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
StarChat	Huggingface H4	null	null	null	null	null	> [HuggingFaceH4/starchat-beta · Hugging Face](https://huggingface.co/HuggingFaceH4/starchat-beta) > [TheBloke/starchat-beta-GPTQ at main](https://huggingface.co/TheBloke/starchat-beta-GPTQ/tree/main)	ea316ec509564a51b56ad92b22831220
HuggingFace Docker Space	HuggingFace Space SDK	Jul 13, 2023	Alan Jo	Alan Jo	Jul 13, 2023		### HuggingFace Docker Space Templates \|Title\| \|:-:\| \|[HuggingFace ChatUI Template](https://texonom.com/huggingface-chatui-template-9cd1b4ab300f44b0bc91c87f33edf6bb)\|	0f250609ad1f493bb91146813de8d8a6
HuggingFace ChatUI Template	HuggingFace Docker Space Templates	Jul 13, 2023	Alan Jo	Alan Jo	Jul 13, 2023			9cd1b4ab300f44b0bc91c87f33edf6bb
fastbook	Deep Learning Tools	Nov 25, 2019	Alan Jo	Alan Jo	May 29, 2023		> [fastai/fastbook](https://github.com/fastai/fastbook)	d022baa785a14ce5af5cf1ef59995cb1
Netron	Deep Learning Tools	Oct 6, 2021	Alan Jo	Alan Jo	May 29, 2023		### visualize Neural Network > [lutzroeder/netron](https://github.com/lutzroeder/netron)	2e410ef29aaa4a7d9e2a2777ce4dd3ee
Activation Function	Neural Network Components	Apr 27, 2023	Alan Jo	Alan Jo	Jun 6, 2023		### Unlike sign function for [Back Propagation](https://texonom.com/back-propagation-18f4493692ad43449d4271f1bb293781), use non zero derivative - sigmoid function - probability - sign function - predict class ### Activation Functions \|Title\| \|:-:\| \|[Sigmoid Function](https://texonom.com/sigmoid-function-417e17b995cf40c996cec78f699ea59b)\| \|[GELU](https://texonom.com/gelu-8f18023a80b84a06b5afd648811df6bc)\| \|[ReLU](https://texonom.com/relu-e582549804da48b893758895e446ffb9)\| \|[Sign Function](https://texonom.com/sign-function-b95e191cc2dc43c0befd6231f022555f)\| \|[ELU Function](https://texonom.com/elu-function-c7306a8566014b4e9d0733b52bd399fa)\| \|[Mish Function](https://texonom.com/mish-function-39eaae692b284adbadde08d3d707f13d)\| \|[Tanh Function](https://texonom.com/tanh-function-679abf35c4eb4fb0bb0e56140cac3d23)\| \|[Maxout Function](https://texonom.com/maxout-function-ad1f081fa0c34a6fbee08259c82ea4a1)\| \|[ELU Function](https://texonom.com/elu-function-e6f30f8fff2c422993a85565f8ce5ee7)\| > [0023 Loss & Metric - Deepest Documentation](https://deepestdocs.readthedocs.io/en/latest/002_deep_learning_part_1/0023/)	8e52ee5f83a244d88abeeee3fb9497a8
Forward Forward Algorithm	Neural Network Components	Apr 3, 2023	Alan Jo	Alan Jo	May 11, 2023		[Back Propagation](https://texonom.com/back-propagation-18f4493692ad43449d4271f1bb293781) 보다 대뇌피질 작동방식에 유사 시냅스의 처음으로 돌아가기 전에 cortical layer를 지나는 과정에서 모종의 루프를 형성한다 네트워크의 레이어 수가 많아질수록 FF가 Backprop보다 메모리를 덜 사용합니다. 하지만 20개 이하의 레이어를 가진 thin network에서는 FF가 backprop보다 훨씬 더 많은 메모리를 사용 ### Forward Forward Algorithm Notion \|Title\| \|:-:\| > [[딥러닝] The Forward-Forward Algorithm: Some Preliminary Investigations (FF)](https://velog.io/@nochesita/딥러닝-The-Forward-Forward-Algorithm-Some-Preliminary-Investigations) > [Welcome to JunYoung's blog \| 딥러닝의 체계를 바꾼다! The Forward-Forward Algorithm 논문 리뷰 (1)](https://junia3.github.io/blog/ffalgorithm) > [The Forward-Forward Algorithm: Some Preliminary Investigations](https://arxiv.org/abs/2212.13345)	6c989313e382466bba02d15c412fb17f
