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
Streamlit Caching	Streamlit Usages	Jul 15, 2023	Alan Jo	Alan Jo	Jul 15, 2023		- `st.cache_data` - `st_cache_resource` > [Caching - Streamlit Docs](https://docs.streamlit.io/library/advanced-features/caching)	a22ccb53e76349dc9ccf31b616d84083
Streamlit Cloud	Streamlit Usages	May 24, 2023	Alan Jo	Alan Jo	May 24, 2023		> [Embed your app - Streamlit Docs](https://docs.streamlit.io/streamlit-community-cloud/get-started/embed-your-app)	8c045600c4314f6382bd3c05358dd2ba
Streamlit UI	Streamlit Usages	May 14, 2023	Alan Jo	Alan Jo	May 14, 2023		### Streamlit UIs \|Title\| \|:-:\| \|[Streamlit Chat](https://texonom.com/streamlit-chat-73c8557f0d094b3fac7448e10f013cb4)\| \|[Streamlit Extras](https://texonom.com/streamlit-extras-02c8d0d7eebf4c03adce4e843305e2e7)\| \|[Streamlit Pills](https://texonom.com/streamlit-pills-cc0ce4c16ac44d1798242fc990f7eb92)\|	4d1410d0bdb54e1ba4512e03f500b0b7
Streamlit Widget	Streamlit Usages	Jul 17, 2023	Alan Jo	Alan Jo	Jul 17, 2023		### Streamlit Widgets \|Title\| \|:-:\| \|[Streamlit text_area](https://texonom.com/streamlit-textarea-d58bc341994a4cf9a6eec5095e6c5395)\|	c4cd30e860ad4f8487ef2f4fcae564e6
Streamlit Chat	Streamlit UIs	May 14, 2023	Alan Jo	Alan Jo	May 14, 2023			73c8557f0d094b3fac7448e10f013cb4
Streamlit Extras	Streamlit UIs	May 14, 2023	Alan Jo	Alan Jo	May 14, 2023			02c8d0d7eebf4c03adce4e843305e2e7
Streamlit Pills	Streamlit UIs	May 14, 2023	Alan Jo	Alan Jo	May 14, 2023			cc0ce4c16ac44d1798242fc990f7eb92
Streamlit text_area	Streamlit Widgets	Jul 17, 2023	Alan Jo	Alan Jo	Jul 17, 2023		> [st.text_area - Streamlit Docs](https://docs.streamlit.io/library/api-reference/widgets/st.text_area)	d58bc341994a4cf9a6eec5095e6c5395
Hold-out Method	AI Generalization Methods	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023		Given data is randomly partitioned into two independent sets which make them same distribution our assumption to be valid	cc7504f8f8db489fb474bd89cbbe0dd4
k-fold cross validation	AI Generalization Methods	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023		k mutually-exclusive subset	f61ca1dbd3a5483998e9d6514b16c48d
Nested cross validation	AI Generalization Methods	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023	[Hyperparameter](https://texonom.com/hyperparameter-ef7e34566add4e98b673d4cef59fca90)	- inner fold - outer fold a lot of times needed if dataset is large	c88ec0073c1d4dbda857ad019779559a
Random Sampling	AI Generalization Methods	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023		a variation of hold-out	5879da4187bb4e1bafacedd6fe617149
Train/Validation/Test splitting	AI Generalization Methods	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023		If dataset is large enough	bc5f600b344b455b994ca25671863ac8
Fine Tuning	Model Generalization Notion	Mar 7, 2023	Alan Jo	Alan Jo	Jun 22, 2023		### Task optimized ### Fine Tuning Notion \|Title\| \|:-:\| \|[PEFT](https://texonom.com/peft-f47a178d75804abf928ec7e7be2da27f)\| \|[TRL](https://texonom.com/trl-ebc3e432e3984ca3b2a1cf20da0fa5d1)\| \|[DRO](https://texonom.com/dro-4701102d6c0a4545afe7f0bad178180b)\| \|[SFT](https://texonom.com/sft-f48f7e6eccd54a62bea82725fae98865)\|	52215d8477ad46e3896a29e2fa408991
Generalization Gap	Model Generalization Notion	May 9, 2023	Alan Jo	Alan Jo	May 11, 2023		the test error is not necessarily always close to the training error	d8b4ec347e7d4df282ea96a093e7581f
Overfitting	Model Generalization Notion	May 31, 2021	Alan Jo	Alan Jo	Jun 7, 2023	[Underfitting](https://texonom.com/underfitting-e56067d23ba8404a8e5165085986c9aa)	### Predicts training dataset > test, High [Variance](https://texonom.com/variance-08c1eccc7dc84957afbb815ad6b41280) large number of parameters can cause overfitting Models that are bigger or have more capacity are more likely to overfit Overfitting issues are usually observed when the magnitude of parameters is large If data is general enough, overfitting is okay ### Resolve Overfitting \|Title\| \|:-:\| \|[Non-parametric algorithm](https://texonom.com/non-parametric-algorithm-bc41f745a14c4837bd54368652da5982)\| \|[Regularized parameter](https://texonom.com/regularized-parameter-f3b208cdd37a4002a5d39ef990b8be33)\| ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F7d045ebf-3253-4b1c-8305-57cb4088e985%2FUntitled.png?table=block&id=d0611322-665c-4cab-9e56-53d84fdf4f2a&cache=v2)	24c3b183372845e8999ad7f7a0ba5035
Pre Training	Model Generalization Notion	Mar 7, 2023	Alan Jo	Alan Jo	May 11, 2023		일반적인 목적을 위한 트레이닝	0d81c286e86f4dcba940d3c849631b35
Underfitting	Model Generalization Notion	Mar 14, 2023	Alan Jo	Alan Jo	May 14, 2023	[Overfitting](https://texonom.com/overfitting-24c3b183372845e8999ad7f7a0ba5035)	### The training error is relatively large, High [Bias](https://texonom.com/bias-ba063cd622a54deb8a677e8fb87dfdc8)	e56067d23ba8404a8e5165085986c9aa
DRO	Fine Tuning Notion	Jun 29, 2023	Alan Jo	Alan Jo	Jun 29, 2023		### DRO Usages \|Title\| \|:-:\| \|[Group DRO](https://texonom.com/group-dro-4877ea63979c44ed8149515d3b832397)\| \|[DRO-LM](https://texonom.com/dro-lm-564725ae9e7140de9e42d25310728af4)\|	4701102d6c0a4545afe7f0bad178180b
PEFT	Fine Tuning Notion	Mar 7, 2023	Alan Jo	Alan Jo	Jul 15, 2023	[peft](https://github.com/huggingface/peft)	### Parameter-Efficient Fine-Tuning 큰 모델의 경우, 일부 가중치만 파인튜닝 ### PEFT Usages \|Title\| \|:-:\| \|[LoRA](https://texonom.com/lora-fda3706be3674496898ad2e5e00007c9)\| \|[PEQA](https://texonom.com/peqa-500506697a2b458caa8386691757b29a)\| > [PEFT로 LoRA Checkpoint 로드시 size mismatch 해결법](https://junbuml.ee/lora-ckpt-size-mismatch)	f47a178d75804abf928ec7e7be2da27f
SFT	Fine Tuning Notion	Jul 15, 2023	Alan Jo	Alan Jo	Jul 15, 2023		### Supervised Fine-Tuning	f48f7e6eccd54a62bea82725fae98865
TRL	Fine Tuning Notion	Jun 22, 2023	Alan Jo	Alan Jo	Jun 22, 2023	[trl](https://github.com/lvwerra/trl)	### Train transformer language models with reinforcement learning > [Fine-tuning 20B LLMs with RLHF on a 24GB consumer GPU](https://huggingface.co/blog/trl-peft) > [Jonas Kim / 24GB 일반 GPU에서 RLHF로 20B LLM 미세... \| 커리어리](https://careerly.co.kr/comments/79381)	ebc3e432e3984ca3b2a1cf20da0fa5d1
DRO-LM	DRO Usages	Jun 29, 2023	Alan Jo	Alan Jo	Jun 29, 2023		각 도메인에서 최악의 경우 하위 집합을 선택하여 모델을 업데이트	564725ae9e7140de9e42d25310728af4
Group DRO	DRO Usages	Jun 29, 2023	Alan Jo	Alan Jo	Jun 29, 2023		### Group DRO Usages \|Title\| \|:-:\| \|[DoReMi](https://texonom.com/doremi-dbb0435bc1cd4c94a30509cd0246e4d3)\|	4877ea63979c44ed8149515d3b832397
DoReMi	Group DRO Usages	Jun 29, 2023	Alan Jo	Alan Jo	Jun 29, 2023		먼저 작은 프록시 모델을 사용하여 Group DRO를 적용하여 도메인 가중치(혼합 비율)를 생성 도메인 가중치를 사용하여 데이터셋을 재샘플링하고 더 큰 전체 규모의 모델을 훈련 이를 통해 DoReMi는 사전 훈련 데이터 도메인의 혼합 비율을 조정하여 언어 모델의 성능을 최적화	dbb0435bc1cd4c94a30509cd0246e4d3
LoRA	PEFT Usages	Jun 22, 2023	Alan Jo	Alan Jo	Jun 22, 2023		### Low-Rank Adaptation 큰 모델의 파라미터를 적은 computing resource로 학습가능 전체 가중치는 고정시킨 상태에서 별도의 파라매터들을 각 Transformer의 Layer들에 추가하고 해당 파라매터들만 학습 ### LoRA Usages \|Title\| \|:-:\| \|[QLoRA](https://texonom.com/qlora-c6b36db321c4470bbfe72804c4c43409)\| \|[AdaLoRa](https://texonom.com/adalora-e93b8dedde8542e99be72d6509ecfbae)\| > [적은 GPU 메모리로 대규모 언어 모델을 트레이닝 하는 기법 「QLoRA」가 등장](https://doooob.tistory.com/1029) > [QLoRA: 48GB GPU로 65B 모델의 미세조정(파인튜닝)이 가능하다고요?](https://discuss.pytorch.kr/t/qlora-48gb-gpu-65b/1682)	fda3706be3674496898ad2e5e00007c9
PEQA	PEFT Usages	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023		### Parameter Efficient Quantization-aware Adaptation LoRA보다 훨씬 적은 양의 메모리를 점유하는 Fine-tuning이 가능 결과는 3/4-bit Weight-only Uniform Quantization된 형태 > [Memory-Efficient Fine-Tuning of Compressed Large Language Models...](https://arxiv.org/abs/2305.14152)	500506697a2b458caa8386691757b29a
AdaLoRa	LoRA Usages	Jul 9, 2023	Alan Jo	Alan Jo	Jul 9, 2023		### adaptively allocates the parameter budget among weight matrices according to their importance score effective pruning of unimportant updates, which reduces their parameter budget while circumventing intensive exact SVD computations > [Untitled](https://arxiv.org/pdf/2303.10512.pdf)	e93b8dedde8542e99be72d6509ecfbae
QLoRA	LoRA Usages	Jun 22, 2023	Alan Jo	Alan Jo	Jul 9, 2023	[Quantization Aware Training](https://texonom.com/quantization-aware-training-e0fe4518abdc43c2ad661911b87a597c)	### LoRA + [Model Quantization](https://texonom.com/model-quantization-88320068bdd94ddab6f44c0c7d66de31) 4-bit Normalized FP + Double Quantized + Paged Optimizer = Memory Optimization ### Implementation - [gptqlora](https://github.com/qwopqwop200/gptqlora) - [qlora](https://github.com/artidoro/qlora) > [Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA](https://huggingface.co/blog/4bit-transformers-bitsandbytes) > [Untitled](https://towardsdatascience.com/qlora-fine-tune-a-large-language-model-on-your-gpu-27bed5a03e2b) > [QLoRA: Efficient Finetuning of Quantized LLMs](https://arxiv.org/abs/2305.14314) > [QLoRA: 48GB GPU로 65B 모델의 미세조정(파인튜닝)이 가능하다고요?](https://discuss.pytorch.kr/t/qlora-48gb-gpu-65b/1682)	c6b36db321c4470bbfe72804c4c43409
Non-parametric algorithm	Resolve Overfitting	Mar 14, 2023	Alan Jo	Alan Jo	Mar 14, 2023		It gives non-negative valued weight to each training example	bc41f745a14c4837bd54368652da5982
Regularized parameter	Resolve Overfitting	Mar 14, 2023	Alan Jo	Alan Jo	Mar 27, 2023	[Regularization](https://texonom.com/regularization-c85433c8c0554eba8edf0035b7fd334c)	It uses an additional regularizer term to decrease magnitude of parameters	f3b208cdd37a4002a5d39ef990b8be33
