef87c797) [ml-compiler-opt](https://github.com/google/ml-compiler-opt)

### Infrastructure for Machine Learning Guided Optimization > [MLGO: A Machine Learning Framework for Compiler Optimization](https://ai.googleblog.com/2022/07/mlgo-machine-learning-framework-for.html)

60d88727f0b944ec9ce62d254c2e8a76

XLA

ML Compiler Optimization Tools

Mar 11, 2023

Alan Jo

Mar 11, 2023

[xla](https://github.com/openxla/xla)

- pytorch - tensorflow - jax

4f75dff43dc0451aa5ca92e9218a3028

Data Parallelism

Parallel Training Notion

Apr 25, 2023

Alan Jo

Apr 25, 2023

학습 데이터를 여러 GPU에 나눠 학습 ### Data Parallelism Usages |Title| |:-:|

8e90f1c595a84dad9f4e921e74f86ba6

Model Parallelism

Parallel Training Notion

Apr 25, 2023

Alan Jo

Apr 25, 2023

모델을 여러 GPU에 나누는 ### Model Parallelism Usages |Title| |:-:|

76dd813ada7b4e50b645af2f05821d48

Parallel Learning Tool

Parallel Training Usages

Apr 25, 2023

Alan Jo

Apr 25, 2023

### Parallel Training System |Title| |:-:| |[Colossal AI](https://texonom.com/colossal-ai-031c5480b24249ce903fea4e0f8d435c)| |[Megatrom LM](https://texonom.com/megatrom-lm-87be23e9b623465395a0d6a4e94470ae)| |[DeepSpeed](https://texonom.com/deepspeed-3866b23c00eb4d529de6e33dc48ffae7)|

2a9741aa76c14c16a1240f3422f11421

Parallel Training Example

Parallel Training Usages

Apr 25, 2023

Alan Jo

Apr 25, 2023

3c16ddf97fbe43359e7da3dbd3ce96ee

Colossal AI

Parallel Training System

Mar 15, 2022

Alan Jo

Apr 25, 2023

[ColossalAI](https://github.com/hpcaitech/ColossalAI)

> [Colossal-AI](https://colossalai.org/)

031c5480b24249ce903fea4e0f8d435c

DeepSpeed

Parallel Training System

Feb 19, 2021

Alan Jo

Apr 25, 2023

[DeepSpeed](https://github.com/microsoft/DeepSpeed)

### Pipeline Parallelism > [DeepSpeed Pipeline Parallelism](https://velog.io/@nawnoes/DeepSpeed-Pipeline-Parallelism) > [PyTorch Lightning DeepSpeed](https://velog.io/@nawnoes/PyTorch-Lightning-DeepSpeed)

3866b23c00eb4d529de6e33dc48ffae7

Megatrom LM

Parallel Training System

Apr 25, 2023

Alan Jo

Apr 25, 2023

[Megatron-LM](https://github.com/NVIDIA/Megatron-LM)

87be23e9b623465395a0d6a4e94470ae

[gpt-neox](https://github.com/EleutherAI/gpt-neox)

Parallel Learning Examples

Apr 25, 2023

Alan Jo

Apr 25, 2023

[Megatrom LM](https://texonom.com/megatrom-lm-87be23e9b623465395a0d6a4e94470ae) [DeepSpeed](https://texonom.com/deepspeed-3866b23c00eb4d529de6e33dc48ffae7)

48f191fcffa04c068978381a78b4ca8d

**He initialization**

Weight Initialization Usages

Jul 6, 2023

Alan Jo

Jul 6, 2023

[ReLU](https://texonom.com/relu-e582549804da48b893758895e446ffb9)

ReLU + He 초기화 방법이 보편적 > [Untitled](https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/He_Delving_Deep_into_ICCV_2015_paper.pdf)

a255fdaeec8e485faf3215a28ed5fdb9

**Xavier Initialization**

Weight Initialization Usages

Jul 6, 2023

Alan Jo

Jul 6, 2023

## Glorot Initialization 이전 층의 뉴런 개수와 다음 층의 뉴런 개수 이용 여러 층의 기울기 분산 사이에 균형을 맞춘다 S자 형태인 활성화 함수와 함께 사용할 경우에는 좋은 성능을 보이지만, ReLU와 함께 사용할 경우에는 성능이 좋지 않다 ### Uniform Distribution ### Normal distribution > [07-07 기울기 소실(Gradient Vanishing)과 폭주(Exploding)](https://wikidocs.net/61375)

40045bfdf72343aea3a234214145f9dd

Function Transformers

Machine Learning Tools

Jun 13, 2021

Alan Jo

Jan 9, 2023

### Composable transformations > [Transformers](https://huggingface.co/transformers/)

c4396d81a01a4425ba0c6702501c911a

ML Accelerator

Machine Learning Tools

Jun 1, 2022

Alan Jo

Jun 1, 2022

### ML Accelerators |Title| |:-:|

0df3a271237b4e63a342aa7ce704870d

ML Analyze Tool

Machine Learning Tools

Aug 5, 2021

Alan Jo

Apr 19, 2022

5aeb48a9850245eb97e20ec56448a15f

ML Container Tool

Machine Learning Tools

May 11, 2022

Alan Jo

May 11, 2022

14223d2021c748268fba092dee1fa357

ML Feature Store

Machine Learning Tools

Apr 19, 2022

Alan Jo

Apr 19, 2022

f53069a7083b4719b1e5fab18a5a9bbd

ML Platform

Machine Learning Tools

Sep 8, 2021

Alan Jo

Aug 4, 2022

[](https://texonom.com/fee31eaaf53a45c28b0b305c3874b856)

9d4142db8db042ed9e4a79085348cc55

Evidently

ML Analyze Tools

Aug 5, 2021

null

> [GitHub - evidentlyai/evidently: Interactive reports to analyze machine learning models during validation or production monitoring.](https://github.com/evidentlyai/evidently?ref=producthunt?utm_source=tldrnewsletter)

0500b048a7c34d8e91e18207e43cefd1

Uptrain

ML Analyze Tools

Mar 9, 2023

null

[uptrain](https://github.com/uptrain-ai/uptrain)

a12b6a1c3410434ea318cb76a2f99a98

Cog

ML Container Tools

May 11, 2022

Alan Jo

May 11, 2022

[cog](https://github.com/replicate/cog)

1b9175a7b63c43b5b6cd69a223f9d99d

Feathr

ML Feature Stores

Apr 19, 2022

Alan Jo

Apr 19, 2022

[LinkedIn](https://texonom.com/linkedin-1c0eb8ae1ca346a388e79c15b34355dc) [feathr](https://github.com/linkedin/feathr)

> [Open sourcing Feathr - LinkedIn's feature store for productive machine learning](https://engineering.linkedin.com/blog/2022/open-sourcing-feathr linkedin-s-feature-store-for-productive-m) ### Template Gallery |Title| |:-:| |[Template Page](https://texonom.com/template-page-b6dd128730be402fbf47e98d1a81c5f2)|