Neural Network Layer	Neural Network Components	Apr 27, 2023	Alan Jo	Alan Jo	Jun 7, 2023		### Do not count input layer as perceptron (no weights & bias) $$ learnable\;parameters = weights(edges) + perceptron(bias)$$ ### Neural Network Layer Notion \|Title\| \|:-:\| \|[Layer Normalization](https://texonom.com/layer-normalization-c5111411c0594c04847b0eb4d8fa6682)\| \|[Batch Normalization](https://texonom.com/batch-normalization-0e5e919e98b74b168a248a9fd3e0cb63)\| \|[Fully Connected Layer](https://texonom.com/fully-connected-layer-3fd5a241e2b84a53a2339c60fbcd957a)\| \|[Input Layer](https://texonom.com/input-layer-abdd1cda0d124843a34c66dc5ec7eba1)\| \|[Hidden Layer](https://texonom.com/hidden-layer-a22cafa21026469094d260b473fd4789)\| \|[Output Layer](https://texonom.com/output-layer-184eb47c3a98437daceba11104377194)\| \|[Residual Connection](https://texonom.com/residual-connection-f6778b3cfbee4b0cab02f7fad533372f)\|	e10ef2afe1954cf6b909f8aa40077393
skip connections	Neural Network Components	May 23, 2023	Alan Jo	Alan Jo	May 23, 2023		help dampen the impact of poor layer	d7ad187f3487468db6eea278f0236a22
ELU Function	Activation Functions	Apr 27, 2023	Alan Jo	Alan Jo	Apr 27, 2023			c7306a8566014b4e9d0733b52bd399fa
ELU Function	Activation Functions	Jun 7, 2023	Alan Jo	Alan Jo	Jun 7, 2023			e6f30f8fff2c422993a85565f8ce5ee7
GELU	Activation Functions	Apr 8, 2023	Alan Jo	Alan Jo	Apr 27, 2023			8f18023a80b84a06b5afd648811df6bc
Maxout Function	Activation Functions	Jun 7, 2023	Alan Jo	Alan Jo	Jun 7, 2023			ad1f081fa0c34a6fbee08259c82ea4a1
Mish Function	Activation Functions	May 18, 2023	Alan Jo	Alan Jo	May 18, 2023			39eaae692b284adbadde08d3d707f13d
ReLU	Activation Functions	Mar 7, 2023	Alan Jo	Alan Jo	Jun 7, 2023		## Rectified Linear Unit 입력값이 음수인 뉴런은 다시 회생시키기 어렵다는 한계 이러한 문제를 Dying ReLU ### ReLU Variants \|Title\| \|:-:\| \|[PReLU](https://texonom.com/prelu-67cdafbec19248bfad4225946a4c9fc3)\| \|[Leaky ReLU](https://texonom.com/leaky-relu-fa25fa8a234044068fd839a93512a762)\|	e582549804da48b893758895e446ffb9
Sigmoid Function	Activation Functions	Nov 19, 2020	Alan Jo	Alan Jo	Jun 14, 2023		### inverse of the natural logit function logistic은 s자 모양을 의미하는 시그모이드중 대표적 ReLU가 등장하기 이전에 활발하게 사용되었던 activation function hidden 노드 바로 뒤에 부착 ### Sigmoid Functions \|Title\| \|:-:\| \|[SoftMax Function](https://texonom.com/softmax-function-d8364b5e839145e89df423e0c5365dc1)\| \|[Logit Function](https://texonom.com/logit-function-104e55a12b234813a020cb09ec51c754)\| $$\sigma(x) = \frac{1}{1 + e^{-(wx + b)}}$$ > [어떻게 하면 "부드럽게" 연결할까?](https://youtube.com/watch?v=LS-C8KBvJYY&feature=shares) > [Logistic function](https://en.wikipedia.org/wiki/Logistic_function) > [logit, sigmoid, softmax의 관계 - 한 페이지 머신러닝](https://opentutorials.org/module/3653/22995)	417e17b995cf40c996cec78f699ea59b
Sign Function	Activation Functions	Apr 27, 2023	Alan Jo	Alan Jo	Apr 27, 2023			b95e191cc2dc43c0befd6231f022555f
Tanh Function	Activation Functions	May 23, 2023	Alan Jo	Alan Jo	May 23, 2023			679abf35c4eb4fb0bb0e56140cac3d23
Leaky ReLU	ReLU Variants	Apr 27, 2023	Alan Jo	Alan Jo	May 23, 2023		음수일 때 출력값을 0이 아닌 0.001과 같은 매우 작은 값을 출력하도록 설정	fa25fa8a234044068fd839a93512a762
PReLU	ReLU Variants	May 18, 2023	Alan Jo	Alan Jo	May 18, 2023			67cdafbec19248bfad4225946a4c9fc3
Logit Function	Sigmoid Functions	Nov 1, 2020	Alan Jo	Alan Jo	Apr 27, 2023		$$g(z) = \frac{1}{1 + e^{-z}}$$ ### This is important $$g'(z) = g(z)(1 - g(z))$$ [Logit](https://texonom.com/logit-84e304d3bf104e3fbc6029bcc7eea6c2)	104e55a12b234813a020cb09ec51c754
SoftMax Function	Sigmoid Functions	Nov 1, 2020	Alan Jo	Alan Jo	Apr 27, 2023	[Cross Entropy](https://texonom.com/cross-entropy-fefc71fd930a41a7842f39bccb3abcc9)	### softmax함수는 sigmoid의 일반형 - 2개 클래스를 대상으로 정의하던 logit을 K개의 클래스를 대상으로 일반화하면 softmax함수가 유도(derived) - softmax함수에서 K를 2로 놓으면 sigmoid함수로 환원 - sigmoid함수를 K개의 클래스를 대상으로 일반화하면 softmax함수가 유도 ### classification softmax함수는 인공신경망이 내놓은 K개의 클래스 구분 결과를 확률처럼 해석하도록 sigmoid는 activation에, softmax는 classification에 사용되지만 수학적으로 같다 > 다루는 클래스가 2개냐 K개냐로 차이가 있을 뿐 > [logit, sigmoid, softmax의 관계 - 한 페이지 머신러닝](https://opentutorials.org/module/3653/22995)	d8364b5e839145e89df423e0c5365dc1