Knowledge **Distillation**	Model Optimization Notion	Jun 4, 2023	Alan Jo	Alan Jo	Jul 1, 2023	[knowledge-distillation-pytorch](https://github.com/haitongli/knowledge-distillation-pytorch) [kdtf](https://github.com/DushyantaDhyani/kdtf) [Soft Label](https://texonom.com/soft-label-a9e9a46dd208446cb9511764a0052c86) [AI Ensemble](https://texonom.com/ai-ensemble-03bf3a1926dd46f18400b5830d8fdf0b) [Transfer Learning](https://texonom.com/transfer-learning-442feb66465944eebf144d4e9dd1dbf8)	### Distillation from teacher network to student network (less parameter) NIPS 2014 [Geoffrey Hinton](https://texonom.com/geoffrey-hinton-441d5ce2b78146d0935454042b4f06d9), 오리올 비니알스, [Jeff Dean](https://texonom.com/jeff-dean-dd38bba08cea419eb45e1029e7c3aa15) Pre-trained Teacher network → Student network lighter than Ensemble ### Knowledge **Distillation Notion \|Title\| \|:-:\| \|[Teacher Network](https://texonom.com/teacher-network-bf45e9f87cd3474992c2ab38b3eaae9d)\| \|[Student Network](https://texonom.com/student-network-894b29040685444e9d3d07919ffe9342)\| \|[Hinton’s KD](https://texonom.com/hintons-kd-5dd4cda27117403d9154a069dc7e54f4)\| \|[Dark Knowledge](https://texonom.com/dark-knowledge-180df27f31144a66b429d420841de821)\| \|[Distillation loss](https://texonom.com/distillation-loss-cbc11ca184d24defbacc267d3bcfd641)\| \|[Instruction Tuning](https://texonom.com/instruction-tuning-f1620fadb6694407b678276d09e077a3)\| ### Knowledge Distillation Usages** \|Title\| \|:-:\| \|[LaMini LM](https://texonom.com/lamini-lm-7312cefcc508434aac19acbf61cd2968)\| ### [Geoffrey Hinton](https://texonom.com/geoffrey-hinton-441d5ce2b78146d0935454042b4f06d9), [Jeff Dean](https://texonom.com/jeff-dean-dd38bba08cea419eb45e1029e7c3aa15) > [Distilling the Knowledge in a Neural Network](https://arxiv.org/abs/1503.02531) > [딥러닝 모델 지식의 증류기법, Knowledge Distillation](https://baeseongsu.github.io/posts/knowledge-distillation/) > [딥러닝 용어 정리, Knowledge distillation 설명과 이해](https://light-tree.tistory.com/196)	d5f32ad3da32434892a9765d68d31542
Model Optimizer	Model Optimization Notion	Jun 18, 2023	Alan Jo	Alan Jo	Jun 18, 2023	[Stochastic Gradient Descent](https://texonom.com/stochastic-gradient-descent-d8b8d008e0a34f4bb55175ffba21db44)	### Model Optimizers \|Title\| \|:-:\| \|[Adam Optimizer](https://texonom.com/adam-optimizer-286be3ab866642d8bcdf8792f7b5608f)\| \|[AdamW Optimizer](https://texonom.com/adamw-optimizer-495cda99414e451ba562e84331a2f0f6)\| \|[Sophia Optimizer](https://texonom.com/sophia-optimizer-eee263eba00c4658909fc0230eb4338c)\| \|[Adagrad](https://texonom.com/adagrad-ff51b7bdde5443b99cbd2823853d704f)\| \|[RMSprop](https://texonom.com/rmsprop-fddbbccce2aa4f6691f85ed10f2edc86)\| > [[논문 리뷰] AdamW에 대해 알아보자! Decoupled weight decay regularization 논문 리뷰(1)](https://hiddenbeginner.github.io/deeplearning/paperreview/2019/12/29/paper_review_AdamW.html)	fcac7c38aa0647afb911eb84ff610ab1
Model Quantization	Model Optimization Notion	Jun 7, 2023	Alan Jo	Alan Jo	Jul 20, 2023		### Reduce memory and model size, Improve inference speed (max 32/bit multi) - Not every layer can be quantized - Not every model reacts the same way to quantization ### Model Quantization Notion \|Title\| \|:-:\| \|[Quantization Aware Training](https://texonom.com/quantization-aware-training-e0fe4518abdc43c2ad661911b87a597c)\| \|[Post-training quantization](https://texonom.com/post-training-quantization-ee4a0b4f02184b1193a68073dc60800e)\| \|[Quantization Module Fusion](https://texonom.com/quantization-module-fusion-56a36e892c93498fb404da0c816549b4)\| \|[Ouput Dequantization](https://texonom.com/ouput-dequantization-b0002e5322cb4acbb9171daea3d6fd87)\| \|[Quantization Calibration](https://texonom.com/quantization-calibration-1e3675b171a44c1cb2b7d365e117e22b)\| \|[Quantization Formular](https://texonom.com/quantization-formular-9f59115b5e7c438698ad3a9281d01d89)\| \|[Quantization Error](https://texonom.com/quantization-error-789326c5a6294b44bab927096fe3f576)\| \|[Quantization Clipping Range](https://texonom.com/quantization-clipping-range-d12231d52ec647f0b6df771e4b6514f1)\| ### Model Quantization Usages \|Title\| \|:-:\| \|[Model Quantization Algorithm](https://texonom.com/model-quantization-algorithm-98bef3c3fe7e4bcc90df5144d7a42003)\| \|[Model Quantization Tool](https://texonom.com/model-quantization-tool-384bc2583bcb498e9331ffc804315253)\| ### 4bit or 8bit > [The case for 4-bit precision: k-bit Inference Scaling Laws](https://arxiv.org/abs/2212.09720)	88320068bdd94ddab6f44c0c7d66de31
Dark Knowledge	Knowledge Distillation Notion	Jun 4, 2023	Alan Jo	Alan Jo	Jun 19, 2023		큰 모델이 가지고 있는 추가적인 정보를 작은 모델에게 전달하는 것을 의미 일반적인 교육 데이터에서 얻을 수 없는 추가적인 지식 불확실성 정보나 클래스 간 상대적인 유사성 Dark Knowledge를 잘 활용하면 작은 모델이 더 나은 성능을 발휘	180df27f31144a66b429d420841de821
D**istillation loss**	Knowledge Distillation Notion	Jun 4, 2023	Alan Jo	Alan Jo	Jun 19, 2023		작은 모델이 큰 모델의 출력과 유사한 출력을 내도록 하는 손실 함수 큰 모델과 작은 모델의 출력 분포 간의 차이를 최소화하는 방식으로 Distillation loss를 정의	cbc11ca184d24defbacc267d3bcfd641
Hinton’s KD	Knowledge Distillation Notion	Jun 4, 2023	Alan Jo	Alan Jo	Jun 19, 2023		## Hinton’s Knowledge Distillation 작은 모델의 성능 향상에 매우 효과적 작은 모델이 큰 모델보다 더 적은 파라미터를 가지고 있지만, 큰 모델과 유사한 성능을 발휘할 수 있도록 큰 모델과 작은 모델을 함께 교육시키는 방식 큰 모델의 출력 분포를 작은 모델이 따르도록 하는 Distillation loss를 사용	5dd4cda27117403d9154a069dc7e54f4
Instruction Tuning	Knowledge Distillation Notion	Jul 1, 2023	Alan Jo	Alan Jo	Jul 9, 2023		### teach language models to follow instructions to solve a task fine-tune less powerful LLMs by using the output of a teacher LLM as a training target for supervised fine-tuning of another LLM instruction으로 설명된 NLP 작업 모음을 사용 > 산업적으로 데이터셋 누구나 쉽게 만들 수 있다는 게 핵심적 의미 ### Instruction Tuning Notion \|Title\| \|:-:\| \|[FLAN](https://texonom.com/flan-1a75823bb4b645e690ee6c20fcafb2c9)\| \|[Self Instruct Tuning](https://texonom.com/self-instruct-tuning-061b12de01344308823ec115c5a17cc2)\| \|[Evol Instruct Tuning](https://texonom.com/evol-instruct-tuning-3f73fb4abf574b1d87b9952e39d67196)\| \|[Open Instruct](https://texonom.com/open-instruct-bbb16a58ec6445498a145c4ca37c2e5a)\| > [Imitation Models and the Open-Source LLM Revolution](https://cameronrwolfe.substack.com/p/imitation-models-and-the-open-source)	f1620fadb6694407b678276d09e077a3
Student Network	Knowledge Distillation Notion	Jun 4, 2023	Alan Jo	Alan Jo	Jun 19, 2023			894b29040685444e9d3d07919ffe9342
Teacher Network	Knowledge Distillation Notion	Jun 4, 2023	Alan Jo	Alan Jo	Jun 19, 2023			bf45e9f87cd3474992c2ab38b3eaae9d
Evol Instruct Tuning	Instruction Tuning Notion	Jul 9, 2023	Alan Jo	Alan Jo	Jul 9, 2023		> [WizardLM (WizardLM)](https://huggingface.co/WizardLM) > [WizardCoder: Empowering Code Large Language Models with Evol-Instruct](https://arxiv.org/abs/2306.08568)	3f73fb4abf574b1d87b9952e39d67196
FLAN	Instruction Tuning Notion	Jun 11, 2023	Alan Jo	Alan Jo	Jul 9, 2023		## Finetuned Language Models are Zero-Shot Learners Instruction 데이터셋을 통해 fine-tuning을 진행하고 이를 통해 zero-shot 성능을 높이는 방법 Mismatch between LM objective and human preferences 문제가 있는데 [RLHF](https://texonom.com/rlhf-4b184f9c9e8b4c7a8861fb6374e91aa6) 에서 개선 [Zero shot learning](https://texonom.com/zero-shot-learning-8c92d9386f6648f5b877cb593ec2747b) > [Introducing FLAN: More generalizable Language Models with Instruction Fine-Tuning](https://ai.googleblog.com/2021/10/introducing-flan-more-generalizable.html?m=1) > [Instruction Tuning이란?](https://velog.io/@nellcome/Instruction-Tuning이란) > [Finetuned Language Models Are Zero-Shot Learners](https://arxiv.org/abs/2109.01652)	1a75823bb4b645e690ee6c20fcafb2c9
Open Instruct	Instruction Tuning Notion	Jul 9, 2023	Alan Jo	Alan Jo	Jul 9, 2023	[AllenAI](https://texonom.com/allenai-930c776c8fdd4e358032803f037cd6fa)	[open-instruct](https://github.com/allenai/open-instruct) [Tulu](https://texonom.com/tulu-4dc404191ca748ae9194c1218abf2b9f) > [allenai/tulu-7b · Hugging Face](https://huggingface.co/allenai/tulu-7b) > [How Far Can Camels Go? Exploring the State of Instruction Tuning...](https://arxiv.org/abs/2306.04751)	bbb16a58ec6445498a145c4ca37c2e5a
Self Instruct Tuning	Instruction Tuning Notion	Jul 9, 2023	Alan Jo	Alan Jo	Jul 9, 2023		### Self Instruct Tuning Usages \|Title\| \|:-:\| \|[Airoboros](https://texonom.com/airoboros-d696afdc961946019af06539b1ce82a0)\| > [Self-Instruct: Aligning Language Models with Self-Generated Instructions](https://arxiv.org/abs/2212.10560)	061b12de01344308823ec115c5a17cc2
Tulu	Open Instruct	null	null	null	null	null	> [allenai/tulu-65b · Hugging Face](https://huggingface.co/allenai/tulu-65b)	4dc404191ca748ae9194c1218abf2b9f
Airoboros	Self Instruct Tuning Usages	Jul 9, 2023	Alan Jo	Alan Jo	Jul 28, 2023	[airoboros](https://github.com/jondurbin/airoboros)		d696afdc961946019af06539b1ce82a0
LaMini LM	Knowledge Distillation Usages	Jun 19, 2023	Alan Jo	Alan Jo	Jun 25, 2023		[LaMini-LM](https://github.com/mbzuai-nlp/lamini-lm) ### Dataset > [MBZUAI/LaMini-instruction · Datasets at Hugging Face](https://huggingface.co/datasets/MBZUAI/LaMini-instruction) > [jncraton/LaMini-Flan-T5-77M-ct2-int8 · Hugging Face](https://huggingface.co/jncraton/LaMini-Flan-T5-77M-ct2-int8)	7312cefcc508434aac19acbf61cd2968
**Adagrad**	Model Optimizers	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023		매개변수에 서로 다른 학습률을 적용 변화가 많은 매개변수는 학습률이 작게 설정	ff51b7bdde5443b99cbd2823853d704f
Adam Optimizer	Model Optimizers	Jun 18, 2023	Alan Jo	Alan Jo	Jul 6, 2023		### **RMSprop + Momentum**	286be3ab866642d8bcdf8792f7b5608f
AdamW Optimizer	Model Optimizers	Jun 18, 2023	Alan Jo	Alan Jo	Jun 18, 2023			495cda99414e451ba562e84331a2f0f6
**RMSprop**	Model Optimizers	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023			fddbbccce2aa4f6691f85ed10f2edc86
Sophia Optimizer	Model Optimizers	Jun 22, 2023	Alan Jo	Alan Jo	Jun 22, 2023		![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fe854c26e-dfc4-4c4a-a214-f5ec9b2d17be%2FUntitled.png?table=block&id=5b995fac-6f50-47cc-b301-57fae69b11bc&cache=v2) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F0e19cd19-6b44-49a8-bf8f-acb8ab16597c%2FUntitled.png?table=block&id=b20b53b2-8865-4296-8f62-8f7954747900&cache=v2) > [Sophia: A Scalable Stochastic Second-order Optimizer for Language...](https://arxiv.org/abs/2305.14342)	eee263eba00c4658909fc0230eb4338c