5ff843203b4745df949b956391bf9423

Template Page

Template Gallery

Apr 19, 2022

Alan Jo

Apr 19, 2022

b6dd128730be402fbf47e98d1a81c5f2

diffgram

ML Platforms

Sep 8, 2021

null

> [GitHub - diffgram/diffgram: Complete training data platform for machine learning delivered as a single application.](https://github.com/diffgram/diffgram)

99e0ffaf73f84615a556b6bbe71c4572

Wandb

ML Platforms

Aug 4, 2022

null

[wandb](https://github.com/wandb/wandb)

2219d228212940068aa5a604af7d5dbc

AdaBoost

ML Meta Algorithms

Oct 6, 2021

Alan Jo

Oct 6, 2021

성능을 향상시키기 위하여 다른 많은 형태의 학습 알고리즘과 결합하여 사용

79feac6204384e7299e08e3cfa40d05e

Deep Learning

Neural Network Notion

Nov 5, 2019

Alan Jo

Seong-lae Cho

Jul 5, 2023

[Neuroscience ](https://texonom.com/neuroscience-b45d9f638a2b4330906556c402307925)

### Neural Network based Machine Learning method composition of differentiable functions Brain Algorithm + [Neural Network](https://texonom.com/neural-network-86f54f9f1de848c1a29c56c24f7d5094) + [Big Data](https://texonom.com/big-data-236ec9f0ed844f4d8a5ca3236dfa442c) limit - explainability, fairness, generalizability, causality Main feature of neural network It can find non-heuristic feature representation ### Deep Learning Notion |Title| |:-:| |[Deep Learning Math](https://texonom.com/deep-learning-math-57c204edef3042568bbd0d5268b877fd)| |[Deep Learning Network](https://texonom.com/deep-learning-network-f368faf8fa634699aeda503d01c193f0)| |[End-to-end Deep Learning](https://texonom.com/end-to-end-deep-learning-16bd62f447a144f1b251acf25f1f8789)| ### Deep Learning Usages |Title| |:-:| |[Deep Learning Tool](https://texonom.com/deep-learning-tool-a14ea6f4574342ef974443634e27c6ce)| |[Deep Learning Compiler](https://texonom.com/deep-learning-compiler-7d79af9683764b6b983793c1856578c6)| |[Sentiment Neuron](https://texonom.com/sentiment-neuron-2e44e9754b894534af6c121b6d6074d6)| |[Learn Deep Learning](https://texonom.com/learn-deep-learning-4382083fb54d4705984e7f45a9af2d86)| ### Interview from popular people > [제프리 힌턴이 말하는 AI의 영향력과 잠재력](https://www.youtube.com/watch?v=IvUw9um4Bv8) > [OpenAI의 핵심, Ilya Sutskever 인터뷰](https://www.youtube.com/watch?v=SGCFeIbpGlU&t=722s)

7d3c8b9ce05b49cf9eed92dbcdc80cfd

Neural Network History

Neural Network Notion

May 11, 2023

Alan Jo

Jun 7, 2023

> [[미라클레터] AI시대 필수품 아날로그 컴퓨터!!](https://stibee.com/api/v1.0/emails/share/KeeaUnUVO5o8muoWpT9bjEtpsL5hny0=)

2127bdb56da54a659e570f47704e40b1

Neural Network Structure

Neural Network Notion

May 11, 2023

Alan Jo

Jul 4, 2023

### Neural Network Components |Title| |:-:| |[Activation Function](https://texonom.com/activation-function-8e52ee5f83a244d88abeeee3fb9497a8)| |[Neural Network Layer](https://texonom.com/neural-network-layer-e10ef2afe1954cf6b909f8aa40077393)| |[Forward Forward Algorithm](https://texonom.com/forward-forward-algorithm-6c989313e382466bba02d15c412fb17f)| |[skip connections](https://texonom.com/skip-connections-d7ad187f3487468db6eea278f0236a22)| ### Neural Networks |Title| |:-:| |[Perceptron](https://texonom.com/perceptron-1deb66f486d54c93bb928d8afba1864c)| |[FFNN](https://texonom.com/ffnn-89ecee87d8b7482e86995950db90eb31)| |[CNN](https://texonom.com/cnn-002bf81a77bc40d1858740d26b61d97b)| |[RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16)| |[ANN](https://texonom.com/ann-d4232205ecf9463c95a911d179c87a84)| |[SNN](https://texonom.com/snn-6cf69239a87e4df9a32b3494862374e4)| |[GNN](https://texonom.com/gnn-58adb81cb2b649d9af19019182960bb2)|