Logit	Logit Function	null	null	null	null	null	‘logistic’ +‎ ‘probit’ > [logit, sigmoid, softmax의 관계 - 한 페이지 머신러닝](https://opentutorials.org/module/3653/22995)	84e304d3bf104e3fbc6029bcc7eea6c2
Batch Normalization	Neural Network Layer Notion	Mar 7, 2023	Alan Jo	Alan Jo	Jul 6, 2023	[Layer Normalization](https://texonom.com/layer-normalization-c5111411c0594c04847b0eb4d8fa6682)	레이어마다 Normalization을 하는 레이어를 두어, 변형된 분포가 나오지 않도록 하는 것 ### Limitation 1. dependent to mini batch size 2. hard to apply to RNN ### Batch Normalization Methods \|Title\| \|:-:\| \|[Internal Covariate Shift](https://texonom.com/internalcovariate-shift-827b94e336a141b4a2ca100fc42dab66)\| \|[Whitening](https://texonom.com/whitening-5bfaed3572904c208125cc9df6c664b0)\| > [배치 정규화(Batch Normalization)](https://gaussian37.github.io/dl-concept-batchnorm/) > [Batch normalization](https://en.wikipedia.org/wiki/Batch_normalization)	0e5e919e98b74b168a248a9fd3e0cb63
Fully Connected Layer	Neural Network Layer Notion	Apr 27, 2023	Alan Jo	Alan Jo	Jun 1, 2023		## FC Layer input node is connected to each input node is connected to each output node number of parameter is $m\times n$	3fd5a241e2b84a53a2339c60fbcd957a
Hidden Layer	Neural Network Layer Notion	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023			a22cafa21026469094d260b473fd4789
Input Layer	Neural Network Layer Notion	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023			abdd1cda0d124843a34c66dc5ec7eba1
Layer Normalization	Neural Network Layer Notion	Nov 19, 2019	Alan Jo	Alan Jo	Jul 6, 2023	[Batch Normalization](https://texonom.com/batch-normalization-0e5e919e98b74b168a248a9fd3e0cb63)	BN normalizes the activations of each batch, while LN normalizes the activations of each layer	c5111411c0594c04847b0eb4d8fa6682
Output Layer	Neural Network Layer Notion	May 11, 2023	Alan Jo	Alan Jo	May 11, 2023			184eb47c3a98437daceba11104377194
Residual Connection	Neural Network Layer Notion	Mar 7, 2023	Alan Jo	Alan Jo	May 11, 2023		### Skip connection 하위 층에서 학습된 정보가 데이터 처리 과정에서 손실되는 것을 방지하기 위한 방법 > [[DL] Exploding & Vanishing Gradient 문제와 Residual Connection](https://heeya-stupidbutstudying.tistory.com/entry/DL-Exploding-Vanishing-gradient-문제와-Residual-Connection잔차연결)	f6778b3cfbee4b0cab02f7fad533372f
Internal Covariate Shift	Batch Normalization Methods	Jul 5, 2023	Alan Jo	Alan Jo	Jul 6, 2023		Internal distribution could have bias or variance even if input data distribution is normalized > [[PyTorch로 시작하는 딥러닝 기초] 09-4 Batch-Normalization](https://wegonnamakeit.tistory.com/47)	827b94e336a141b4a2ca100fc42dab66
Whitening	Batch Normalization Methods	Mar 7, 2023	Alan Jo	Alan Jo	Jul 6, 2023		들어오는 입력값의 특징들을 uncorrelated 하게 만들어주고, 각각의 분산을 1로 만들어주는 작업 covariance matrix의 계산과 inverse의 계산이 필요하기 때문에 계산량이 많을 뿐더러, Whitening은 일부 파라미터들의 영향이 무시된다 특정 파라미터가 계속 커지는 상태로 유지될수 있다는 문제도 > [[Deep Learning] Batch Normalization (배치 정규화)](https://eehoeskrap.tistory.com/430)	5bfaed3572904c208125cc9df6c664b0
ANN	Neural Networks	Mar 5, 2023	Alan Jo	Alan Jo	Mar 6, 2023		### Artificial Neural Network 기계학습과 인지과학에서 생물학의 신경망에서 영감을 얻은 통계학적 학습 알고리즘	d4232205ecf9463c95a911d179c87a84