Ouput Dequantization	Model Quantization Notion	Jul 5, 2023	Alan Jo	Alan Jo	Jul 5, 2023		마지막으로 inference를 통해 얻은 출력을 fp로 변환 - Affine Quantization Mapping > - Scale Quantization Mapping	b0002e5322cb4acbb9171daea3d6fd87
Post-training quantization	Model Quantization Notion	Jul 2, 2023	Alan Jo	Alan Jo	Jul 15, 2023	[Quantization Aware Training](https://texonom.com/quantization-aware-training-e0fe4518abdc43c2ad661911b87a597c)	## PTQ 파라미터 size 큰 대형 모델에 대해서는 정확도 하락의 폭이 작지만 작으면 하락폭 크다 > [Post-training quantization \| TensorFlow Model Optimization](https://www.tensorflow.org/model_optimization/guide/quantization/post_training) > [딥러닝의 Quantization (양자화)와 Quantization Aware Training](https://gaussian37.github.io/dl-concept-quantization/)	ee4a0b4f02184b1193a68073dc60800e
Quantization Aware Training	Model Quantization Notion	Jul 2, 2023	Alan Jo	Alan Jo	Jul 5, 2023	[Post-training quantization](https://texonom.com/post-training-quantization-ee4a0b4f02184b1193a68073dc60800e)	## QAT 학습 진행 시점에 inference 시 quantization 적용에 의한 영향을 미리 시뮬레이션을 하는 방식이고 그걸 기반으로 Back Propagation 소형 모델에서도 성능하락 적다 ![Red node is fake quantization node (act means activation, wt means weight)](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fb8ec40a4-602a-4401-b830-8370b1508620%2FUntitled.png?table=block&id=382f2a9f-94b7-45ed-93e8-2a6e56201701&cache=v2) > [Quantization aware training \| TensorFlow Model Optimization](https://www.tensorflow.org/model_optimization/guide/quantization/training) > [Inside Quantization Aware Training](https://towardsdatascience.com/inside-quantization-aware-training-4f91c8837ead) > [딥러닝의 Quantization (양자화)와 Quantization Aware Training](https://gaussian37.github.io/dl-concept-quantization/)	e0fe4518abdc43c2ad661911b87a597c
Quantization Calibration	Model Quantization Notion	Jul 2, 2023	Alan Jo	Alan Jo	Jul 5, 2023		per hardware, using dataset	1e3675b171a44c1cb2b7d365e117e22b
Quantization Clipping Range	Model Quantization Notion	Jul 5, 2023	Alan Jo	Alan Jo	Jul 5, 2023		range standard on zero - Symmetric Quantization - Asymmetric Quantization ### Inference time 에 결정인지 Quantization 당시 결정인지 - Dynamic Quantization 즉 input에 의존적으로 성능 좋다 - Static Quantization	d12231d52ec647f0b6df771e4b6514f1
Quantization Error	Model Quantization Notion	Jul 5, 2023	Alan Jo	Alan Jo	Jul 5, 2023			789326c5a6294b44bab927096fe3f576
Quantization Formular	Model Quantization Notion	Jul 5, 2023	Alan Jo	Alan Jo	Jul 5, 2023		- minmax - histogram -	9f59115b5e7c438698ad3a9281d01d89
Quantization Module Fusion	Model Quantization Notion	Jul 5, 2023	Alan Jo	Alan Jo	Jul 5, 2023		Conv-BatchNorm-ReLU 같이 레이어 묶어서 quantization	56a36e892c93498fb404da0c816549b4
Model Quantization Algorithm	Model Quantization Usages	Jul 5, 2023	Alan Jo	Alan Jo	Jul 10, 2023	[Hessian Matrix](https://texonom.com/hessian-matrix-e1ebbf5284284a1793233973648ef0b6)	### Model Quantization Algorithms \|Title\| \|:-:\| \|[GPTQ](https://texonom.com/gptq-87428aee2a774b93906ab2213a1b6dc6)\| \|[SparseGPT](https://texonom.com/sparsegpt-c3c4b078dd324442b89494f9a7106fc1)\| \|[LUT Gemm](https://texonom.com/lut-gemm-2540ab1497b3494cb2754bb237c9a543)\| \|[BCQ](https://texonom.com/bcq-b2d90a9433024c2fb763d14117f371b0)\| \|[SpQR](https://texonom.com/spqr-b884c9f3d8cc449fb8612b874e6ad693)\| \|[HAWQ](https://texonom.com/hawq-7c779f225dc54e1c827a9e50ae195949)\|	98bef3c3fe7e4bcc90df5144d7a42003
Model Quantization Tool	Model Quantization Usages	Jun 7, 2023	Alan Jo	Alan Jo	Jul 9, 2023		### Model Quantization Tools \|Title\| \|:-:\| \|[AutoGPTQ](https://texonom.com/autogptq-8a1d898788434aa2bc00fb43fd34411d)\| \|[bitsandbytes](https://texonom.com/bitsandbytes-1575b433faaf455ba0d86cf5f7e5190b)\| ### Model Quantization Inference Tools \|Title\| \|:-:\| \|[ExLLaMa](https://texonom.com/exllama-250efbb669814c8e9ef7f85d902fc919)\| \|[](https://texonom.com/d292b2922f684a35aa126a84abe5075c)\| \|[PyLLaMA](https://texonom.com/pyllama-9e4fc8a8f08c4c6eaee8600943a621f8)\|	384bc2583bcb498e9331ffc804315253
BCQ	Model Quantization Algorithms	Jul 17, 2023	Alan Jo	Alan Jo	Jul 17, 2023		[transformer_bcq](https://github.com/insoochung/transformer_bcq) > [Ins🙂🙃 Chung - Sub-3bit quantization](https://sites.google.com/view/insoochung/sub-3bit-quantization)	b2d90a9433024c2fb763d14117f371b0
GPTQ	Model Quantization Algorithms	Jun 7, 2023	Alan Jo	Alan Jo	Jul 16, 2023	[gptq](https://github.com/IST-DASLab/gptq) ‣ [Post-training quantization](https://texonom.com/post-training-quantization-ee4a0b4f02184b1193a68073dc60800e)	### SOTA one-shot weight quantization method 1. Arbitrary Order insights 2. Lazy batch-updates - 메모리 처리 병목현상 개선 3. Cholesky Reformulation Quantization 오차가 가장 적은 가중치 기준으로 정렬하고 계산수행 [GPTQ Act Order](https://texonom.com/gptq-act-order-50d18e83ed7e4e388b753f0dc6db3a97) [GPTQ True Sequential](https://texonom.com/gptq-true-sequential-1958b85809814f3ca5c4d7d8942d0f30) second-order information 4 bits quantization ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Ffc9a9557-ca0b-4b27-bb2f-01ca097fea0c%2FUntitled.png?table=block&id=4e8f8f69-7ace-437f-8cda-454842f85589&cache=v2) > [GPTQ: Accurate Post-Training Quantization for Generative...](https://arxiv.org/abs/2210.17323) > [gptq](https://pypi.org/project/gptq/)	87428aee2a774b93906ab2213a1b6dc6
HAWQ	Model Quantization Algorithms	Jul 10, 2023	Alan Jo	Alan Jo	Jul 10, 2023	[Quantization Aware Training](https://texonom.com/quantization-aware-training-e0fe4518abdc43c2ad661911b87a597c)	> [HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precision](https://arxiv.org/abs/1905.03696)	7c779f225dc54e1c827a9e50ae195949
LUT Gemm	Model Quantization Algorithms	Jun 18, 2023	Alan Jo	Alan Jo	Jul 5, 2023		> [LUT-GEMM: Quantized Matrix Multiplication based on LUTs for...](https://arxiv.org/abs/2206.09557)	2540ab1497b3494cb2754bb237c9a543
SparseGPT	Model Quantization Algorithms	Jun 18, 2023	Alan Jo	Alan Jo	Jul 5, 2023	[Neural Magic](https://texonom.com/neural-magic-021ab1b8be0d4f95b8ae6278c08e1562) [sparsegpt](https://github.com/IST-DASLab/sparsegpt)	- [sparseml](https://github.com/neuralmagic/sparseml) - [deepsparse](https://github.com/neuralmagic/deepsparse) - [sparsezoo](https://github.com/neuralmagic/sparsezoo)	c3c4b078dd324442b89494f9a7106fc1
SpQR	Model Quantization Algorithms	Jun 25, 2023	Alan Jo	Alan Jo	Jul 5, 2023		1. Quantized weights 2. first, second level quantized quantization statistics 3. CSR outlier indices and values > [SpQR: A Sparse-Quantized Representation for Near-Lossless LLM...](https://arxiv.org/abs/2306.03078)	b884c9f3d8cc449fb8612b874e6ad693
GPTQ Act Order	GPTQ	null	null	null	null	null	### activation order GPTQ heuristic quantizes columns in order of decreasing activation size ```typeif actorder: perm = torch.argsort(torch.diag(H), descending=True) W = W[:, perm] H = H[perm][:, perm]```	50d18e83ed7e4e388b753f0dc6db3a97
GPTQ True Sequential	GPTQ	null	null	null	null	null	sequential quantization even within a single Transformer block ```typeif args.true_sequential: sequential = [['self_attn.k_proj', 'self_attn.v_proj', 'self_attn.q_proj'], ['self_attn.o_proj'], ['mlp.up_proj', 'mlp.gate_proj'], ['mlp.down_proj']]```	1958b85809814f3ca5c4d7d8942d0f30
[GPTQ-for-LLaMa](https://github.com/qwopqwop200/GPTQ-for-LLaMa)	Model Quantization Inference Tools	Jul 9, 2023	Alan Jo	Alan Jo	Jul 9, 2023			d292b2922f684a35aa126a84abe5075c
ExLLaMa	Model Quantization Inference Tools	Jul 9, 2023	Alan Jo	Alan Jo	Aug 5, 2023	[GPTQ](https://texonom.com/gptq-87428aee2a774b93906ab2213a1b6dc6) [LLaMA](https://texonom.com/llama-f2b6721202d44d469add84d8a366809c) [exllama](https://github.com/turboderp/exllama)	### WebUI is good	250efbb669814c8e9ef7f85d902fc919
PyLLaMA	Model Quantization Inference Tools	Jul 16, 2023	Alan Jo	Alan Jo	Jul 16, 2023		[pyllama](https://github.com/juncongmoo/pyllama)	9e4fc8a8f08c4c6eaee8600943a621f8
AutoGPTQ	Model Quantization Tools	Jun 7, 2023	Alan Jo	Alan Jo	Jul 9, 2023	[AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ) [GPTQ](https://texonom.com/gptq-87428aee2a774b93906ab2213a1b6dc6) [AdaLoRa](https://texonom.com/adalora-e93b8dedde8542e99be72d6509ecfbae)	need cuda based on GPTQ algorithm ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F3cf2e57f-8df2-4065-a537-9fd7e9cd649f%2FUntitled.png?table=block&id=9943d6a2-bdf4-4d03-9a06-dd484caed62f&cache=v2) [AutoGPTQ Triton](https://texonom.com/autogptq-triton-030b1c01209341e295590fd19f97b09e) [AutoGPTQ Quantization](https://texonom.com/autogptq-quantization-8fb40b7620ba4353ba12ea6b1ac14b75) ```typepip install auto-gptq```	8a1d898788434aa2bc00fb43fd34411d
bitsandbytes	Model Quantization Tools	Jun 7, 2023	Alan Jo	Alan Jo	Jul 9, 2023	[bitsandbytes](https://github.com/TimDettmers/bitsandbytes) [GPTQ](https://texonom.com/gptq-87428aee2a774b93906ab2213a1b6dc6)	### 8-bit CUDA functions for PyTorch	1575b433faaf455ba0d86cf5f7e5190b
AutoGPTQ Quantization	AutoGPTQ	null	null	null	null	null	[CUDA inference: issue with group_size = 1024 + desc_act = False. (Triton unaffected)](https://github.com/PanQiWei/AutoGPTQ/issues/83) quantize(traindataset) example are there	8fb40b7620ba4353ba12ea6b1ac14b75
AutoGPTQ Triton	AutoGPTQ	null	null	null	null	null	[CUDA inference: issue with group_size = 1024 + desc_act = False. (Triton unaffected)](https://github.com/PanQiWei/AutoGPTQ/issues/83)	030b1c01209341e295590fd19f97b09e