400bbea8029c4eb1a97c0dd063735551

Deep Learning Math

Deep Learning Notion

Nov 5, 2019

Alan Jo

Seong-lae Cho

Mar 26, 2023

### Sub-field of ML: learning representations of data Existing ML uses manually designed features - often over-specified and incomplete - take a long time to design and validate DP - Learned Features are easy to adapt, fast - Deep learning provides a very flexible, (almost?) universal - Effective end-to-end joint system learning > speach recog not good > visual perception good > Question Answering good > 4 + 2 = 6 neurons (not counting inputs) - biases has same number (for each resulting node) > [3 x 4] + [4 x 2] = 20 weights > > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F5986fe3d-198a-4608-959f-6bd5bc738065%2FUntitled.png?table=block&id=8f6601d1-344a-4488-b982-926c53650f7f&cache=v2) > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F4110b2cb-bf0f-4221-b2c9-cc8c6f2c8786%2FUntitled.png?table=block&id=0e436d3f-24b4-4328-bc11-1e91e5aa5747&cache=v2) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Ff7503f3f-0f31-446e-b91b-32842c7f119d%2FUntitled.png?table=block&id=4e369bcf-c0fd-4d82-9523-8156a1c29aac&cache=v2) Optimize (min. or max.)** objective/cost function 𝐽(𝜃)** Generate **error signal **that measures difference between predictions and target values word representation is 74 72 65 65 in x0 but it is pool so we need new presentation of word ## wordnet WordNet: contains the list of synonyms/hypernyms ← using human resources Word’s meaning is given by words that frequently appear close-by ### Context: set of words that appears nearby in the fixed-size window (ex: before/after 5 words) Vector dimension = Number of words * Number of words memroy O(n^2) poor so we use dimensionality reduction (pca, svd) ← almost all value is 0 so l;we can → word embedding 이제 벡터곱으로 similar판단가능 (become dense vector) - from one hot vector > We can make embedding for between other languagex > vector similarity (cosine) > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F30c1772d-b6df-4eed-ad3f-2777aebcef9a%2FUntitled.png?table=block&id=e97e701f-21dd-4ab7-8e3e-d039ad4d6d08&cache=v2) # Language Modeling: Models ofP(text) - sentance score sentance > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F3c48b3c2-5971-4a4d-8c78-896f6b721be1%2FUntitled.png?table=block&id=70af901f-dd8c-4e16-a2b5-408adf172390&cache=v2) > add all is exampel (linearfeatures) > usually softmax the sum > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F66fd7479-7a77-416a-9684-a30972d659e4%2FUntitled.png?table=block&id=4f96ddb9-9fa4-48df-b29f-2012e73945f1&cache=v2) > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F7bcd216a-04a9-411b-9c95-c2c8728c44e7%2FUntitled.png?table=block&id=2b6a582b-5666-47cd-8fa5-0185c4b045bb&cache=v2) soft max is for before output ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F52fd4795-45f9-4e20-8f4a-b9a18df4d4d7%2FUntitled.png?table=block&id=5a5b7c77-f67f-486b-b148-0265844a4461&cache=v2) # 1. P(text) - linear model ## CBOW Predict word based on sum of surroundingembeddings ## Skip-gram Predict each word in the context given theword [very good, good, neutral, bad, verybad] Linear Models can’tLearn FeatureCombinations # 2. Models of P(label |text) ## BOW each word has its own 5 elements corresponding to ## DeepCBOW CombinationFeatures - Each vector has “features” (e.g. is this an animate object? is this a positive word, etc.) # ConvolutionalNetworks ### x/ CNN - pooling weak for long distance feature extractor don't have holistic view of # x/ RNN good for long disatnce fraeture extractor weakness - Indirect passing of information, credit assignment moredifficult Can be slow, due to incrementalprocessing ## ModelingSentences w/n-grams # P(text |text) > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F547f55b2-840c-4e50-80a0-8b936f4a64c3%2FUntitled.png?table=block&id=7fa3ac7f-fd55-4575-b6b1-afbae72438fe&cache=v2) > Conditional LanguageModels > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fba8eaa2d-8c36-4f5b-b222-e60b091aa063%2FUntitled.png?table=block&id=6f689ac1-dfc7-45ac-9f72-5704d81f6dcf&cache=v2) > Calculating the Probability of aSentence > RNN is frequently used in language modeling since RNN can capture long-distance dependencies

57c204edef3042568bbd0d5268b877fd

Deep Learning Network

Deep Learning Notion

Oct 6, 2021

Alan Jo

May 29, 2023

### Deep Learning Models |Title| |:-:| |[Seq2Seq](https://texonom.com/seq2seq-01a9854dffa6417c87d92c11a607250c)| |[GAN](https://texonom.com/gan-66482b5f518d47f6b337eba9a30ff792)| |[Capsule Network](https://texonom.com/capsule-network-a14fc7e569864154aae1ef44106e8991)| |[MANN](https://texonom.com/mann-84ed691391c8439ba8e6b297623c9c0e)|

f368faf8fa634699aeda503d01c193f0

End-to-end Deep Learning

Deep Learning Notion

Apr 3, 2023

Alan Jo

Apr 3, 2023

입력에서 바로 출력을 구할 수 있다 모든 매개변수가 하나의 손실함수에 대해 동시에 훈련되는 경로가 가능한 네트워크를 뜻한다 [Text Tokenizer](https://texonom.com/text-tokenizer-2bbc41eaa76c4674a4f4b9127fbe5da1) [Text Encoding](https://texonom.com/text-encoding-ab7377cfc5c648059de4860510ad9134) 하지않고 메모리 많이 사용 > [What is end-to-end deep learning?](https://velog.io/@jeewoo1025/What-is-end-to-end-deep-learning)

16bd62f447a144f1b251acf25f1f8789

Capsule Network

Deep Learning Models

Aug 21, 2021

Alan Jo

Oct 6, 2021

[CNN](https://texonom.com/cnn-002bf81a77bc40d1858740d26b61d97b)

## CapsNet cnn에서 이미지 recognition 에서 생기는 문제가 없는 인공신경망 detects the rotation and leans it as one of the activation vector ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fb1ec5a35-c647-46dd-90c3-07dcc5e94b7c%2FUntitled.png?table=block&id=9eb6f960-e681-498c-8c63-3b6096168011&cache=v2) > [What is a CapsNet or Capsule Network?](https://medium.com/hackernoon/what-is-a-capsnet-or-capsule-network-2bfbe48769cc) > [Why Do Capsule Networks Work Better Than Convolutional Neural Networks?](https://medium.com/@ashukumar27/why-do-capsule-networks-work-better-than-convolutional-neural-networks-f4a105a53aff)

a14fc7e569864154aae1ef44106e8991

GAN

Deep Learning Models

Nov 18, 2019

Alan Jo

Jun 1, 2023

[Unsupervised learning](https://texonom.com/unsupervised-learning-8cb6e253bfa845b5931d22963ea93019) [Generative Model](https://texonom.com/generative-model-6e5204d2982b4042847aa42e88eb8fb5) [Transfer Learning](https://texonom.com/transfer-learning-442feb66465944eebf144d4e9dd1dbf8)

## Generative Adversarial Network Learn how to generate samples, 위조지폐와 경찰같이 경쟁하며 발전 ### GAN Notion |Title| |:-:| |[Generator Network](https://texonom.com/generator-network-3d3fb237d8f149979c7ed172aca65529)| |[Discriminator Network](https://texonom.com/discriminator-network-a02287522a264d779574871285eed3b6)| |[GAN Minmax Game](https://texonom.com/gan-minmax-game-d3b08fa32fc34cf5920bfc0e80c34b90)| |[GAN Issues](https://texonom.com/gan-issues-be6312c3a8184b55ba443167164101ba)| ### GANs |Title| |:-:| |[3D GAN](https://texonom.com/3d-gan-672af0e2639041e484ed1a26b56f84cb)| |[DCGAN](https://texonom.com/dcgan-ae93b511fed5402c926e818243d8a966)| |[DragGan](https://texonom.com/draggan-bd335de6767243e7b7ec48ff88dea800)| > [Generative adversarial network](https://en.wikipedia.org/wiki/Generative_adversarial_network)

66482b5f518d47f6b337eba9a30ff792

****MANN****

Deep Learning Models

Apr 29, 2023

Alan Jo

May 29, 2023

### Memory Augmented Neural Networks RNN, CNN 등의 기반 모델과 메모리 구조를 결합한 모델 ### ****MANN Notion**** |Title| |:-:| |[Differentiable Neural Computer](https://texonom.com/differentiable-neural-computer-2c182410c5c34222b41605b63c37c777)| |[Neural Turing Machine](https://texonom.com/neural-turing-machine-c03efee39e1942f197f4b3d6553e4ac1)|