CNN	Neural Networks	Nov 5, 2019	Alan Jo	Seong-lae Cho	Jul 4, 2023	[Capsule Network](https://texonom.com/capsule-network-a14fc7e569864154aae1ef44106e8991) [BCI](https://texonom.com/bci-e7b6d2a55d19435b9d3f8990441cab9e)	## Convolutional Neural Network Actually we use [Correlation](https://texonom.com/correlation-d75ea902ec294bf19e1180b6bfd22137) not the [Convolution](https://texonom.com/convolution-54def9d34de64e199ea46393675af2ce) Convolutional Layer와 Pooling Layer를 번갈아 가며 적용하여 입력 이미지의 특징을 추출하고 (Fully Connected) Layer를 통해 분류 작업을 수행 ### Hyperparameters - Convolution layer’s filter type, size, stride which result each activation map - Placements and types of convolution layer, pooling layer and activation layer So the typical architecture look like this where N is usually up to ~5, M is large, 0 <= K <= 2. but recent advances challenge this paradigm ```type[(CONV-RELU)N-POOL?]M-(FC-RELU)*K,SOFTMAX``` ### CNN Notion \|Title\| \|:-:\| \|[CNN History](https://texonom.com/cnn-history-76ffc514584846348136ea794f37ea9a)\| \|[Pooling Layer](https://texonom.com/pooling-layer-c36550932e8c4436b4b6da0d4d2fc946)\| \|[Convolutional Layer](https://texonom.com/convolutional-layer-0880b6c2089b4b20bb9a6232b6cb4b19)\| \|[CNN bottleneck layer](https://texonom.com/cnn-bottleneck-layer-c67427171be5400db198dd6e7ef644ce)\| \|[Training CNN](https://texonom.com/training-cnn-dc4f33cdf45544668d6f9cb6d059f431)\| \|[Local Contrast Normalization](https://texonom.com/local-contrast-normalization-036406a996c94edab249e7ff8a32b2bc)\| \|[ctivation map](https://texonom.com/ctivation-map-d0117ee2531b44dcbb0d19d33dff23ac)\| ### CNN Models \|Title\| \|:-:\| \|[ResNet](https://texonom.com/resnet-d8c7707445684e848ebef939500672d4)\| \|[ALexNet](https://texonom.com/alexnet-e33d746e06a540d4b8900be7ccede18c)\| \|[ConvNets](https://texonom.com/convnets-1846cf4fac8b4feea6fa8cf4ac083df9)\| \|[SuperVision](https://texonom.com/supervision-9d75ddb2bccd4dc9857f4332ae096298)\| \|[GoogLeNet](https://texonom.com/googlenet-64b5ad51aafc48d492ba489f47ffa58a)\| \|[VGG](https://texonom.com/vgg-c59b523e5e894bf2842474c9022dceec)\| \|[MSRA](https://texonom.com/msra-9024cb9533e54bdc9ada6d30601f847b)\|	002bf81a77bc40d1858740d26b61d97b
FFNN	Neural Networks	Mar 5, 2023	Alan Jo	Alan Jo	Mar 7, 2023	[Tree Data](https://texonom.com/tree-data-45f2e08b1f184e60b2e47c96e4cb535f) [RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16)	### *Feed-Forward Neural Network* 고안된 인공 신경망의 최초의 가장 단순한 형태 ### FFNN Usages \|Title\| \|:-:\| \|[Position-wise FFNN](https://texonom.com/position-wise-ffnn-6fe30c96aa9245d898c73ec34625377d)\|	89ecee87d8b7482e86995950db90eb31
GNN	Neural Networks	Apr 30, 2023	Alan Jo	Alan Jo	Apr 30, 2023		### Graph neural network 그래프로 나타낼 수 있는 데이터를 처리하기 위한 인공 신경망 기존의 특정 신경망 아키텍처는 적절하게 정의된 그래프에서 작동하는 GNN으로 해석가능 ### GNN Notion \|Title\| \|:-:\| > [Learnable Structural Semantic Readout for Graph Classification](https://arxiv.org/abs/2111.11523)	58adb81cb2b649d9af19019182960bb2
Perceptron	Neural Networks	Nov 5, 2019	Alan Jo	Seong-lae Cho	Jun 7, 2023	[Discriminant Function](https://texonom.com/discriminant-function-ed98e4c8d7284c34824f332770ccaf52) [Neuron](https://texonom.com/neuron-35ebe3514dc64647a74a0c3bad47b51b)	Can mimic OR function but cannot separate XOR by manipulating weights (flexibility) 입력값에 대한 이진 분류를 수행하거나 선형 예측기로 사용 based on only mistakes gradient descent update weight vector ### Perceptron Notion \|Title\| \|:-:\| \|[Perceptron Criterion](https://texonom.com/perceptron-criterion-5524bf3059d244f2b2437e74120d1b2b)\| \|[Optimization for Perceptron Criterion](https://texonom.com/optimization-for-perceptron-criterion-e64193867ff34deebc0b319d13be3545)\| \|[Perceptron convergence theorem](https://texonom.com/perceptron-convergence-theorem-7837263544974463b5bcc6a812fc6dce)\| \|[Universal Approximation Theorem](https://texonom.com/universal-approximation-theorem-872ac98a16c94f1f8a29f49496a1dfcd)\| \|[Multi Layer Perceptron](https://texonom.com/multi-layer-perceptron-9de4da2c68f24d76824e792b0c949ab8)\| \|[Perceptron History](https://texonom.com/perceptron-history-24992383cc2c44718e86d1cb53133e41)\|	1deb66f486d54c93bb928d8afba1864c