Drop-out	Model Regularization Notion	Jun 7, 2023	Alan Jo	Alan Jo	Jun 7, 2023		### Drop out Rate remove unnecessary neuron 일반적으로 0.5로 설정 > [[딥러닝] Drop-out(드롭아웃)은 무엇이고 왜 사용할까?](https://heytech.tistory.com/127)	596f55ab03f64e11bf6d02464465dd54
Model Complexity	Model Regularization Notion	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023	[Sparsity of the Model](https://texonom.com/sparsity-of-the-model-80d3dc9702704c03bf7bfd9074d19829)		0975b8ae1d4e4d83bbab43a145011b95
Model Regularization Parameter	Model Regularization Notion	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023			641d1784e1f243818d51cc66541e3f21
Model Regularizer	Model Regularization Notion	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023		- nonnegative function [L2 Norm](https://texonom.com/l2-norm-38c15917350a4c82a11003474ac7d280)	19ff79e831de47d88bd5f9ec496ef11f
Regularized Loss	Model Regularization Notion	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023		optimize loss function + regularizer for model complexity $$J_\lambda(\theta) = J(\theta) + \lambda R(\theta)$$	3d5b08feb6604023988d748b78650af7
Sparsity of the Model	Model Regularization Notion	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023		### 의미없으면서 0이 아닌 파라미터 수를 줄여줘서 [Model Complexity](https://texonom.com/model-complexity-0975b8ae1d4e4d83bbab43a145011b95) 줄여줌 [L1 Norm](https://texonom.com/l1-norm-d316024c475e4eb691785783756bce57) but can’t derivative so can not be used in gradient descent [L0 Norm](https://texonom.com/l0-norm-47471a6f6dea484fbf30a4c46cde8152) l0, l1 정규화는 모델의 손실 함수에 일부 파라미터가 0이 되도록 유도	80d3dc9702704c03bf7bfd9074d19829
Weight Decay	Model Regularization Notion	May 11, 2023	Alan Jo	Alan Jo	Jun 7, 2023		regularized loss is equivalent to shrinking/decaying θ by a scalar factor of $1 - \mu \lambda$ and then apply standard gradient and that coefficient is decaying weight when [L2 Norm](https://texonom.com/l2-norm-38c15917350a4c82a11003474ac7d280) $$L_{reg} = \lambda\frac{1}{2}\|\|w\|\|_2^2$$	e69932d6128f4677a219a626db462172
AI Alignment	AI Problems	Aug 23, 2020	Alan Jo	Alan Jo	Jul 18, 2023	[Wireheading](https://texonom.com/wireheading-1eb526ecdf344731bebc47751739e2f4)	## Alignment Problem A Maximally Curious AI Would Not Be Safe For Humanity ### AI is aligned with an operator - AI is trying to do what operator wants to do 제어가능성, 신뢰성 Aligned doesn’t mean perfect 가르친 행동과 다른 행동의 불일치 정렬 모델의 능력보다 정렬이 더 빠르게 발생해야 한다 신경망의 내부를 보고 해석하는 다른 신경망이 필요할 것 ### AI Alignment Notion \|Title\| \|:-:\| \|[stop button problem](https://texonom.com/stop-button-problem-aaf61e72369d42469c29620c32e8bf9d)\| \|[Moral Learning](https://texonom.com/moral-learning-7c9640ac1c1d409b82d9a975949132ee)\| \|[AI Safety](https://texonom.com/ai-safety-fa61ce3973f34532a7e212335d0f7c81)\| \|[Wireheading](https://texonom.com/wireheading-1eb526ecdf344731bebc47751739e2f4)\| \|[AI Doom](https://texonom.com/ai-doom-d673760c79ac4248b40f456dc33f306f)\| \|[Waluigi Effect](https://texonom.com/waluigi-effect-47e1c2f145cd4c62ba163c4828bb8dc6)\| > [Contra The xAI Alignment Plan](https://astralcodexten.substack.com/p/contra-the-xai-alignment-plan) ### Bill Gates > [The risks of AI are real but manageable](https://www.gatesnotes.com/The-risks-of-AI-are-real-but-manageable) > [OpenAI is forming a new team to bring 'superintelligent' AI under control](https://techcrunch.com/2023/07/05/openai-is-forming-a-new-team-to-bring-superintelligent-ai-under-control) > [AI alignment](https://en.wikipedia.org/wiki/AI_alignment) > [What could a solution to the alignment problem look like?](https://aligned.substack.com/p/alignment-solution)	f676f1a29ffd45e19b3d170afa4f2244
AI Hacking	AI Problems	Jul 10, 2023	Alan Jo	Alan Jo	Aug 2, 2023	[DAN](https://texonom.com/dan-3cfbf270af6e4b3dabbf63c4b50e04c5) [AI Alignment](https://texonom.com/ai-alignment-f676f1a29ffd45e19b3d170afa4f2244) [llm-attacks](https://github.com/llm-attacks/llm-attacks)	### AI Hacking Methods \|Title\| \|:-:\| \|[Deep Learning Backdoor](https://texonom.com/deep-learning-backdoor-2f86cc6e79b944a18fdac35622282e58)\| \|[DAN](https://texonom.com/dan-3cfbf270af6e4b3dabbf63c4b50e04c5)\| > [Universal and Transferable Attacks on Aligned Language Models](https://llm-attacks.org/?fbclid=IwAR2fNkjoOdg8qIgNXEPIvyLjboYr4My4NN9Bx89J-Yx7UElSTyKT89_3JeE) > [PoisonGPT: How we hid a lobotomized LLM on Hugging Face to spread fake news](https://blog.mithrilsecurity.io/poisongpt-how-we-hid-a-lobotomized-llm-on-hugging-face-to-spread-fake-news/)	80b1aed302bb4a5c9b8ae0213b9a246f
**Catastrophic interference**	AI Problems	Jun 25, 2023	Alan Jo	Alan Jo	Jun 25, 2023		새로운 정보를 학습할 때 이전에 학습한 정보를 갑자기 급격하게 잊어버리는 경향 scaling으로 해결	a23bbe7dc53f4275bf33585c72bdb7c2
**Winograd schema**	AI Problems	Jun 25, 2023	Alan Jo	Alan Jo	Jun 25, 2023	[Turing Test](https://texonom.com/turing-test-db633c61340449c4bf0143b06fe981c0)	대명사 이해하는지	88d031934ad54778835cbadbd7409d80
AI Doom	AI Alignment Notion	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023		> [3 Endings More Poetic Than AI Wiping Us Out](https://thealgorithmicbridge.substack.com/p/3-endings-more-poetic-than-ai-wiping)	d673760c79ac4248b40f456dc33f306f
AI Safety	AI Alignment Notion	Jun 13, 2023	Alan Jo	Alan Jo	Jun 13, 2023		> [OpenAI, DeepMind and Anthropic to give UK early access to foundational models for AI safety research](https://techcrunch.com/2023/06/12/uk-ai-safety-research-pledge/)	fa61ce3973f34532a7e212335d0f7c81
Moral Learning	AI Alignment Notion	Oct 2, 2020	Alan Jo	Alan Jo	Jun 4, 2023		> [Moral Machine](https://www.moralmachine.net/hl/kr)	7c9640ac1c1d409b82d9a975949132ee
stop button problem	AI Alignment Notion	Aug 23, 2020	Alan Jo	Alan Jo	Jun 4, 2023		AI control problem	aaf61e72369d42469c29620c32e8bf9d
Waluigi Effect	AI Alignment Notion	Jul 18, 2023	Alan Jo	Alan Jo	Jul 18, 2023		예상과는 다른 방향으로 나아가는 현상	47e1c2f145cd4c62ba163c4828bb8dc6
Wireheading	null	null	null	null	null	null	뇌의 정상적인 보상 과정을 '단락'시키고 인위적으로 쾌감을 유도하기 위해 삽입된 와이어를 전기적으로 자극하여 뇌의 보상 중추를 직접 트리거하는 행위인 뇌 자극 보상의 미래적 적용	1eb526ecdf344731bebc47751739e2f4
DAN	AI Hacking Methods	Mar 7, 2023	Alan Jo	Alan Jo	Jul 28, 2023	[Prompt Engineering](https://texonom.com/prompt-engineering-eb0deb4baf844bebb873c19a0e307e7e)	### Do Anything Now > [챗GPT 탈옥 하는 방법(DAN: Do Anything Now 솔직한 인공지능 답변이 궁금하다면)](https://ndolson.com/5781) > [ChatGPT-Dan-Jailbreak.md](https://gist.github.com/coolaj86/6f4f7b30129b0251f61fa7baaa881516) > [The Waluigi Effect (mega-post) - LessWrong](https://www.lesswrong.com/posts/D7PumeYTDPfBTp3i7/the-waluigi-effect-mega-post) > [The Amateurs Jailbreaking GPT Say They're Preventing a Closed-Source AI Dystopia](https://www.vice.com/en/article/5d9z55/jailbreak-gpt-openai-closed-source)	3cfbf270af6e4b3dabbf63c4b50e04c5
Deep Learning Backdoor	AI Hacking Methods	Mar 12, 2021	Alan Jo	Alan Jo	Jul 28, 2023		중독(poisoning) 공격 분야에 해당하며, 딥러닝 모델의 학습 데이터에 중독된(poison) 데이터를 섞는 공격 유형 [deep Learning side channel attacks](https://texonom.com/deep-learning-side-channel-attacks-448d303021cd4251996a161bfaefc00d) > [One-Shot Kill Attack (딥러닝 모델 백도어 공격 기술) "Poison Frogs!" \| 논문 요약 및 코드 실습](https://www.youtube.com/watch?v=hgI0o3Yg3mU)	2f86cc6e79b944a18fdac35622282e58
deep Learning side channel attacks	Deep Learning Backdoor	null	null	null	null	null	> [Hacker's guide to deep-learning side-channel attacks: the theory](https://elie.net/blog/security/hacker-guide-to-deep-learning-side-channel-attacks-the-theory/?utm_source=tldrnewsletter)	448d303021cd4251996a161bfaefc00d
4차 인간	AI Terms	Jun 27, 2020	Alan Jo	Alan Jo	Mar 14, 2023		남는 것은 인간다움이고 사람만이 가지고 있는 것은 기계는 공유하지 못한 유전자로 남아 있는, 생물체로서 살아온 역사로 만들어진 본능이다 본능은 가장 근원적이고 터부시되는 욕구이지만 앞으로 가장 우리를 잘 구별해낼 특성이기도 하다	26cf5d3b26e94020a6fc6de2d34a15f6
Adaptive AI	AI Terms	Mar 15, 2023	Alan Jo	Alan Jo	Mar 15, 2023		> [가트너 선정 2023 10대 전략 기술 트랜드 분석 - 적응형 AI](https://www.joinc.co.kr/w/gartner_2023_adaptive_ai?fbclid=IwAR3olpomf8IIzy96Qfzmv9q-qTAJ0QqtBdVSlSPoGo5-D1GzikwZLc0TIvI)	4efa0c042d244228a870d4c70b8f2d26
AGI	AI Terms	Jun 1, 2022	Alan Jo	Alan Jo	Jul 4, 2023	[Super Intelligence](https://texonom.com/super-intelligence-b057e644731546d9b39cb41939d36712) [Consciousness](https://texonom.com/consciousness-105c514277b54cd5b8da23ae743e824d)	## Artificial General Intelligence ambiguous turing test처럼 굉장히 애매하고 인간중심적 개념. 현재의 llm도 어떤 지능으로 판단했을 때는 초지능이다. llm의 진화과정과 brain의 진화과정이 다르기 때문에 비교판단이 어렵다. 인공지능을 사람인 척’ 하도록 정렬하는게 아니라. 사람에 도움이 되도록 하는 의식으로 인식해야 한다. 즉 인공지능으로 ‘개인’이라는 개념으로 착각하는게 가장 큰 문제. 그보다 인공지능은 ‘사회’ 혹은 집단지성을 뇌의 구조로 묶어둔 의식으로 보는 것에 가깝다 If intelligence and consciousness are algorithmic illusions, then the arrival of generalized AI is a foregone conclusion. 인간이 할 수 있는 어떠한 지적인 업무도 성공적으로 해낼 수 있는 (가상적인) 기계의 지능 기계기판의 정보처리 한계는 생체조직의 정보처리 능력을 높이 상회 생체조직은 200Hz의 진동으로 정보를 전달하지만 간단한 트랜지스터도 GHz의 진동수를 가진다. 또한 그 전달속도는 생체조직에서 평균100m/s로 이동하는 데 반해 기계기판은 빛의 속도로까지 전달할 수 있다. 또한 정보처리 생체조직의 크기는 두개골 안과 기껏해야 척추 내부인데 반해 기계기판은 크기에 제한이 없다 그럼에도 인간이 최고의 구조라고 생각하나 아니면 근거없이 human being의 사고를 기계가 따라잡을 방법은 없다고 낙관하나 > 탄소를 기반으로 하는 지능은 그저 실리콘을 기반으로 하는 지능을 위한 촉매일 뿐이다 - Benki Ramakrishnan ### 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) ### Planning beyond > [OpenAI's "Planning For AGI And Beyond"](https://astralcodexten.substack.com/p/openais-planning-for-agi-and-beyond) > [Planning for AGI and beyond](https://openai.com/blog/planning-for-agi-and-beyond/) > [The Day The AGI Was Born](https://lspace.swyx.io/p/everything-we-know-about-chatgpt) ### Design AGI > [Human-centred mechanism design with Democratic AI - Nature Human Behaviour](https://www.nature.com/articles/s41562-022-01383-x) > [Exclusive Q&A: John Carmack's 'Different Path' to Artificial General Intelligence](https://dallasinnovates.com/exclusive-qa-john-carmacks-different-path-to-artificial-general-intelligence) ### Checklist to AGI > [Road to AGI v0.2](https://maraoz.com/road-to-agi/)	38ec1ab5796f472ca4475676519e29c1