84ed691391c8439ba8e6b297623c9c0e

Seq2Seq

Deep Learning Models

Mar 4, 2023

Alan Jo

Jul 30, 2023

### Variable Length of inputs and outputs 인코더-디코더 구조는 주로 입력 문장과 출력 문장의 길이가 다를 경우에 사용 - Encoder takes the input sequence and converts it into a fixed length vector representation - Decoder uses this vector to generate the output sequence ### Seq2Seq Notion |Title| |:-:| |[Attention Mechanism](https://texonom.com/attention-mechanism-762711860abb45f59904f1ac4e4af285)| |[Copy mechanism](https://texonom.com/copy-mechanism-b5dee9c80ca24bb993b5f152129b3577)| ### Seq2Seq Models |Title| |:-:| |[Decoder Model](https://texonom.com/decoder-model-36e78a40265c473d90197089aebfa83b)| |[Transformer Model](https://texonom.com/transformer-model-f3e8053cc5b447a2bc9c6b5d0874dafc)| |[Encoder Model](https://texonom.com/encoder-model-321d79943c8940fcaac0c9ccca0b5f6f)| > [14-01 시퀀스-투-시퀀스(Sequence-to-Sequence, seq2seq)](https://wikidocs.net/24996)

01a9854dffa6417c87d92c11a607250c

Discriminator Network

GAN Notion

Nov 18, 2019

Alan Jo

Jun 1, 2023

tries to distinguish between real and fake images Discriminator $\phi$ aims at maximizing the objective - $D(x)$ to be close to 1 for real - $D(G(z))$ to be close to 0 for fake

a02287522a264d779574871285eed3b6

GAN Issues

GAN Notion

Jun 1, 2023

Alan Jo

Jun 14, 2023

### 1. Non-convergence minmax objective **cycle **generated so repeated. Even with a small learning rate, it will not converge example. $min_xmax_yV(x, y ) = xy$ ### 2. Mode-Collapse What if the generator keeps generating a **single realistic image**? The discriminator will be always fooled by the single sample ### Mini Batch Trick **Compute the similarity** of the image $x$ with other images in the same batch to avoid Mode-Collapse. **This measures the diversity of the batch.** **Feed the similarity score** along with the image to the discriminator as an input feature. This penalizes the generator and encourages it to generate less similar images Many more advanced techniques have also been proposed so far.

be6312c3a8184b55ba443167164101ba

GAN Minmax Game

GAN Notion

Jun 1, 2023

Alan Jo

Jun 1, 2023

## Training GAN Aims to $D(G(z))$ to be close to 1, i.e. discriminator is fooled ### Gradient descent on generator ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F19a41272-13ca-4ad9-b94a-c3fe8dbd031f%2FUntitled.png?table=block&id=17e51d57-7290-4e22-8703-8a64cf6e1b95&cache=v2) If $G$ is very bad compared to $D$, then we have almost zero gradient. Hence, $-D$ term can be used ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F5c0d0cce-916a-44e9-a971-c7201830e7db%2FUntitled.png?table=block&id=5841e261-a075-413d-860c-0186cb106361&cache=v2) ### Gradient ascent on discriminator Generator $\theta$ and Discriminator $\phi$ has very different objective so hard to train stable

d3b08fa32fc34cf5920bfc0e80c34b90

Generator Network

GAN Notion

Nov 18, 2019

Alan Jo

Jun 1, 2023

tries to fool the discriminator by generating realistic sample Generator $\theta $ aims at minimizing the objective

3d3fb237d8f149979c7ed172aca65529

3D GAN

GANs

Feb 24, 2022

Alan Jo

Jun 1, 2023

> [3D Generative Adversarial Network](http://3dgan.csail.mit.edu/)

672af0e2639041e484ed1a26b56f84cb

DCGAN

GANs

Mar 5, 2023

Alan Jo

Jun 1, 2023

[DCGAN-tensorflow](https://github.com/carpedm20/DCGAN-tensorflow)

ae93b511fed5402c926e818243d8a966

DragGan

GANs

May 29, 2023

Alan Jo

Jun 1, 2023

[DragGAN](https://github.com/XingangPan/DragGAN)

### Unofficial Huggingface implementation > [DragGAN - a Hugging Face Space by fffiloni](https://huggingface.co/spaces/fffiloni/DragGAN) > [Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold](https://vcai.mpi-inf.mpg.de/projects/DragGAN/)

bd335de6767243e7b7ec48ff88dea800

Differentiable Neural Computer

MANN Notion

May 29, 2023

Alan Jo

May 29, 2023

2c182410c5c34222b41605b63c37c777

Neural Turing Machine

MANN Notion

May 29, 2023

Alan Jo

May 29, 2023

c03efee39e1942f197f4b3d6553e4ac1

Decoder Model

Seq2Seq Models

Mar 6, 2023

Alan Jo

Jul 17, 2023

[Encoder Model](https://texonom.com/encoder-model-321d79943c8940fcaac0c9ccca0b5f6f) [Text Generation](https://texonom.com/text-generation-cb0e475216c043fdbb4f309d324c9d19)

## **Autoregressive Model, **Causal language model 이전 시점의 출력을 현재 시점의 입력으로 사용하여, 시퀀스 데이터를 생성하는 모델 단방향 모델을 통해 다음 단어를 예측, 추측하는 언어모델 [Decoder Input IDs](https://texonom.com/decoder-input-ids-2b1e1d3c7c274b0880191a54290bed27) ### Autoregressive 현재 값이 이전 값들의 선형 조합으로 나타나는 모델을 의미 Transformer의 Decoder만 사용하므로 이전 시점의 출력을 현재 시점의 입력으로 사용하여, 시퀀스 데이터를 생성하는 방식을 사용 이 방식은 문장 전체를 한 번에 처리하지 않고, 단어나 토큰 단위로 처리 이전 단어들을 기반으로 다음 단어를 예측 즉 문장 전체를 한 번에 처리하지 않기 때문에, 문장 전체를 볼 수 없다 > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/) > [Decoder models - Hugging Face NLP Course](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt)

36e78a40265c473d90197089aebfa83b

Encoder Model

Seq2Seq Models

Mar 6, 2023

Alan Jo

Jul 30, 2023

[Decoder Model](https://texonom.com/decoder-model-36e78a40265c473d90197089aebfa83b)

## Auto*-encoding model, Masked Language Model* transform text or images into a condensed numerical representation called an embedding ### These models are often characterized as having bi-directional attention 입력 문장을 벡터화하고 다시 복원하는 모델 입력 문장의 의미를 보존하면서, 모델이 처리하기 쉬운 형태로 변환 The pretraining of these models usually revolves around somehow corrupting a given sentence and tasking the model with finding or reconstructing the initial sentence ### *auto-encoding* 입력 시퀀스를 인코딩하여 전부 [Latent Space](https://texonom.com/latent-space-d67b6bdef18b4058bfbc3d25f87ec087) 에 매핑한다 입력 시퀀스를 토큰 단위로 처리하지만 Self-Attention으로 입력 시퀀스의 각 위치마다 해당 위치와 다른 모든 위치 간의 유사도를 계산하기 때문에 문맥고려 가능하다. 이때 유사도가 높은 위치들은 해당 위치의 임베딩 벡터에 더 많은 가중치가 부여 > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/) > [Encoder models - Hugging Face NLP Course](https://huggingface.co/learn/nlp-course/chapter1/5?fw=pt)