RNN	Neural Networks	Mar 12, 2021	Alan Jo	Alan Jo	Jul 30, 2023	[Sequential Data](https://texonom.com/sequential-data-784908b963384da3b3b603ebf1068e4d) [Tanh Function](https://texonom.com/tanh-function-679abf35c4eb4fb0bb0e56140cac3d23) [Vanishing Gradient](https://texonom.com/vanishing-gradient-6dc21442840841b3879d6502c5efd98a)	## Recurrent Neural Networ sequential 하거나 time series data 에 적합한 모델 RNN기반 모델인 LSTM과 GRU는 구조적으로 고정 길이의 hidden state 벡터에 모든 단어의 의미를 담아야 하기 때문에 문장이 길어지면 모든 단어 정보를 고정 길이 벡터에 담기 어렵다는 문제 Input Seq’s information lost so 보정하기 위해 [Attention Mechanism](https://texonom.com/attention-mechanism-762711860abb45f59904f1ac4e4af285) 사용 어텐션을 RNN의 보정을 위한 용도로서 사용하는 것이 아니라 인코더 디코더에도 사용하게 ### RNN Notion \|Title\| \|:-:\| \|[Recurrent Neurons](https://texonom.com/recurrent-neurons-341ce79b6a0648bb8c57fafc806573cc)\| \|[Long-term dependency](https://texonom.com/long-term-dependency-0b4e9db8d0cd4fc4889b83e32408e5f8)\| \|[RWKV](https://texonom.com/rwkv-639fb6fcbaba43e398a6e2fe5495373a)\| \|[Hidden State](https://texonom.com/hidden-state-7b65c9b953ce4a9b843d2a0b381a5453)\| ### RNNs \|Title\| \|:-:\| \|[LSTM](https://texonom.com/lstm-7946d20b7c394827b02e7b6b297bbadf)\| \|[GRU](https://texonom.com/gru-280a4b083b0b49989e3cc36ef493410c)\| \|[Bidirectional RNN](https://texonom.com/bidirectional-rnn-4f44262a3cc3417e8b31f24040cef871)\| \|[Vanilla RNN](https://texonom.com/vanilla-rnn-f4032d5ead424efaaf6e1fb210b546cf)\| > [07-1.순환 신경망(RNN, Recurrent Neural Network) - (1)](https://excelsior-cjh.tistory.com/183)	f7aad56acb5542b2ac26c2908be4ce16
SNN	Neural Networks	May 30, 2022	Alan Jo	Alan Jo	Mar 6, 2023		### Spiking neural network Closely mimic natural neural networks 자연 신경망을 더 가깝게 모방한 인공 신경망. 뉴런 및 시냅스 상태 외에도 SNN은 작동 모델에 시간 개념을 통합 ### SNN Notion \|Title\| \|:-:\| > [Untitled](http://www.koreascience.kr/article/JAKO202025465017052.pdf)	6cf69239a87e4df9a32b3494862374e4
ALexNet	CNN Models	Sep 14, 2020	Alan Jo	Alan Jo	May 31, 2023	[ImageNet](https://texonom.com/imagenet-d5542cdb730f4e218bacb583b5e2576b)	### GPU CNN 구조를 기반으로 한 ALexNet, ILSVRC 오차율이 10%로 줄어든 혁신적인 알고리즘 > [CS231n 1강 요약](https://taeyoung96.github.io/cs231n/CS231n_1/)	e33d746e06a540d4b8900be7ccede18c
ConvNets	CNN Models	May 23, 2023	Alan Jo	Alan Jo	May 23, 2023		CNN ⇒ [Local Contrast Normalization](https://texonom.com/local-contrast-normalization-6357cfe8ce294c69ae102053dcb5aaff) ⇒ [Pooling Layer](https://texonom.com/pooling-layer-c36550932e8c4436b4b6da0d4d2fc946)	1846cf4fac8b4feea6fa8cf4ac083df9
GoogLeNet	CNN Models	May 23, 2023	Alan Jo	Alan Jo	Jun 7, 2023		2014	64b5ad51aafc48d492ba489f47ffa58a
MSRA	CNN Models	May 23, 2023	Alan Jo	Alan Jo	Jun 7, 2023		2015	9024cb9533e54bdc9ada6d30601f847b
ResNet	CNN Models	May 30, 2023	Alan Jo	Alan Jo	Jun 6, 2023	[CNN bottleneck layer](https://texonom.com/cnn-bottleneck-layer-c67427171be5400db198dd6e7ef644ce)	### Microsoft 2015 [COCO](https://texonom.com/coco-c7524d6e059444c09a5067e07caf94e6) > [ResNet - Azure Machine Learning](https://learn.microsoft.com/en-us/azure/machine-learning/component-reference/resnet?view=azureml-api-2) > [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) > [Pytorch로 ResNet 구현, torch summary 살펴보기](https://velog.io/@gibonki77/ResNetwithPyTorch)	d8c7707445684e848ebef939500672d4