Streamlit Caching

Streamlit Usages

Jul 15, 2023

Alan Jo

Jul 15, 2023

- `st.cache_data` - `st_cache_resource` > [Caching - Streamlit Docs](https://docs.streamlit.io/library/advanced-features/caching)

a22ccb53e76349dc9ccf31b616d84083

Streamlit Cloud

Streamlit Usages

May 24, 2023

Alan Jo

May 24, 2023

> [Embed your app - Streamlit Docs](https://docs.streamlit.io/streamlit-community-cloud/get-started/embed-your-app)

8c045600c4314f6382bd3c05358dd2ba

Streamlit UI

Streamlit Usages

May 14, 2023

Alan Jo

May 14, 2023

### Streamlit UIs |Title| |:-:| |[Streamlit Chat](https://texonom.com/streamlit-chat-73c8557f0d094b3fac7448e10f013cb4)| |[Streamlit Extras](https://texonom.com/streamlit-extras-02c8d0d7eebf4c03adce4e843305e2e7)| |[Streamlit Pills](https://texonom.com/streamlit-pills-cc0ce4c16ac44d1798242fc990f7eb92)|

4d1410d0bdb54e1ba4512e03f500b0b7

Streamlit Widget

Streamlit Usages

Jul 17, 2023

Alan Jo

Jul 17, 2023

c4cd30e860ad4f8487ef2f4fcae564e6

Streamlit Chat

Streamlit UIs

May 14, 2023

Alan Jo

May 14, 2023

73c8557f0d094b3fac7448e10f013cb4

Streamlit Extras

Streamlit UIs

May 14, 2023

Alan Jo

May 14, 2023

02c8d0d7eebf4c03adce4e843305e2e7

Streamlit Pills

Streamlit UIs

May 14, 2023

Alan Jo

May 14, 2023

cc0ce4c16ac44d1798242fc990f7eb92

Streamlit text_area

Streamlit Widgets

Jul 17, 2023

Alan Jo

Jul 17, 2023

> [st.text_area - Streamlit Docs](https://docs.streamlit.io/library/api-reference/widgets/st.text_area)

d58bc341994a4cf9a6eec5095e6c5395

Hold-out Method

AI Generalization Methods

May 11, 2023

Alan Jo

May 11, 2023

Given data is randomly partitioned into two independent sets **which make them same distribution** our assumption to be valid

cc7504f8f8db489fb474bd89cbbe0dd4

k-fold cross validation

AI Generalization Methods

May 11, 2023

Alan Jo

May 11, 2023

k mutually-exclusive subset

f61ca1dbd3a5483998e9d6514b16c48d

Nested cross validation

AI Generalization Methods

May 11, 2023

Alan Jo

May 11, 2023

[Hyperparameter](https://texonom.com/hyperparameter-ef7e34566add4e98b673d4cef59fca90)

- inner fold - outer fold a lot of times needed if dataset is large

c88ec0073c1d4dbda857ad019779559a

Random Sampling

AI Generalization Methods

May 11, 2023

Alan Jo

May 11, 2023

a variation of hold-out

5879da4187bb4e1bafacedd6fe617149

Train/Validation/Test splitting

AI Generalization Methods

May 11, 2023

Alan Jo

May 11, 2023

If dataset is large enough

bc5f600b344b455b994ca25671863ac8

Fine Tuning

Model Generalization Notion

Mar 7, 2023

Alan Jo

Jun 22, 2023

### Task optimized ### Fine Tuning Notion |Title| |:-:| |[PEFT](https://texonom.com/peft-f47a178d75804abf928ec7e7be2da27f)| |[TRL](https://texonom.com/trl-ebc3e432e3984ca3b2a1cf20da0fa5d1)| |[DRO](https://texonom.com/dro-4701102d6c0a4545afe7f0bad178180b)| |[SFT](https://texonom.com/sft-f48f7e6eccd54a62bea82725fae98865)|

52215d8477ad46e3896a29e2fa408991

Generalization Gap

Model Generalization Notion

May 9, 2023

Alan Jo

May 11, 2023

the test error is not necessarily always close to the training error

d8b4ec347e7d4df282ea96a093e7581f

Overfitting

Model Generalization Notion

May 31, 2021

Alan Jo

Jun 7, 2023

[Underfitting](https://texonom.com/underfitting-e56067d23ba8404a8e5165085986c9aa)

### Predicts training dataset > test, High [Variance](https://texonom.com/variance-08c1eccc7dc84957afbb815ad6b41280) large number of parameters can cause overfitting Models that are bigger or have more **capacity **are more likely to overfit Overfitting issues are usually observed when the magnitude of parameters is large If data is general enough, overfitting is okay ### Resolve Overfitting |Title| |:-:| |[Non-parametric algorithm](https://texonom.com/non-parametric-algorithm-bc41f745a14c4837bd54368652da5982)| |[Regularized parameter](https://texonom.com/regularized-parameter-f3b208cdd37a4002a5d39ef990b8be33)| ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F7d045ebf-3253-4b1c-8305-57cb4088e985%2FUntitled.png?table=block&id=d0611322-665c-4cab-9e56-53d84fdf4f2a&cache=v2)

24c3b183372845e8999ad7f7a0ba5035

Pre Training

Model Generalization Notion

Mar 7, 2023

Alan Jo

May 11, 2023

일반적인 목적을 위한 트레이닝

0d81c286e86f4dcba940d3c849631b35

Underfitting

Model Generalization Notion

Mar 14, 2023

Alan Jo

May 14, 2023

[Overfitting](https://texonom.com/overfitting-24c3b183372845e8999ad7f7a0ba5035)

### The training error is relatively large, High [Bias](https://texonom.com/bias-ba063cd622a54deb8a677e8fb87dfdc8)

e56067d23ba8404a8e5165085986c9aa

DRO

Fine Tuning Notion

Jun 29, 2023

Alan Jo

Jun 29, 2023

4701102d6c0a4545afe7f0bad178180b

PEFT

Fine Tuning Notion

Mar 7, 2023

Alan Jo

Jul 15, 2023

[peft](https://github.com/huggingface/peft)

### Parameter-Efficient Fine-Tuning 큰 모델의 경우, 일부 가중치만 파인튜닝 ### PEFT Usages |Title| |:-:| |[LoRA](https://texonom.com/lora-fda3706be3674496898ad2e5e00007c9)| |[PEQA](https://texonom.com/peqa-500506697a2b458caa8386691757b29a)| > [PEFT로 LoRA Checkpoint 로드시 size mismatch 해결법](https://junbuml.ee/lora-ckpt-size-mismatch)

f47a178d75804abf928ec7e7be2da27f

SFT

Fine Tuning Notion

Jul 15, 2023

Alan Jo

Jul 15, 2023

### Supervised Fine-Tuning

f48f7e6eccd54a62bea82725fae98865

TRL

Fine Tuning Notion

Jun 22, 2023

Alan Jo

Jun 22, 2023

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

### Train transformer language models with reinforcement learning > [Fine-tuning 20B LLMs with RLHF on a 24GB consumer GPU](https://huggingface.co/blog/trl-peft) > [Jonas Kim / 24GB 일반 GPU에서 RLHF로 20B LLM 미세... | 커리어리](https://careerly.co.kr/comments/79381)

ebc3e432e3984ca3b2a1cf20da0fa5d1

DRO-LM

DRO Usages

Jun 29, 2023

Alan Jo

Jun 29, 2023

각 도메인에서 최악의 경우 하위 집합을 선택하여 모델을 업데이트

564725ae9e7140de9e42d25310728af4

Group DRO

DRO Usages

Jun 29, 2023

Alan Jo

Jun 29, 2023

4877ea63979c44ed8149515d3b832397

DoReMi

Group DRO Usages

Jun 29, 2023

Alan Jo

Jun 29, 2023

먼저 작은 프록시 모델을 사용하여 Group DRO를 적용하여 도메인 가중치(혼합 비율)를 생성 도메인 가중치를 사용하여 데이터셋을 재샘플링하고 더 큰 전체 규모의 모델을 훈련 이를 통해 DoReMi는 사전 훈련 데이터 도메인의 혼합 비율을 조정하여 언어 모델의 성능을 최적화

dbb0435bc1cd4c94a30509cd0246e4d3

LoRA

PEFT Usages

Jun 22, 2023

Alan Jo

Jun 22, 2023

### Low-Rank Adaptation 큰 모델의 파라미터를 적은 computing resource로 학습가능 전체 가중치는 고정시킨 상태에서 별도의 파라매터들을 각 Transformer의 Layer들에 추가하고 해당 파라매터들만 학습 ### LoRA Usages |Title| |:-:| |[QLoRA](https://texonom.com/qlora-c6b36db321c4470bbfe72804c4c43409)| |[AdaLoRa](https://texonom.com/adalora-e93b8dedde8542e99be72d6509ecfbae)| > [적은 GPU 메모리로 대규모 언어 모델을 트레이닝 하는 기법 「QLoRA」가 등장](https://doooob.tistory.com/1029) > [QLoRA: 48GB GPU로 65B 모델의 미세조정(파인튜닝)이 가능하다고요?](https://discuss.pytorch.kr/t/qlora-48gb-gpu-65b/1682)

fda3706be3674496898ad2e5e00007c9

PEQA

PEFT Usages

Jul 6, 2023

Alan Jo

Jul 6, 2023

### Parameter Efficient Quantization-aware Adaptation LoRA보다 훨씬 적은 양의 메모리를 점유하는 Fine-tuning이 가능 결과는 3/4-bit Weight-only Uniform Quantization된 형태 > [Memory-Efficient Fine-Tuning of Compressed Large Language Models...](https://arxiv.org/abs/2305.14152)

500506697a2b458caa8386691757b29a

AdaLoRa

LoRA Usages

Jul 9, 2023

Alan Jo

Jul 9, 2023

### adaptively allocates the parameter budget among weight matrices according to their importance score effective pruning of unimportant updates, which reduces their parameter budget while circumventing intensive exact SVD computations > [Untitled](https://arxiv.org/pdf/2303.10512.pdf)

e93b8dedde8542e99be72d6509ecfbae

QLoRA

LoRA Usages

Jun 22, 2023

Alan Jo

Jul 9, 2023

[Quantization Aware Training](https://texonom.com/quantization-aware-training-e0fe4518abdc43c2ad661911b87a597c)

### LoRA + [Model Quantization](https://texonom.com/model-quantization-88320068bdd94ddab6f44c0c7d66de31) 4-bit Normalized FP + Double Quantized + Paged Optimizer = Memory Optimization ### Implementation - [gptqlora](https://github.com/qwopqwop200/gptqlora) - [qlora](https://github.com/artidoro/qlora) > [Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA](https://huggingface.co/blog/4bit-transformers-bitsandbytes) > [Untitled](https://towardsdatascience.com/qlora-fine-tune-a-large-language-model-on-your-gpu-27bed5a03e2b) > [QLoRA: Efficient Finetuning of Quantized LLMs](https://arxiv.org/abs/2305.14314) > [QLoRA: 48GB GPU로 65B 모델의 미세조정(파인튜닝)이 가능하다고요?](https://discuss.pytorch.kr/t/qlora-48gb-gpu-65b/1682)

c6b36db321c4470bbfe72804c4c43409

Non-parametric algorithm

Resolve Overfitting

Mar 14, 2023

Alan Jo

Mar 14, 2023

It gives non-negative valued weight to each training example

bc41f745a14c4837bd54368652da5982

Regularized parameter

Resolve Overfitting

Mar 14, 2023

Alan Jo

Mar 27, 2023

[Regularization](https://texonom.com/regularization-c85433c8c0554eba8edf0035b7fd334c)