321d79943c8940fcaac0c9ccca0b5f6f

Transformer Model

Seq2Seq Models

Aug 17, 2020

Alan Jo

May 29, 2023

[Attention Mechanism](https://texonom.com/attention-mechanism-762711860abb45f59904f1ac4e4af285) [Attention is all you need](https://texonom.com/attention-is-all-you-need-f52e664318eb47c0aa6cbd27b9a4c491)

## Self Attention is the core feature the Transformer gains a wider perspective and can attend to multiple interaction levels within the input sentence. 어텐션을 Encoder 보정을 위한 용도로서 사용하는 것이 아니라 어텐션만으로 인코더와 디코더 생성 모든 토큰을 동시에 받아 연산하기 때문에 병렬연산이 가능 인코더 블럭과 디코더 블럭이 각각 6개씩 모여있는 구조 인코더 블록은 2개의 sub-layer(Multi-Head(self) Attention, Feed Forward)로 나눌 수 있고 디코더 블록은 3개의 sub-layer(Masked Multi-Head(self) Attention, Multi-Head (Encoder-Decoder) Attention, Feed Forward)로 나눌 수 있다 기존의 Attention과는 다르게 각 벡터가 모두 가중치 벡터라는 점 ### Transformer Model Notion |Title| |:-:| |[Attention is all you need](https://texonom.com/attention-is-all-you-need-f52e664318eb47c0aa6cbd27b9a4c491)| |[NLP Transformer Encoder](https://texonom.com/nlp-transformer-encoder-d5bc8cebe43d4fc7896d8e0290683b18)| |[NLP Transformer Decoder](https://texonom.com/nlp-transformer-decoder-1ee79f4754d24aa8b64312399782cbec)| |[Transformer Attention](https://texonom.com/transformer-attention-84d340d69722409a96ec1d806970608a)| |[Transformer Model Tool ](https://texonom.com/transformer-model-tool-01c749c1e9254d8c85ec5ef2feb0566b)| ### Transformer Models |Title| |:-:| |[RETRO Transformer](https://texonom.com/retro-transformer-e3efdb06419948ebb1be444f4e124867)| |[Torchscale](https://texonom.com/torchscale-c4b82c1245d74c7092617e70b9ddab51)| |[BART](https://texonom.com/bart-48bd2258eee74af3b2d929b03bb9553b)| |[RMT](https://texonom.com/rmt-f076841449f2419a82b3ce09281b9bb9)| |[Transformer-XL](https://texonom.com/transformer-xl-eb2128e01dc744539b6825f3880da761)| ### Architecture > [Transformer’s Encoder-Decoder: Let’s Understand The Model Architecture - KiKaBeN](https://kikaben.com/transformers-encoder-decoder/) ### Pseudo Source code > [Transformers for software engineers](https://blog.nelhage.com/post/transformers-for-software-engineers/) ### Korean > [점프 투 파이썬](https://wikidocs.net/31379) > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리)

f3e8053cc5b447a2bc9c6b5d0874dafc

Decoder Input IDs

Decoder Model

null

## token indices

2b1e1d3c7c274b0880191a54290bed27

Attention is all you need

Transformer Model Notion

Aug 23, 2020

Alan Jo

May 30, 2023

[Transformer Model](https://texonom.com/transformer-model-f3e8053cc5b447a2bc9c6b5d0874dafc)

2017년 발표된 트랜스포머 구조를 처음 발표한 논문 해당 논문은 Non-recurrent sequence to sequence encoder-decoder model을 만드는 것이 목표 Replaced [RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16) Encoder Decoder Model ## **Background** ### **1. Sequential computation** sequence to sequence한 문제를 푸는 과정에서, Encoder-Decoder 구조의 RNN 모델들이 좋은 성능을 냈다. ### **2. Long term dependency** RNN의 경우, Long term dependency의 문제가 항상 따라다니고, CNN의 경우 kernel 안에서 O(1)이나, kernel 간 정보가 공유되지 않는다. # Model Architecture 6개의 stack - 하나의 인코더는 Self-Attention Layer와 Feed Forward Neural Network(2개의 Sub-layer) - Encoder - Multi-Head Attention - Positional Encoding - Relative Positioning - The Residuals - Decoder ### Author > [ashVaswani](https://twitter.com/ashVaswani) ### pdf > [Untitled](https://arxiv.org/pdf/1706.03762.pdf) > [Attention Is All You Need(transformer) paper 정리](https://medium.com/@omicro03/attention-is-all-you-need-transformer-paper-%EC%A0%95%EB%A6%AC-83066192d9ab) > [점프 투 파이썬](https://wikidocs.net/31379)

f52e664318eb47c0aa6cbd27b9a4c491

NLP Transformer Decoder

Transformer Model Notion

Mar 7, 2023

Alan Jo

Mar 7, 2023

### 3 Sub Layer [Masked Self-Attention](https://texonom.com/masked-self-attention-cb0e29589c93423780cc0ca60260f3e4) + [Multi-head Attention](https://texonom.com/multi-head-attention-d9dfc39b27494123ae4c81f3b98e50b5) + [Position-wise FFNN](https://texonom.com/position-wise-ffnn-6fe30c96aa9245d898c73ec34625377d) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fa99c4277-03ed-4ec6-95ff-c4e91a06f844%2FUntitled.png?table=block&id=00eb4408-f746-4d0e-8313-6f5a8e6b94e1&cache=v2)

1ee79f4754d24aa8b64312399782cbec

NLP Transformer Encoder

Transformer Model Notion

Mar 7, 2023

Alan Jo

Mar 7, 2023

### 2 Sub Layer [Multi-head Attention](https://texonom.com/multi-head-attention-d9dfc39b27494123ae4c81f3b98e50b5) + [Position-wise FFNN](https://texonom.com/position-wise-ffnn-6fe30c96aa9245d898c73ec34625377d) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fa99c4277-03ed-4ec6-95ff-c4e91a06f844%2FUntitled.png?table=block&id=d48db9af-a165-4db5-842f-80d8046d66c6&cache=v2)

d5bc8cebe43d4fc7896d8e0290683b18

Transformer Attention

Transformer Model Notion

Apr 3, 2023

Alan Jo

Apr 3, 2023

### Transformer Attentions |Title| |:-:| |[Encoder-Decoder Attention](https://texonom.com/encoder-decoder-attention-4b3bb5b4aeac4681bf6bdca66ea79e04)| |[Encoder Self-Attention](https://texonom.com/encoder-self-attention-c1e03ab0c57f44c395f2056e201f4326)| |[Masked Self-Attention](https://texonom.com/masked-self-attention-cb0e29589c93423780cc0ca60260f3e4)|