SuperVision	CNN Models	May 23, 2023	Alan Jo	Alan Jo	Jun 7, 2023		2012	9d75ddb2bccd4dc9857f4332ae096298
VGG	CNN Models	May 23, 2023	Alan Jo	Alan Jo	Jun 7, 2023		2014	c59b523e5e894bf2842474c9022dceec
CNN bottleneck layer	CNN Notion	May 30, 2023	Alan Jo	Alan Jo	Jun 7, 2023		## inception module 핵심은 1x1 Convolution 연산량을 최소화 연산량과 정보손실은 서로 tradeoff 관계라 연산량과 정보 손실의 관계를 잘 보면서 Error가 최소가 되는 지점 ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F59552677-4c80-46c9-b3fc-9678673bf88f%2FUntitled.png?table=block&id=4344af8f-0f0c-4470-bc0e-482e0f837efe&cache=v2) [ResNet](https://texonom.com/resnet-d8c7707445684e848ebef939500672d4) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Facc5b086-6258-4b9c-9aba-d9e0bcd5a220%2FUntitled.png?table=block&id=559e2275-d78b-4266-91ce-47404eab13c1&cache=v2) ### [GoogLeNet](https://texonom.com/googlenet-64b5ad51aafc48d492ba489f47ffa58a) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F75fa731a-374c-4ea7-b572-2086555c79ca%2FUntitled.png?table=block&id=8c337e84-43f0-47cf-acf7-9b683cebfc88&cache=v2) > [CNN의 Bottleneck에 대한 이해](https://velog.io/@lighthouse97/CNN의-Bottleneck에-대한-이해)	c67427171be5400db198dd6e7ef644ce
CNN History	CNN Notion	Sep 14, 2020	Alan Jo	Alan Jo	May 30, 2023		1998년에 이미 CNN 알고리즘이 제안되었다.하지만 여러 문제들에 부딫혀 그 동안 발전하지 못했다. > 데이터의 양이 너무 적었다.연산 속도가 굉장히 느렸다. ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F12554e84-bfcd-4125-8bed-f31d172c4bf4%2FUntitled.png?table=block&id=4d668c12-9d63-4e37-a8b5-0b23fdeee53c&cache=v2) ### 1980 Fukushima “sandwich” architecture (SCSCSC…) > [Untitled](https://www.rctn.org/bruno/public/papers/Fukushima1980.pdf) ### 1998 [Yann LeCun](https://texonom.com/yann-lecun-7866102480474d939d45d97ddd318c3a), Bottou, Bengio, Haffner Gradient-based learning applied to document recognition > [Untitled](https://ieeexplore.ieee.org/document/726791) ### 2012 [ALexNet](https://texonom.com/alexnet-e33d746e06a540d4b8900be7ccede18c) [Geoffrey Hinton](https://texonom.com/geoffrey-hinton-441d5ce2b78146d0935454042b4f06d9), [Ilya Sutskever](https://texonom.com/ilya-sutskever-f38707abee824d21b2ec9824a1706342) ImageNet Classification with Deep Convolutional Neural Networks > [Untitled](https://proceedings.neurips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf)	76ffc514584846348136ea794f37ea9a
Convolutional Layer	CNN Notion	May 18, 2023	Alan Jo	Alan Jo	Jun 7, 2023		### Share the same parameters across different locations (assume input stationary) In practice usually zero pad the border parameter 가 fully connected 나 locally connected 보다 훨씬 작아서 There are weight variables which are same as the size of filter including all that channels Using small filter is better [CNN Stride](https://texonom.com/cnn-stride-b84e6cf4abd347289dc8dbe6ffc1097d) 1x1 is also valid because we can reduce or increase the Output Channel of feature map with different scaling 채널별 필터 총합이 최종필터 Commonly number of filters K is powers of 2 K가 늘면 지수적으로 연관 픽셀이 늘어나는게 아니라 선형적으로 가까운 거리로 늘어남 ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F2b27a771-bf6c-4c73-bb6c-6dc888a9d56e%2FUntitled.png?table=block&id=04353e66-16fb-48a3-a472-44bf24a7fe77&cache=v2) ### Multiple small sized filter layer’s parameters have much smaller than one large sized filter layer (also more nonlinearity, less compute) so we use mostly 3x3 after [VGG](https://texonom.com/vgg-c59b523e5e894bf2842474c9022dceec) ([ALexNet](https://texonom.com/alexnet-e33d746e06a540d4b8900be7ccede18c) used 11x11 or 7x7 convolution layer) 1x1 으로 더 줄일 수 있는데 그게 [CNN bottleneck layer](https://texonom.com/cnn-bottleneck-layer-c67427171be5400db198dd6e7ef644ce) ### total parameter $$filter \; size ^2 +1\;for\;bias$$	0880b6c2089b4b20bb9a6232b6cb4b19