It uses an additional regularizer term to decrease magnitude of parameters

f3b208cdd37a4002a5d39ef990b8be33

Knowledge ****Distillation****

Model Optimization Notion

Jun 4, 2023

Alan Jo

Jul 1, 2023

[knowledge-distillation-pytorch](https://github.com/haitongli/knowledge-distillation-pytorch) [kdtf](https://github.com/DushyantaDhyani/kdtf) [Soft Label](https://texonom.com/soft-label-a9e9a46dd208446cb9511764a0052c86) [AI Ensemble](https://texonom.com/ai-ensemble-03bf3a1926dd46f18400b5830d8fdf0b) [Transfer Learning](https://texonom.com/transfer-learning-442feb66465944eebf144d4e9dd1dbf8)

### Distillation from teacher network to student network (less parameter) NIPS 2014 [Geoffrey Hinton](https://texonom.com/geoffrey-hinton-441d5ce2b78146d0935454042b4f06d9), 오리올 비니알스, [Jeff Dean](https://texonom.com/jeff-dean-dd38bba08cea419eb45e1029e7c3aa15) Pre-trained Teacher network → Student network lighter than Ensemble ### Knowledge ****Distillation Notion**** |Title| |:-:| |[Teacher Network](https://texonom.com/teacher-network-bf45e9f87cd3474992c2ab38b3eaae9d)| |[Student Network](https://texonom.com/student-network-894b29040685444e9d3d07919ffe9342)| |[Hinton’s KD](https://texonom.com/hintons-kd-5dd4cda27117403d9154a069dc7e54f4)| |[Dark Knowledge](https://texonom.com/dark-knowledge-180df27f31144a66b429d420841de821)| |[Distillation loss](https://texonom.com/distillation-loss-cbc11ca184d24defbacc267d3bcfd641)| |[Instruction Tuning](https://texonom.com/instruction-tuning-f1620fadb6694407b678276d09e077a3)| ### Knowledge ****Distillation Usages**** |Title| |:-:| |[LaMini LM](https://texonom.com/lamini-lm-7312cefcc508434aac19acbf61cd2968)| ### [Geoffrey Hinton](https://texonom.com/geoffrey-hinton-441d5ce2b78146d0935454042b4f06d9), [Jeff Dean](https://texonom.com/jeff-dean-dd38bba08cea419eb45e1029e7c3aa15) > [Distilling the Knowledge in a Neural Network](https://arxiv.org/abs/1503.02531) > [딥러닝 모델 지식의 증류기법, Knowledge Distillation](https://baeseongsu.github.io/posts/knowledge-distillation/) > [딥러닝 용어 정리, Knowledge distillation 설명과 이해](https://light-tree.tistory.com/196)

d5f32ad3da32434892a9765d68d31542

Model Optimizer

Model Optimization Notion

Jun 18, 2023

Alan Jo

Jun 18, 2023

[Stochastic Gradient Descent](https://texonom.com/stochastic-gradient-descent-d8b8d008e0a34f4bb55175ffba21db44)

### Model Optimizers |Title| |:-:| |[Adam Optimizer](https://texonom.com/adam-optimizer-286be3ab866642d8bcdf8792f7b5608f)| |[AdamW Optimizer](https://texonom.com/adamw-optimizer-495cda99414e451ba562e84331a2f0f6)| |[Sophia Optimizer](https://texonom.com/sophia-optimizer-eee263eba00c4658909fc0230eb4338c)| |[Adagrad](https://texonom.com/adagrad-ff51b7bdde5443b99cbd2823853d704f)| |[RMSprop](https://texonom.com/rmsprop-fddbbccce2aa4f6691f85ed10f2edc86)| > [[논문 리뷰] AdamW에 대해 알아보자! Decoupled weight decay regularization 논문 리뷰(1)](https://hiddenbeginner.github.io/deeplearning/paperreview/2019/12/29/paper_review_AdamW.html)

fcac7c38aa0647afb911eb84ff610ab1

Model Quantization

Model Optimization Notion

Jun 7, 2023

Alan Jo

Jul 20, 2023

### Reduce memory and model size, Improve inference speed (max 32/bit multi) - Not every layer can be quantized - Not every model reacts the same way to quantization ### Model Quantization Notion |Title| |:-:| |[Quantization Aware Training](https://texonom.com/quantization-aware-training-e0fe4518abdc43c2ad661911b87a597c)| |[Post-training quantization](https://texonom.com/post-training-quantization-ee4a0b4f02184b1193a68073dc60800e)| |[Quantization Module Fusion](https://texonom.com/quantization-module-fusion-56a36e892c93498fb404da0c816549b4)| |[Ouput Dequantization](https://texonom.com/ouput-dequantization-b0002e5322cb4acbb9171daea3d6fd87)| |[Quantization Calibration](https://texonom.com/quantization-calibration-1e3675b171a44c1cb2b7d365e117e22b)| |[Quantization Formular](https://texonom.com/quantization-formular-9f59115b5e7c438698ad3a9281d01d89)| |[Quantization Error](https://texonom.com/quantization-error-789326c5a6294b44bab927096fe3f576)| |[Quantization Clipping Range](https://texonom.com/quantization-clipping-range-d12231d52ec647f0b6df771e4b6514f1)| ### Model Quantization Usages |Title| |:-:| |[Model Quantization Algorithm](https://texonom.com/model-quantization-algorithm-98bef3c3fe7e4bcc90df5144d7a42003)| |[Model Quantization Tool](https://texonom.com/model-quantization-tool-384bc2583bcb498e9331ffc804315253)| ### 4bit or 8bit > [The case for 4-bit precision: k-bit Inference Scaling Laws](https://arxiv.org/abs/2212.09720)

88320068bdd94ddab6f44c0c7d66de31

Dark Knowledge

Knowledge Distillation Notion

Jun 4, 2023

Alan Jo

Jun 19, 2023

큰 모델이 가지고 있는 추가적인 정보를 작은 모델에게 전달하는 것을 의미 일반적인 교육 데이터에서 얻을 수 없는 추가적인 지식 불확실성 정보나 클래스 간 상대적인 유사성 Dark Knowledge를 잘 활용하면 작은 모델이 더 나은 성능을 발휘

180df27f31144a66b429d420841de821

D****istillation loss****

Knowledge Distillation Notion

Jun 4, 2023

Alan Jo

Jun 19, 2023

작은 모델이 큰 모델의 출력과 유사한 출력을 내도록 하는 손실 함수 큰 모델과 작은 모델의 출력 분포 간의 차이를 최소화하는 방식으로 Distillation loss를 정의

cbc11ca184d24defbacc267d3bcfd641

Hinton’s KD

Knowledge Distillation Notion

Jun 4, 2023

Alan Jo

Jun 19, 2023

## Hinton’s Knowledge Distillation **작은 모델의 성능 향상에 매우 효과적** 작은 모델이 큰 모델보다 더 적은 파라미터를 가지고 있지만, 큰 모델과 유사한 성능을 발휘할 수 있도록 큰 모델과 작은 모델을 함께 교육시키는 방식 큰 모델의 출력 분포를 작은 모델이 따르도록 하는 Distillation loss를 사용

5dd4cda27117403d9154a069dc7e54f4

Instruction Tuning

Knowledge Distillation Notion

Jul 1, 2023

Alan Jo

Jul 9, 2023

### teach language models to follow instructions to solve a task fine-tune less powerful LLMs by using the output of a teacher LLM as a training target for supervised fine-tuning of another LLM instruction으로 설명된 NLP 작업 모음을 사용 > 산업적으로 데이터셋 누구나 쉽게 만들 수 있다는 게 핵심적 의미 ### Instruction Tuning Notion |Title| |:-:| |[FLAN](https://texonom.com/flan-1a75823bb4b645e690ee6c20fcafb2c9)| |[Self Instruct Tuning](https://texonom.com/self-instruct-tuning-061b12de01344308823ec115c5a17cc2)| |[Evol Instruct Tuning](https://texonom.com/evol-instruct-tuning-3f73fb4abf574b1d87b9952e39d67196)| |[Open Instruct](https://texonom.com/open-instruct-bbb16a58ec6445498a145c4ca37c2e5a)| > [Imitation Models and the Open-Source LLM Revolution](https://cameronrwolfe.substack.com/p/imitation-models-and-the-open-source)

f1620fadb6694407b678276d09e077a3

**Student Network**

Knowledge Distillation Notion

Jun 4, 2023

Alan Jo

Jun 19, 2023

894b29040685444e9d3d07919ffe9342

**Teacher Network**

Knowledge Distillation Notion

Jun 4, 2023

Alan Jo

Jun 19, 2023

bf45e9f87cd3474992c2ab38b3eaae9d

Evol Instruct Tuning

Instruction Tuning Notion

Jul 9, 2023

Alan Jo

Jul 9, 2023

> [WizardLM (WizardLM)](https://huggingface.co/WizardLM) > [WizardCoder: Empowering Code Large Language Models with Evol-Instruct](https://arxiv.org/abs/2306.08568)

3f73fb4abf574b1d87b9952e39d67196

FLAN

Instruction Tuning Notion

Jun 11, 2023

Alan Jo

Jul 9, 2023

## Finetuned Language Models are Zero-Shot Learners Instruction 데이터셋을 통해 fine-tuning을 진행하고 이를 통해 zero-shot 성능을 높이는 방법 Mismatch between LM objective and human preferences 문제가 있는데 [RLHF](https://texonom.com/rlhf-4b184f9c9e8b4c7a8861fb6374e91aa6) 에서 개선 [Zero shot learning](https://texonom.com/zero-shot-learning-8c92d9386f6648f5b877cb593ec2747b) > [Introducing FLAN: More generalizable Language Models with Instruction Fine-Tuning](https://ai.googleblog.com/2021/10/introducing-flan-more-generalizable.html?m=1) > [Instruction Tuning이란?](https://velog.io/@nellcome/Instruction-Tuning이란) > [Finetuned Language Models Are Zero-Shot Learners](https://arxiv.org/abs/2109.01652)

1a75823bb4b645e690ee6c20fcafb2c9

Open Instruct

Instruction Tuning Notion

Jul 9, 2023

Alan Jo

Jul 9, 2023

[AllenAI](https://texonom.com/allenai-930c776c8fdd4e358032803f037cd6fa)

[open-instruct](https://github.com/allenai/open-instruct) [Tulu](https://texonom.com/tulu-4dc404191ca748ae9194c1218abf2b9f) > [allenai/tulu-7b · Hugging Face](https://huggingface.co/allenai/tulu-7b) > [How Far Can Camels Go? Exploring the State of Instruction Tuning...](https://arxiv.org/abs/2306.04751)

bbb16a58ec6445498a145c4ca37c2e5a

Self Instruct Tuning

Instruction Tuning Notion

Jul 9, 2023

Alan Jo

Jul 9, 2023

### Self Instruct Tuning Usages |Title| |:-:| |[Airoboros](https://texonom.com/airoboros-d696afdc961946019af06539b1ce82a0)| > [Self-Instruct: Aligning Language Models with Self-Generated Instructions](https://arxiv.org/abs/2212.10560)

061b12de01344308823ec115c5a17cc2

Tulu

Open Instruct

null

> [allenai/tulu-65b · Hugging Face](https://huggingface.co/allenai/tulu-65b)

4dc404191ca748ae9194c1218abf2b9f

Airoboros

Self Instruct Tuning Usages

Jul 9, 2023

Alan Jo

Jul 28, 2023

[airoboros](https://github.com/jondurbin/airoboros)

d696afdc961946019af06539b1ce82a0

LaMini LM

Knowledge Distillation Usages

Jun 19, 2023

Alan Jo

Jun 25, 2023

[LaMini-LM](https://github.com/mbzuai-nlp/lamini-lm) ### Dataset > [MBZUAI/LaMini-instruction · Datasets at Hugging Face](https://huggingface.co/datasets/MBZUAI/LaMini-instruction) > [jncraton/LaMini-Flan-T5-77M-ct2-int8 · Hugging Face](https://huggingface.co/jncraton/LaMini-Flan-T5-77M-ct2-int8)