84d340d69722409a96ec1d806970608a

Transformer Model Tool

Transformer Model Notion

May 22, 2023

Alan Jo

May 22, 2023

[Transformers.js](https://texonom.com/transformersjs-3e7f8bb1d88940298936f13ee1ce7ed7)

01c749c1e9254d8c85ec5ef2feb0566b

Encoder-Decoder Attention

Transformer Attentions

Mar 5, 2023

Alan Jo

Apr 3, 2023

input과 output의 정보를 엮어주는 역할

4b3bb5b4aeac4681bf6bdca66ea79e04

Encoder Self-Attention

Transformer Attentions

Mar 5, 2023

Alan Jo

Apr 3, 2023

c1e03ab0c57f44c395f2056e201f4326

Masked Self-Attention

Transformer Attentions

Mar 7, 2023

Alan Jo

Apr 3, 2023

디코더 블럭에서 사용되는 특수한 Self-Attention 디코더는 [Autoregressive](https://texonom.com/autoregressive-93cf5710b4b54730a7e7efcc6e0fc642) 하기 때문에 이후단어 보지않고 예측해야 그래서 뒤에 보지 않도록 Masking한다 > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리)

cb0e29589c93423780cc0ca60260f3e4

[trl](https://github.com/lvwerra/trl)

Transformer Model Tools

May 22, 2023

Alan Jo

May 22, 2023

3ddce975d3aa4f6dbca6e3d4b3eb6e6e

BART

Transformer Models

Mar 25, 2023

Alan Jo

May 31, 2023

[AI Summarization](https://texonom.com/ai-summarization-a314a6fb3162447086b8d2526ae8ef16) [KoBART-summarization](https://github.com/seujung/KoBART-summarization) [BERT](https://texonom.com/bert-e282abe8a34543988a0e71f6c8701ad2)

****Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension**** 2019 > [BART: Denoising Sequence-to-Sequence Pre-training for Natural...](https://arxiv.org/abs/1910.13461) > [BART Text Summarization vs. GPT-3 vs. BERT: An In-Depth Comparison | Width.ai](https://www.width.ai/post/bart-text-summarization) > [BART 논문 리뷰](https://dladustn95.github.io/nlp/BART_paper_review/)

48bd2258eee74af3b2d929b03bb9553b

RETRO Transformer

Transformer Models

Jul 5, 2022

Alan Jo

May 22, 2023

[Vector Database](https://texonom.com/vector-database-5dfdb6e2bc294fed8ae80eaea2ee5c26) [Deepmind](https://texonom.com/deepmind-5eb171c77b344d4786a9a5b23ae70eca)

### Retrieval-Enhanced ### Fast 밖에다 retrieval database를 두는 형태 ### Implementations [RETRO-pytorch](https://github.com/lucidrains/RETRO-pytorch) > [Improving language models by retrieving from trillions of tokens](https://www.deepmind.com/publications/improving-language-models-by-retrieving-from-trillions-of-tokens) > [The Illustrated Retrieval Transformer](https://jalammar.github.io/illustrated-retrieval-transformer/) > [RETRO Is Blazingly Fast](http://mitchgordon.me/ml/2022/07/01/retro-is-blazing.html) ### Korean > [RETRO: Improving language models by retrieving from trillions of tokens](https://velog.io/@nawnoes/RETRO-Improving-language-models-by-retrieving-from-trillions-of-tokens)

e3efdb06419948ebb1be444f4e124867

RMT

Transformer Models

May 3, 2023

Alan Jo

May 3, 2023

[LM-RMT](https://github.com/booydar/LM-RMT) [RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16)

## **Recurrent Memory Transformer** GPT-4’s maximum input token for inference is 32000204 This model can 2 million > [Recurrent Memory Transformer](https://arxiv.org/abs/2207.06881) > [“모든 것 바꿀지도”... 기억력 GPT-4 63배 ‘RMT’ 기반 AI 등장](https://contents.premium.naver.com/themiilk/business/contents/230426095632265ym)

f076841449f2419a82b3ce09281b9bb9

Torchscale

Transformer Models

Dec 2, 2022

Alan Jo

May 22, 2023

[Pytorch](https://texonom.com/pytorch-2dd232d99b3a46d5b7d1e4e686070686) [torchscale](https://github.com/microsoft/torchscale)

c4b82c1245d74c7092617e70b9ddab51

**Transformer-XL**

Transformer Models

May 3, 2023

Alan Jo

May 3, 2023

[transformer-xl](https://github.com/kimiyoung/transformer-xl)

**Attentive Language Models Beyond a Fixed-Length Context** > [Recurrent Memory Transformer](https://arxiv.org/abs/2207.06881)

eb2128e01dc744539b6825f3880da761

Attention Mechanism

Seq2Seq Notion

Mar 5, 2023

Alan Jo

Jul 28, 2023

[Encoder Model](https://texonom.com/encoder-model-321d79943c8940fcaac0c9ccca0b5f6f) [Decoder Model](https://texonom.com/decoder-model-36e78a40265c473d90197089aebfa83b) [RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16)

**한 번에 전체 문장을 읽고 문장을 기반으로 각 단어의 표현을 병렬로 계산가능** Imagine yourself in a library. You have a specific question (**query**). Books on the shelves have titles on their spines (**keys**) that suggest their content. You compare your question to these titles to decide how relevant each book is, and how much **attention** to give each book. Then, you get the information (**value**) from the relevant books to answer your question. the query vector points to the current input word (aka context). **The *****keys***** represent the words in the input sentence. The key vectors help the model understand how each word relates to the context word.** **Attention is how much weight the query word should give each word in the sentence. This is computed via a dot product between the query vector and all the key vectors**. These dot products then go through a softmax which makes the attention scores (across all keys) sum to 1. **Each word is also represented by a *****value***** which contains the information of that word. As a result, each context word is now represented by an attention-based weightage of all the words in the sentence.** ### NLP Attention Notion |Title| |:-:| |[Self-Attention](https://texonom.com/self-attention-d06bde44563f455b951d17955d820f77)| |[Multi-head Attention](https://texonom.com/multi-head-attention-d9dfc39b27494123ae4c81f3b98e50b5)| |[Cross-Attention](https://texonom.com/cross-attention-6365d422d2bf4be78dac038df3a19aae)| ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F36626d64-ac93-430d-aa3b-11ead7cb27e3%2FUntitled.png?table=block&id=6ce231d7-c03f-4556-b56f-0d2d3032f7a3&cache=v2) ### NLP Attention Usages |Title| |:-:| |[Flash Attention](https://texonom.com/flash-attention-ca8093deb6c648059aff1580b0ddbf68)| |[Dilated Attention](https://texonom.com/dilated-attention-d9000e3460f547bca3f15ac9b8d7e36c)| |[Multi Query Attention](https://texonom.com/multi-query-attention-5641aba38a8b47caa3a9f364c13789f1)| |[Group Query Attentiion](https://texonom.com/group-query-attentiion-d72d2da59fba4e9d937a4ec856dac90f)| |[PagedAttention](https://texonom.com/pagedattention-abf197357343437681fb878ae5700926)| > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/) > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리) > [16-01 트랜스포머(Transformer)](https://wikidocs.net/31379)