ctivation map	CNN Notion	Jun 7, 2023	Alan Jo	Alan Jo	Jun 18, 2023		입력 이미지의 각 위치에서 활성화된 뉴런의 출력값을 시각화한 것 모델이 어떤 부분에서 입력 이미지를 보고 판단하는지 이해하는 데 도움을 줌	d0117ee2531b44dcbb0d19d33dff23ac
Local Contrast Normalization	CNN Notion	Jun 7, 2023	Alan Jo	Alan Jo	Jun 7, 2023		after convolution layer, before pooling layer ConvNet	036406a996c94edab249e7ff8a32b2bc
Pooling Layer	CNN Notion	Apr 30, 2023	Alan Jo	Alan Jo	Jun 7, 2023		### [Image Downsampling](https://texonom.com/image-downsampling-97df12a6eb2949aa808c6f03dac101a6) feature map의 크기를 줄이는 데 사용 ### CNN Pooling Types \|Title\| \|:-:\| \|[CNN Spatial Pooling](https://texonom.com/cnn-spatial-pooling-d191e60632c84b54b0266c2a73f54c51)\| \|[CNN Static Pooling](https://texonom.com/cnn-static-pooling-48534a9ed35e465c94694903dd2c80cb)\| \|[CNN Dynamic Pooling](https://texonom.com/cnn-dynamic-pooling-3412ef080ffb4e949bebfed80a17cf8d)\| \|[CNN Global Pooling](https://texonom.com/cnn-global-pooling-fe9913d4c6674a808ef6c899e778dcd4)\|	c36550932e8c4436b4b6da0d4d2fc946
Training CNN	CNN Notion	Jun 1, 2023	Alan Jo	Alan Jo	Jun 1, 2023		![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F0b17e758-45f3-40f9-986e-592b434eb581%2FUntitled.png?table=block&id=07c57c5d-4ab9-43d3-aab3-dc7fb384178c&cache=v2)	dc4f33cdf45544668d6f9cb6d059f431
CNN Stride	Convolutional Layer	null	null	null	null	null	How many pixels to skip $$Count = (N + 2P - F) / stride + 1$$ P is padding size, N is input size, F is filter size and count is output size+	b84e6cf4abd347289dc8dbe6ffc1097d
CNN Dynamic Pooling	CNN Pooling Types	Apr 30, 2023	Alan Jo	Alan Jo	Apr 30, 2023		segment-level에서 feature를 집계 입력 데이터 내에서 고수준 특징들의 시간 정보를 더 잘 보존 > [Learnable Dynamic Temporal Pooling for Time Series Classification](https://arxiv.org/abs/2104.02577)	3412ef080ffb4e949bebfed80a17cf8d
CNN Global Pooling	CNN Pooling Types	Apr 30, 2023	Alan Jo	Alan Jo	Apr 30, 2023		고수준 특징들의 버려질 수 있다 feature map 내의 모든 값을 집계하여 하나의 숫자로	fe9913d4c6674a808ef6c899e778dcd4
CNN Spatial Pooling	CNN Pooling Types	May 23, 2023	Alan Jo	Alan Jo	Jun 7, 2023		Sum or max - max pooling - average pooling - min pooling - L2 pooling - L2 pooling over features	d191e60632c84b54b0266c2a73f54c51
CNN Static Pooling	CNN Pooling Types	Apr 30, 2023	Alan Jo	Alan Jo	Apr 30, 2023		고정된 크기의 영역을 선택하여 그 영역 내의 값을 집계	48534a9ed35e465c94694903dd2c80cb
**Position-wise FFNN**	FFNN Usages	Mar 5, 2023	Alan Jo	Alan Jo	Mar 7, 2023		## Fully-connected FFNN [Seq2Seq](https://texonom.com/seq2seq-01a9854dffa6417c87d92c11a607250c) 인코더와 디코더에서 공통적으로 가지고 있는 서브층 > [16-01 트랜스포머(Transformer)](https://wikidocs.net/31379)	6fe30c96aa9245d898c73ec34625377d
Multi Layer Perceptron	Perceptron Notion	Mar 5, 2023	Alan Jo	Alan Jo	May 11, 2023	[Neural Network](https://texonom.com/neural-network-86f54f9f1de848c1a29c56c24f7d5094)	## MLP Can have non-linear decision boundary using 3 Perceptron, now available to separate XOR ### Two layer convex open or closed region ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Ffb96ba60-f3ee-44af-9457-8d166f25df99%2FUntitled.png?table=block&id=83c05706-0b14-4c08-a841-60ce8ac8a182&cache=v2) each line means perceptron ### Three layer arbitrary (complexity limited by number of neurons) ### Hidden layer except final layer	9de4da2c68f24d76824e792b0c949ab8