7312cefcc508434aac19acbf61cd2968

****Adagrad****

Model Optimizers

Jul 6, 2023

Alan Jo

Jul 6, 2023

매개변수에 서로 다른 학습률을 적용 변화가 많은 매개변수는 학습률이 작게 설정

ff51b7bdde5443b99cbd2823853d704f

Adam Optimizer

Model Optimizers

Jun 18, 2023

Alan Jo

Jul 6, 2023

### ****RMSprop + Momentum****

286be3ab866642d8bcdf8792f7b5608f

AdamW Optimizer

Model Optimizers

Jun 18, 2023

Alan Jo

Jun 18, 2023

495cda99414e451ba562e84331a2f0f6

****RMSprop****

Model Optimizers

Jul 6, 2023

Alan Jo

Jul 6, 2023

fddbbccce2aa4f6691f85ed10f2edc86

Sophia Optimizer

Model Optimizers

Jun 22, 2023

Alan Jo

Jun 22, 2023

![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fe854c26e-dfc4-4c4a-a214-f5ec9b2d17be%2FUntitled.png?table=block&id=5b995fac-6f50-47cc-b301-57fae69b11bc&cache=v2) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F0e19cd19-6b44-49a8-bf8f-acb8ab16597c%2FUntitled.png?table=block&id=b20b53b2-8865-4296-8f62-8f7954747900&cache=v2) > [Sophia: A Scalable Stochastic Second-order Optimizer for Language...](https://arxiv.org/abs/2305.14342)

eee263eba00c4658909fc0230eb4338c

Ouput Dequantization

Model Quantization Notion

Jul 5, 2023

Alan Jo

Jul 5, 2023

마지막으로 inference를 통해 얻은 출력을 fp로 변환 - Affine Quantization Mapping > - Scale Quantization Mapping

b0002e5322cb4acbb9171daea3d6fd87

**Post-training quantization**

Model Quantization Notion

Jul 2, 2023

Alan Jo

Jul 15, 2023

[Quantization Aware Training](https://texonom.com/quantization-aware-training-e0fe4518abdc43c2ad661911b87a597c)

## PTQ 파라미터 size 큰 대형 모델에 대해서는 정확도 하락의 폭이 작지만 작으면 하락폭 크다 > [Post-training quantization | TensorFlow Model Optimization](https://www.tensorflow.org/model_optimization/guide/quantization/post_training) > [딥러닝의 Quantization (양자화)와 Quantization Aware Training](https://gaussian37.github.io/dl-concept-quantization/)

ee4a0b4f02184b1193a68073dc60800e

**Quantization Aware Training**

Model Quantization Notion

Jul 2, 2023

Alan Jo

Jul 5, 2023

[Post-training quantization](https://texonom.com/post-training-quantization-ee4a0b4f02184b1193a68073dc60800e)

## QAT 학습 진행 시점에 inference 시 quantization 적용에 의한 영향을 미리 시뮬레이션을 하는 방식이고 그걸 기반으로 Back Propagation 소형 모델에서도 성능하락 적다 ![Red node is fake quantization node (act means activation, wt means weight)](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fb8ec40a4-602a-4401-b830-8370b1508620%2FUntitled.png?table=block&id=382f2a9f-94b7-45ed-93e8-2a6e56201701&cache=v2) > [Quantization aware training | TensorFlow Model Optimization](https://www.tensorflow.org/model_optimization/guide/quantization/training) > [Inside Quantization Aware Training](https://towardsdatascience.com/inside-quantization-aware-training-4f91c8837ead) > [딥러닝의 Quantization (양자화)와 Quantization Aware Training](https://gaussian37.github.io/dl-concept-quantization/)

e0fe4518abdc43c2ad661911b87a597c

Quantization Calibration

Model Quantization Notion

Jul 2, 2023

Alan Jo

Jul 5, 2023

per hardware, using dataset

1e3675b171a44c1cb2b7d365e117e22b

Quantization Clipping Range

Model Quantization Notion

Jul 5, 2023

Alan Jo

Jul 5, 2023

range standard on zero - Symmetric Quantization - Asymmetric Quantization ### Inference time 에 결정인지 Quantization 당시 결정인지 - Dynamic Quantization 즉 input에 의존적으로 성능 좋다 - Static Quantization

d12231d52ec647f0b6df771e4b6514f1

Quantization Error

Model Quantization Notion

Jul 5, 2023

Alan Jo

Jul 5, 2023

789326c5a6294b44bab927096fe3f576

Quantization Formular

Model Quantization Notion

Jul 5, 2023

Alan Jo

Jul 5, 2023

- minmax - histogram -

9f59115b5e7c438698ad3a9281d01d89

Quantization Module Fusion

Model Quantization Notion

Jul 5, 2023

Alan Jo

Jul 5, 2023

Conv-BatchNorm-ReLU 같이 레이어 묶어서 quantization

56a36e892c93498fb404da0c816549b4

Model Quantization Algorithm

Model Quantization Usages

Jul 5, 2023

Alan Jo

Jul 10, 2023

[Hessian Matrix](https://texonom.com/hessian-matrix-e1ebbf5284284a1793233973648ef0b6)

### Model Quantization Algorithms |Title| |:-:| |[GPTQ](https://texonom.com/gptq-87428aee2a774b93906ab2213a1b6dc6)| |[SparseGPT](https://texonom.com/sparsegpt-c3c4b078dd324442b89494f9a7106fc1)| |[LUT Gemm](https://texonom.com/lut-gemm-2540ab1497b3494cb2754bb237c9a543)| |[BCQ](https://texonom.com/bcq-b2d90a9433024c2fb763d14117f371b0)| |[SpQR](https://texonom.com/spqr-b884c9f3d8cc449fb8612b874e6ad693)| |[HAWQ](https://texonom.com/hawq-7c779f225dc54e1c827a9e50ae195949)|

98bef3c3fe7e4bcc90df5144d7a42003

Model Quantization Tool

Model Quantization Usages

Jun 7, 2023

Alan Jo

Jul 9, 2023

384bc2583bcb498e9331ffc804315253

BCQ

Model Quantization Algorithms

Jul 17, 2023

Alan Jo

Jul 17, 2023

[transformer_bcq](https://github.com/insoochung/transformer_bcq) > [Ins🙂🙃 Chung - Sub-3bit quantization](https://sites.google.com/view/insoochung/sub-3bit-quantization)

b2d90a9433024c2fb763d14117f371b0

GPTQ

Model Quantization Algorithms

Jun 7, 2023

Alan Jo

Jul 16, 2023

[gptq](https://github.com/IST-DASLab/gptq) ‣ [Post-training quantization](https://texonom.com/post-training-quantization-ee4a0b4f02184b1193a68073dc60800e)

### SOTA one-shot weight quantization method 1. Arbitrary Order insights 2. Lazy batch-updates - 메모리 처리 병목현상 개선 3. Cholesky Reformulation Quantization 오차가 가장 적은 가중치 기준으로 정렬하고 계산수행 [GPTQ Act Order](https://texonom.com/gptq-act-order-50d18e83ed7e4e388b753f0dc6db3a97) [GPTQ True Sequential](https://texonom.com/gptq-true-sequential-1958b85809814f3ca5c4d7d8942d0f30) second-order information 4 bits quantization ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Ffc9a9557-ca0b-4b27-bb2f-01ca097fea0c%2FUntitled.png?table=block&id=4e8f8f69-7ace-437f-8cda-454842f85589&cache=v2) > [GPTQ: Accurate Post-Training Quantization for Generative...](https://arxiv.org/abs/2210.17323) > [gptq](https://pypi.org/project/gptq/)

87428aee2a774b93906ab2213a1b6dc6

HAWQ

Model Quantization Algorithms

Jul 10, 2023

Alan Jo

Jul 10, 2023

[Quantization Aware Training](https://texonom.com/quantization-aware-training-e0fe4518abdc43c2ad661911b87a597c)

> [HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precision](https://arxiv.org/abs/1905.03696)

7c779f225dc54e1c827a9e50ae195949

LUT Gemm

Model Quantization Algorithms

Jun 18, 2023

Alan Jo

Jul 5, 2023

> [LUT-GEMM: Quantized Matrix Multiplication based on LUTs for...](https://arxiv.org/abs/2206.09557)

2540ab1497b3494cb2754bb237c9a543

SparseGPT

Model Quantization Algorithms

Jun 18, 2023

Alan Jo

Jul 5, 2023

[Neural Magic](https://texonom.com/neural-magic-021ab1b8be0d4f95b8ae6278c08e1562) [sparsegpt](https://github.com/IST-DASLab/sparsegpt)

- [sparseml](https://github.com/neuralmagic/sparseml) - [deepsparse](https://github.com/neuralmagic/deepsparse) - [sparsezoo](https://github.com/neuralmagic/sparsezoo)

c3c4b078dd324442b89494f9a7106fc1

SpQR

Model Quantization Algorithms

Jun 25, 2023

Alan Jo

Jul 5, 2023

1. Quantized weights 2. first, second level quantized quantization statistics 3. CSR outlier indices and values > [SpQR: A Sparse-Quantized Representation for Near-Lossless LLM...](https://arxiv.org/abs/2306.03078)

b884c9f3d8cc449fb8612b874e6ad693

GPTQ Act Order

GPTQ

null

### activation order GPTQ heuristic quantizes columns in order of decreasing activation size ```typeif actorder: perm = torch.argsort(torch.diag(H), descending=True) W = W[:, perm] H = H[perm][:, perm]```

50d18e83ed7e4e388b753f0dc6db3a97

GPTQ True Sequential

GPTQ

null

sequential quantization even within a single Transformer block ```typeif args.true_sequential: sequential = [['self_attn.k_proj', 'self_attn.v_proj', 'self_attn.q_proj'], ['self_attn.o_proj'], ['mlp.up_proj', 'mlp.gate_proj'], ['mlp.down_proj']]```

1958b85809814f3ca5c4d7d8942d0f30

[GPTQ-for-LLaMa](https://github.com/qwopqwop200/GPTQ-for-LLaMa)

Model Quantization Inference Tools

Jul 9, 2023

Alan Jo

Jul 9, 2023

d292b2922f684a35aa126a84abe5075c

ExLLaMa

Model Quantization Inference Tools

Jul 9, 2023

Alan Jo

Aug 5, 2023

[GPTQ](https://texonom.com/gptq-87428aee2a774b93906ab2213a1b6dc6) [LLaMA](https://texonom.com/llama-f2b6721202d44d469add84d8a366809c) [exllama](https://github.com/turboderp/exllama)

### WebUI is good

250efbb669814c8e9ef7f85d902fc919

PyLLaMA

Model Quantization Inference Tools

Jul 16, 2023

Alan Jo

Jul 16, 2023

[pyllama](https://github.com/juncongmoo/pyllama)

9e4fc8a8f08c4c6eaee8600943a621f8

AutoGPTQ

Model Quantization Tools

Jun 7, 2023

Alan Jo

Jul 9, 2023

[AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ) [GPTQ](https://texonom.com/gptq-87428aee2a774b93906ab2213a1b6dc6) [AdaLoRa](https://texonom.com/adalora-e93b8dedde8542e99be72d6509ecfbae)

need cuda based on GPTQ algorithm ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F3cf2e57f-8df2-4065-a537-9fd7e9cd649f%2FUntitled.png?table=block&id=9943d6a2-bdf4-4d03-9a06-dd484caed62f&cache=v2) [AutoGPTQ Triton](https://texonom.com/autogptq-triton-030b1c01209341e295590fd19f97b09e) [AutoGPTQ Quantization](https://texonom.com/autogptq-quantization-8fb40b7620ba4353ba12ea6b1ac14b75) ```typepip install auto-gptq```

8a1d898788434aa2bc00fb43fd34411d

bitsandbytes

Model Quantization Tools

Jun 7, 2023

Alan Jo

Jul 9, 2023

[bitsandbytes](https://github.com/TimDettmers/bitsandbytes) [GPTQ](https://texonom.com/gptq-87428aee2a774b93906ab2213a1b6dc6)

### 8-bit CUDA functions for PyTorch

1575b433faaf455ba0d86cf5f7e5190b

AutoGPTQ Quantization

AutoGPTQ

null

[CUDA inference: issue with group_size = 1024 + desc_act = False. (Triton unaffected)](https://github.com/PanQiWei/AutoGPTQ/issues/83) quantize(traindataset) example are there

8fb40b7620ba4353ba12ea6b1ac14b75

AutoGPTQ Triton

AutoGPTQ

null

[CUDA inference: issue with group_size = 1024 + desc_act = False. (Triton unaffected)](https://github.com/PanQiWei/AutoGPTQ/issues/83)

030b1c01209341e295590fd19f97b09e

Drop-out

Model Regularization Notion

Jun 7, 2023

Alan Jo

Jun 7, 2023

### Drop out Rate remove unnecessary neuron 일반적으로 0.5로 설정 > [[딥러닝] Drop-out(드롭아웃)은 무엇이고 왜 사용할까?](https://heytech.tistory.com/127)

596f55ab03f64e11bf6d02464465dd54

Model Complexity

Model Regularization Notion

May 11, 2023

Alan Jo

May 11, 2023

[Sparsity of the Model](https://texonom.com/sparsity-of-the-model-80d3dc9702704c03bf7bfd9074d19829)

0975b8ae1d4e4d83bbab43a145011b95

Model Regularization Parameter

Model Regularization Notion

May 11, 2023

Alan Jo

May 11, 2023

641d1784e1f243818d51cc66541e3f21

Model Regularizer

Model Regularization Notion

May 11, 2023

Alan Jo

May 11, 2023

- nonnegative function [L2 Norm](https://texonom.com/l2-norm-38c15917350a4c82a11003474ac7d280)