762711860abb45f59904f1ac4e4af285

***Copy mechanism***

Seq2Seq Notion

Mar 8, 2023

Alan Jo

May 29, 2023

디코딩 과정에서 문장을 생성할 때 필요한 어휘가 출력 사전(output vocabulary)에 없는 문제(Out-of-Vocabulary)와고유명사들의 출력 확률이 작아지는 문제를 해결하기 위해 고안된 방법으로, 출력에 필요한 어휘를 입력 열에서찾아 출력 없이 복사(copy)하는 사용하는 방법 > [Untitled](https://koreascience.kr/article/CFKO201612470014629.pdf)

b5dee9c80ca24bb993b5f152129b3577

Cross-Attention

NLP Attention Notion

Apr 9, 2023

Alan Jo

Jul 28, 2023

[MLLM](https://texonom.com/mllm-98172e9446c04bcc9cf52b2bc5d0bd17)

## Encoder-Decoder Attention in [Decoder Model](https://texonom.com/decoder-model-36e78a40265c473d90197089aebfa83b) Query : 디코더 벡터 / **Key **= Value : 인코더 벡터 인코더의 출력과 디코더의 현재 상태를 이용하여 디코더가 다음 단어를 예측하는 데 사용 MLLM에서 매우 중요한 부분 이미지나 비디오 데이터와 텍스트 데이터 간의 상호작용을 모델링 $$h_i = softmax((x_i * K_v) / sqrt(d_k)) * V_v$$

6365d422d2bf4be78dac038df3a19aae

****Multi-head Attention****

NLP Attention Notion

Mar 5, 2023

Alan Jo

Jul 28, 2023

### **Multiple heads lets the model consider multiple words simultaneously.** Because we use the softmax function in attention, it amplifies the highest value while squashing the lower ones. As a result, each head tends to focus on a single element. Multiple heads let us attend to several words. **It also provides redundancy**, where if any single head fails, we have the other attention heads to rely on. 입력 벡터를 여러 개의 헤드로 분할하여 각각 어텐션을 수행하고 결과를 결합하는 방식으로 동작 > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/)

d9dfc39b27494123ae4c81f3b98e50b5

Self-Attention

NLP Attention Notion

Mar 5, 2023

Alan Jo

Jul 28, 2023

### 텍스트 데이터 내에서 단어 간의 상호작용을 모델링 Self-attention enables the encoder to weigh the importance of each word and capture both short and long-range dependencies. Self-attention enables the decoder to focus on different parts of the output generated so far. - 인코더의 셀프 어텐션 : Query = **Key **= Value - 디코더의 마스크드 셀프 어텐션 : Query = **Key **= Value self-attention 은 어텐션을 자기 자신에게 수행한다는 의미 문장 내부 요소의 관계를 잘 파악하기 위해서 문장 자신에 대해 어텐션 매커니즘을 적용 - 쿼리 : 분석하고 하는 단어에 대한 가중치 벡터 - 키 : 각 단어가 쿼리에 해당하는 단어와 얼마나 연관있는 지를 비교하기 위한 가중치 벡터 - 밸류 : 각 단어의 의미를 살려주기 위한 가중치 벡터 ### self Attention 과정 1. 특정 단어의 쿼리(q) 벡터와 모든 단어의 키(k) 벡터를 내적한다. 내적해서 나온 값은 Attention Score가 된다. 2. 보정으로 트랜스포머에서는 이 가중치를 q,k,v 벡터 차원 *dk* 의 제곱근인 *d**k*로 나누어 준다. 3. Softmax로 쿼리에 해당하는 단어와 문장 내 다른 단어가 가지는 관계의 비율 계산 4. Value 각 단어의 벡터를 곱해준 후 모두 더한다. > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리)

d06bde44563f455b951d17955d820f77

Dilated Attention

NLP Attention Usages

Jul 13, 2023

Alan Jo

Jul 28, 2023

![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fe84a0f0b-ccda-469a-978d-c0ef60558a4b%2FUntitled.png?table=block&id=6014ac5c-6ab5-4d96-affe-7f504b2c4355&cache=v2) > [Microsoft’s LongNet Scales Transformer to One Billion Tokens](https://medium.com/syncedreview/microsofts-longnet-scales-transformer-to-one-billion-tokens-af02ff657d87)

d9000e3460f547bca3f15ac9b8d7e36c

Flash Attention

NLP Attention Usages

Jun 29, 2023

Alan Jo

Jul 28, 2023

[flash-attention](https://github.com/HazyResearch/flash-attention)

**Tiling**과 **Recomputation**을 사용하여 Attention을 가속화 softmax를 block단위를 쪼개서 계산 CUDA kernel를 사용 > [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness | Wonbeom Jang](https://www.wonbeomjang.kr/blog/2023/fastattention/) > [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness | Wonbeom Jang](https://www.wonbeomjang.kr/blog/2023/fastattention/)

ca8093deb6c648059aff1580b0ddbf68

Group Query Attentiion

NLP Attention Usages

Aug 2, 2023

Alan Jo

Aug 2, 2023

d72d2da59fba4e9d937a4ec856dac90f

Multi Query Attention

NLP Attention Usages

Aug 2, 2023

Alan Jo

Aug 2, 2023

5641aba38a8b47caa3a9f364c13789f1

**PagedAttention**

NLP Attention Usages

Aug 3, 2023

Alan Jo

Aug 3, 2023

Efficient management of attention key and value memory

abf197357343437681fb878ae5700926

Deep Learning Compiler

Deep Learning Usages

Mar 22, 2022

Alan Jo

Mar 22, 2022

7d79af9683764b6b983793c1856578c6

Deep Learning Tool

Deep Learning Usages

Oct 6, 2021

Alan Jo

May 29, 2023

[tuning_playbook](https://github.com/google-research/tuning_playbook)

a14ea6f4574342ef974443634e27c6ce

Learn Deep Learning

Deep Learning Usages

Jun 6, 2023

Alan Jo

Jun 6, 2023

> [0020 DL Terms & Concepts - Deepest Documentation](https://deepestdocs.readthedocs.io/en/latest/002_deep_learning_part_1/0020/)

4382083fb54d4705984e7f45a9af2d86

Sentiment Neuron

Deep Learning Usages

May 20, 2023

Alan Jo

May 20, 2023

### Proof of compressed data in deep network before gpt1 ### Why ImportantWhy Important > [일리야 수츠케버와 AGI의 미싱링크](https://www.youtube.com/watch?v=LQviQS24uQY&t=840) > [Unsupervised sentiment neuron](https://openai.com/research/unsupervised-sentiment-neuron) > [Sentiment Neuron](https://tensorflow.blog/2017/04/07/sentiment-neuron/)

2e44e9754b894534af6c121b6d6074d6

Nebullvm

Deep Learning Compiler Tools

Mar 22, 2022

Alan Jo

Mar 22, 2022

[nebuly](https://github.com/nebuly-ai/nebullvm)

d967865085f84f5ab2b1a4a5ea661dfd

HuggingFace

Deep Learning Hubs

Jun 27, 2022

Alan Jo

Jul 17, 2023

[AI Industry](https://texonom.com/ai-industry-d8709bd0498145e3a66af6da3f963fa7) [Pytorch](https://texonom.com/pytorch-2dd232d99b3a46d5b7d1e4e686070686)

### [Pytorch Hub](https://texonom.com/pytorch-hub-d46f5b09e91241578a0c63b4847396fb) + [Lightening AI](https://texonom.com/lightening-ai-af651491f5474031bc51e6d1a99d5f22) ### HuggingFace Usages |Title| |:-:| |[HuggingFace Hub](https://texonom.com/huggingface-hub-9b16b4841f2d4bdd9827120c47677b29)| |[Huggingface Model](https://texonom.com/huggingface-model-f985be46b10e462f91e65576b81f452f)| |[Huggingface Dataset](https://texonom.com/huggingface-dataset-43582d2257474a83874ecec2d4c6ab44)| |[HuggingFace Space](https://texonom.com/huggingface-space-e37f59c218ca48cebe509d3ab2381b34)| ### LLM Leaderboard > [Open LLM Leaderboard - a Hugging Face Space by HuggingFaceH4](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) > [[D] HuggingFace ecosystem vs. Pytorch Lightning for big research NLP project with many collaborators.](https://www.reddit.com/r/MachineLearning/comments/si08qt/d_huggingface_ecosystem_vs_pytorch_lightning_for/)

eb25c513b432477e9da51ca19bb06833

Pytorch Hub

Deep Learning Hubs

May 29, 2023

Alan Jo

May 29, 2023

[HuggingFace](https://texonom.com/huggingface-eb25c513b432477e9da51ca19bb06833) [Pytorch](https://texonom.com/pytorch-2dd232d99b3a46d5b7d1e4e686070686)

### Pytorch Hub Usages |Title| |:-:|

d46f5b09e91241578a0c63b4847396fb

Huggingface Dataset

HuggingFace Usages

May 23, 2023

Alan Jo

Jul 19, 2023

### Huggingface Dataset Usages |Title| |:-:| |[ HuggingFace Datasets Jax](https://texonom.com/huggingface-datasets-jax-a92d62f342de4ddc83820623059ca02e)| > [Create a dataset](https://huggingface.co/docs/datasets/create_dataset)

43582d2257474a83874ecec2d4c6ab44

HuggingFace Hub

HuggingFace Usages

Jun 19, 2023

Alan Jo

Jul 17, 2023

[huggingface_hub](https://github.com/huggingface/huggingface_hub) > [Quickstart](https://huggingface.co/docs/huggingface_hub/quick-start)

9b16b4841f2d4bdd9827120c47677b29

Huggingface Model

HuggingFace Usages

May 23, 2023

Alan Jo

Jun 29, 2023

### Huggingface Model Usages |Title| |:-:| |[Huggingface Provider](https://texonom.com/huggingface-provider-8bd64d5951774d6f9ed623abbe471b4c)| |[Huggingface H4](https://texonom.com/huggingface-h4-9b33bfe491704142929a794edd95a7df)| |[Huggingface Model Card](https://texonom.com/huggingface-model-card-39d3f0d8805c4cb0939b623b8bac64ea)| > [Models](https://huggingface.co/docs/hub/models)

f985be46b10e462f91e65576b81f452f

HuggingFace Space

HuggingFace Usages

Mar 26, 2023

Alan Jo

Jul 13, 2023

[Gradio](https://texonom.com/gradio-7e1647e9bc174161b4d7e8f44dd53707) [Streamlit](https://texonom.com/streamlit-9e295c64d27e4999878a022b1c538964)

e37f59c218ca48cebe509d3ab2381b34

HuggingFace Datasets Jax

Huggingface Dataset Usages

May 30, 2023

Alan Jo

Jul 17, 2023

> [Use with JAX](https://huggingface.co/docs/datasets/use_with_jax)

a92d62f342de4ddc83820623059ca02e

Huggingface H4

Huggingface Model Usages

Jul 9, 2023

Alan Jo

Jul 9, 2023

### helpful, honest, harmless, and huggy [StarChat](https://texonom.com/starchat-ea316ec509564a51b56ad92b22831220)

9b33bfe491704142929a794edd95a7df

Huggingface Model Card

Huggingface Model Usages

Jul 19, 2023

Alan Jo

Jul 30, 2023

### Disable API ```type inference: false ``` - tags - pipeline tags - etc ### Widgets > [Widgets](https://huggingface.co/docs/hub/models-widgets) > [Disable Hosted inference API](https://discuss.huggingface.co/t/disable-hosted-inference-api/10379)

39d3f0d8805c4cb0939b623b8bac64ea

Huggingface Provider

Huggingface Model Usages

Jun 29, 2023

Alan Jo

Aug 5, 2023

### Companies > [amazon (Amazon Web Services)](https://huggingface.co/amazon) > [stabilityai (Stability AI)](https://huggingface.co/stabilityai) > [EleutherAI (EleutherAI)](https://huggingface.co/EleutherAI) > [allenai (Allen Institute for AI)](https://huggingface.co/allenai) ### Model Users > [ehartford (Eric Hartford)](https://huggingface.co/ehartford) > [psmathur (Pankaj Mathur)](https://huggingface.co/psmathur) > [TheBloke (Tom Jobbins)](https://huggingface.co/TheBloke) > [jncraton (Jon)](https://huggingface.co/jncraton) > [bhenrym14 (Brandon)](https://huggingface.co/bhenrym14) ### Organization > [decapoda-research (Decapoda Research)](https://huggingface.co/decapoda-research) > [openchat (OpenChat)](https://huggingface.co/openchat) > [MBZUAI (Mohamed Bin Zayed University of Artificial Intelligence)](https://huggingface.co/MBZUAI) > [openchat/openchat · Hugging Face](https://huggingface.co/openchat/openchat) > [OpenAssistant (OpenAssistant)](https://huggingface.co/OpenAssistant)

8bd64d5951774d6f9ed623abbe471b4c