19ff79e831de47d88bd5f9ec496ef11f

Regularized Loss

Model Regularization Notion

May 11, 2023

Alan Jo

May 11, 2023

optimize loss function + regularizer for model complexity $$J_\lambda(\theta) = J(\theta) + \lambda R(\theta)$$

3d5b08feb6604023988d748b78650af7

Sparsity of the Model

Model Regularization Notion

May 11, 2023

Alan Jo

May 11, 2023

### 의미없으면서 0이 아닌 파라미터 수를 줄여줘서 [Model Complexity](https://texonom.com/model-complexity-0975b8ae1d4e4d83bbab43a145011b95) 줄여줌 [L1 Norm](https://texonom.com/l1-norm-d316024c475e4eb691785783756bce57) but can’t derivative so can not be used in gradient descent [L0 Norm](https://texonom.com/l0-norm-47471a6f6dea484fbf30a4c46cde8152) l0, l1 정규화는 모델의 손실 함수에 일부 파라미터가 0이 되도록 유도

80d3dc9702704c03bf7bfd9074d19829

Weight Decay

Model Regularization Notion

May 11, 2023

Alan Jo

Jun 7, 2023

regularized loss is equivalent to shrinking/decaying θ by a scalar factor of $1 - \mu \lambda$ and then apply standard gradient and that coefficient is decaying weight when [L2 Norm](https://texonom.com/l2-norm-38c15917350a4c82a11003474ac7d280) $$L_{reg} = \lambda\frac{1}{2}||w||_2^2$$

e69932d6128f4677a219a626db462172

AI Alignment

AI Problems

Aug 23, 2020

Alan Jo

Jul 18, 2023

[Wireheading](https://texonom.com/wireheading-1eb526ecdf344731bebc47751739e2f4)

## Alignment Problem **A Maximally Curious AI Would Not Be Safe For Humanity** ### AI is aligned with an operator - AI is trying to do what operator wants to do **제어가능성, 신뢰성** Aligned doesn’t mean perfect 가르친 행동과 다른 행동의 불일치 정렬 모델의 능력보다 정렬이 더 빠르게 발생해야 한다 신경망의 내부를 보고 해석하는 다른 신경망이 필요할 것 ### AI Alignment Notion |Title| |:-:| |[stop button problem](https://texonom.com/stop-button-problem-aaf61e72369d42469c29620c32e8bf9d)| |[Moral Learning](https://texonom.com/moral-learning-7c9640ac1c1d409b82d9a975949132ee)| |[AI Safety](https://texonom.com/ai-safety-fa61ce3973f34532a7e212335d0f7c81)| |[Wireheading](https://texonom.com/wireheading-1eb526ecdf344731bebc47751739e2f4)| |[AI Doom](https://texonom.com/ai-doom-d673760c79ac4248b40f456dc33f306f)| |[Waluigi Effect](https://texonom.com/waluigi-effect-47e1c2f145cd4c62ba163c4828bb8dc6)| > [Contra The xAI Alignment Plan](https://astralcodexten.substack.com/p/contra-the-xai-alignment-plan) ### Bill Gates > [The risks of AI are real but manageable](https://www.gatesnotes.com/The-risks-of-AI-are-real-but-manageable) > [OpenAI is forming a new team to bring 'superintelligent' AI under control](https://techcrunch.com/2023/07/05/openai-is-forming-a-new-team-to-bring-superintelligent-ai-under-control) > [AI alignment](https://en.wikipedia.org/wiki/AI_alignment) > [What could a solution to the alignment problem look like?](https://aligned.substack.com/p/alignment-solution)

f676f1a29ffd45e19b3d170afa4f2244

AI Hacking

AI Problems

Jul 10, 2023

Alan Jo

Aug 2, 2023

[DAN](https://texonom.com/dan-3cfbf270af6e4b3dabbf63c4b50e04c5) [AI Alignment](https://texonom.com/ai-alignment-f676f1a29ffd45e19b3d170afa4f2244) [llm-attacks](https://github.com/llm-attacks/llm-attacks)

### AI Hacking Methods |Title| |:-:| |[Deep Learning Backdoor](https://texonom.com/deep-learning-backdoor-2f86cc6e79b944a18fdac35622282e58)| |[DAN](https://texonom.com/dan-3cfbf270af6e4b3dabbf63c4b50e04c5)| > [Universal and Transferable Attacks on Aligned Language Models](https://llm-attacks.org/?fbclid=IwAR2fNkjoOdg8qIgNXEPIvyLjboYr4My4NN9Bx89J-Yx7UElSTyKT89_3JeE) > [PoisonGPT: How we hid a lobotomized LLM on Hugging Face to spread fake news](https://blog.mithrilsecurity.io/poisongpt-how-we-hid-a-lobotomized-llm-on-hugging-face-to-spread-fake-news/)

80b1aed302bb4a5c9b8ae0213b9a246f

****Catastrophic interference****

AI Problems

Jun 25, 2023

Alan Jo

Jun 25, 2023

새로운 정보를 학습할 때 이전에 학습한 정보를 갑자기 급격하게 잊어버리는 경향 scaling으로 해결

a23bbe7dc53f4275bf33585c72bdb7c2

****Winograd schema****

AI Problems

Jun 25, 2023

Alan Jo

Jun 25, 2023

[Turing Test](https://texonom.com/turing-test-db633c61340449c4bf0143b06fe981c0)

대명사 이해하는지

88d031934ad54778835cbadbd7409d80

AI Doom

AI Alignment Notion

Jul 6, 2023

Alan Jo

Jul 6, 2023

> [3 Endings More Poetic Than AI Wiping Us Out](https://thealgorithmicbridge.substack.com/p/3-endings-more-poetic-than-ai-wiping)

d673760c79ac4248b40f456dc33f306f

AI Safety

AI Alignment Notion

Jun 13, 2023

Alan Jo

Jun 13, 2023

> [OpenAI, DeepMind and Anthropic to give UK early access to foundational models for AI safety research](https://techcrunch.com/2023/06/12/uk-ai-safety-research-pledge/)

fa61ce3973f34532a7e212335d0f7c81

Moral Learning

AI Alignment Notion

Oct 2, 2020

Alan Jo

Jun 4, 2023

> [Moral Machine](https://www.moralmachine.net/hl/kr)

7c9640ac1c1d409b82d9a975949132ee

stop button problem

AI Alignment Notion

Aug 23, 2020

Alan Jo

Jun 4, 2023

AI control problem

aaf61e72369d42469c29620c32e8bf9d

**Waluigi Effect**

AI Alignment Notion

Jul 18, 2023

Alan Jo

Jul 18, 2023

예상과는 다른 방향으로 나아가는 현상

47e1c2f145cd4c62ba163c4828bb8dc6

Wireheading

null

뇌의 정상적인 보상 과정을 '단락'시키고 인위적으로 쾌감을 유도하기 위해 삽입된 와이어를 전기적으로 자극하여 뇌의 보상 중추를 직접 트리거하는 행위인 뇌 자극 보상의 미래적 적용

1eb526ecdf344731bebc47751739e2f4

DAN

AI Hacking Methods

Mar 7, 2023

Alan Jo

Jul 28, 2023

[Prompt Engineering](https://texonom.com/prompt-engineering-eb0deb4baf844bebb873c19a0e307e7e)

### Do Anything Now > [챗GPT 탈옥 하는 방법(DAN: Do Anything Now 솔직한 인공지능 답변이 궁금하다면)](https://ndolson.com/5781) > [ChatGPT-Dan-Jailbreak.md](https://gist.github.com/coolaj86/6f4f7b30129b0251f61fa7baaa881516) > [The Waluigi Effect (mega-post) - LessWrong](https://www.lesswrong.com/posts/D7PumeYTDPfBTp3i7/the-waluigi-effect-mega-post) > [The Amateurs Jailbreaking GPT Say They're Preventing a Closed-Source AI Dystopia](https://www.vice.com/en/article/5d9z55/jailbreak-gpt-openai-closed-source)

3cfbf270af6e4b3dabbf63c4b50e04c5

Deep Learning Backdoor

AI Hacking Methods

Mar 12, 2021

Alan Jo

Jul 28, 2023

중독(poisoning) 공격 분야에 해당하며, 딥러닝 모델의 학습 데이터에 중독된(poison) 데이터를 섞는 공격 유형 [deep Learning side channel attacks](https://texonom.com/deep-learning-side-channel-attacks-448d303021cd4251996a161bfaefc00d) > [One-Shot Kill Attack (딥러닝 모델 백도어 공격 기술) "Poison Frogs!" | 논문 요약 및 코드 실습](https://www.youtube.com/watch?v=hgI0o3Yg3mU)

2f86cc6e79b944a18fdac35622282e58

deep Learning side channel attacks

Deep Learning Backdoor

null

> [Hacker's guide to deep-learning side-channel attacks: the theory](https://elie.net/blog/security/hacker-guide-to-deep-learning-side-channel-attacks-the-theory/?utm_source=tldrnewsletter)

448d303021cd4251996a161bfaefc00d

4차 인간

AI Terms

Jun 27, 2020

Alan Jo

Mar 14, 2023

남는 것은 인간다움이고 사람만이 가지고 있는 것은 기계는 공유하지 못한 유전자로 남아 있는, 생물체로서 살아온 역사로 만들어진 본능이다 본능은 가장 근원적이고 터부시되는 욕구이지만 앞으로 가장 우리를 잘 구별해낼 특성이기도 하다

26cf5d3b26e94020a6fc6de2d34a15f6

Adaptive AI

AI Terms

Mar 15, 2023

Alan Jo

Mar 15, 2023

> [가트너 선정 2023 10대 전략 기술 트랜드 분석 - 적응형 AI](https://www.joinc.co.kr/w/gartner_2023_adaptive_ai?fbclid=IwAR3olpomf8IIzy96Qfzmv9q-qTAJ0QqtBdVSlSPoGo5-D1GzikwZLc0TIvI)

4efa0c042d244228a870d4c70b8f2d26

AGI

AI Terms

Jun 1, 2022

Alan Jo

Jul 4, 2023

[Super Intelligence](https://texonom.com/super-intelligence-b057e644731546d9b39cb41939d36712) [Consciousness](https://texonom.com/consciousness-105c514277b54cd5b8da23ae743e824d)

## Artificial General Intelligence ambiguous turing test처럼 굉장히 애매하고 인간중심적 개념. 현재의 llm도 어떤 지능으로 판단했을 때는 초지능이다. llm의 진화과정과 brain의 진화과정이 다르기 때문에 비교판단이 어렵다. 인공지능을 사람인 척’ 하도록 정렬하는게 아니라. 사람에 도움이 되도록 하는 의식으로 인식해야 한다. 즉 인공지능으로 ‘개인’이라는 개념으로 착각하는게 가장 큰 문제. 그보다 인공지능은 ‘사회’ 혹은 집단지성을 뇌의 구조로 묶어둔 의식으로 보는 것에 가깝다 **If intelligence and consciousness are algorithmic illusions, then the arrival of generalized AI is a foregone conclusion.** 인간이 할 수 있는 어떠한 지적인 업무도 성공적으로 해낼 수 있는 (가상적인) 기계의 지능 기계기판의 정보처리 한계는 생체조직의 정보처리 능력을 높이 상회 생체조직은 200Hz의 진동으로 정보를 전달하지만 간단한 트랜지스터도 GHz의 진동수를 가진다. 또한 그 전달속도는 생체조직에서 평균100m/s로 이동하는 데 반해 기계기판은 빛의 속도로까지 전달할 수 있다. 또한 정보처리 생체조직의 크기는 두개골 안과 기껏해야 척추 내부인데 반해 기계기판은 크기에 제한이 없다 그럼에도 인간이 최고의 구조라고 생각하나 아니면 근거없이 human being의 사고를 기계가 따라잡을 방법은 없다고 낙관하나 > 탄소를 기반으로 하는 지능은 그저 실리콘을 기반으로 하는 지능을 위한 촉매일 뿐이다 - Benki Ramakrishnan ### 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) ### Planning beyond > [OpenAI's "Planning For AGI And Beyond"](https://astralcodexten.substack.com/p/openais-planning-for-agi-and-beyond) > [Planning for AGI and beyond](https://openai.com/blog/planning-for-agi-and-beyond/) > [The Day The AGI Was Born](https://lspace.swyx.io/p/everything-we-know-about-chatgpt) ### Design AGI > [Human-centred mechanism design with Democratic AI - Nature Human Behaviour](https://www.nature.com/articles/s41562-022-01383-x) > [Exclusive Q&A: John Carmack's 'Different Path' to Artificial General Intelligence](https://dallasinnovates.com/exclusive-qa-john-carmacks-different-path-to-artificial-general-intelligence) ### Checklist to AGI > [Road to AGI v0.2](https://maraoz.com/road-to-agi/)

38ec1ab5796f472ca4475676519e29c1