Optimization for Perceptron Criterion	Perceptron Notion	Apr 27, 2023	Alan Jo	Alan Jo	Apr 27, 2023		### Maximize set of misclassified samples under w only update until convergence ### Weight Updates Start with weight 0 For each training instance > - If correct (i.e., y=y), no change > - If wrong: adjust the weight vector - adding or subtracting the feature vector (f / y exactly > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F96bde13f-ce87-46fb-942e-c9964ec7123f%2FUntitled.png?table=block&id=a35c861a-d054-4419-b50c-c672b7a6ae41&cache=v2) > > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F3e523c5e-9337-4c60-8453-1396f889d8a9%2FUntitled.png?table=block&id=fdbb578d-a453-4351-a61b-5efbfd34f979&cache=v2) > > x0 bias ican be tan add to f ### Error-Driven Linear Classification > Binary case - compare features to weight vector > Learning - figure out weight vector from examples	e64193867ff34deebc0b319d13be3545
Perceptron convergence theorem	Perceptron Notion	Apr 27, 2023	Alan Jo	Alan Jo	Apr 27, 2023		Convergence is guaranteed for linearly separable data Cannot classify non-linearly separable data	7837263544974463b5bcc6a812fc6dce
Perceptron Criterion	Perceptron Notion	Apr 27, 2023	Alan Jo	Alan Jo	Apr 27, 2023		### Inequality ### Linear Classifiers - Inputs data are feature values (feature vector) - (x0, x1, x2, ... , xn) - Each feature has a weight - Sum is the activation > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F97f9d462-0b77-464a-8d69-45b5d1a8d92b%2FUntitled.png?table=block&id=18b03cdd-2d54-42ea-8372-c57a7ff78e2a&cache=v2) - output - if activation is positive → 1 negative → -1 (classification +/-) ### Decision Rule binary decision rule One side corresponds to Y=+1 Other corresponds to Y=-1 ### Learning - binary perceptron 1. Start with weights = 0 > 1. For each training instance >> - Classify with current weights >> - If correct (i.e., y=y), no change >> - If wrong: adjust the weight vector by adding or subtracting the feature vector. Subtract if y is -1. > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fcc561d3b-0dae-4866-bf99-e3d2593dba1f%2FUntitled.png?table=block&id=81c5098b-4384-4058-9615-2678d672d6e7&cache=v2) ## Multiclass dicision rule ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F9af73bdd-20cd-4421-a317-ddd37127ae63%2FUntitled.png?table=block&id=49e83b0c-2442-4df3-acbb-88eefa82cb3f&cache=v2) lower count of wrong class (for this instance), raise count of right class (for this instance) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fc30c5970-b6a9-4c98-a062-562a93d94adb%2FUntitled.png?table=block&id=d15b9c2d-8e80-4688-9a70-675be4443d7f&cache=v2) black box (kernel) K that told us the dot product of two examples x and x’ ### Solution 0. set true value to +/- 2. get data 3. choose true value 4. classification by hypothesis → compare to true value 5. update hypothesis to all data try this until no revise to all data redo these to data set ### multiclass 다른거 똑같은데 분류라서 argmax xf = y 틀리면 그놈은 값 빼주고 맞는놈은 올려주기 f만큼 듀얼이면 argmax alpha K(k is sigma ff) - at sigma only one f change by w if wrong 틀린놈은 알파에서 1 빼고 맞는놈들은 1만큼 올려줌 맞는 클래스 틀린 클래스 ### Feature Vectors ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fd1bbd38a-1fa3-43b3-96de-b4161118d84f%2FUntitled.png?table=block&id=0d54acc4-2dc6-4fdf-ba33-4f798cdf9949&cache=v2) Some (Simplified) Biology - Inputs are feature values - Each feature has a weight - Sum is the activation ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fd20f5cf0-dbee-4d69-b541-77430ef0677e%2FUntitled.png?table=block&id=ad108811-66aa-45fa-a82f-89af6936407a&cache=v2) If the activation is: Positive, output +1 Negative, output -1 ### Weights - Binary case: compare features to a weight vector - Learning: figure out the weight vector from examples ### Decision Rules Binary Decision Rule in the space of feature vectors > - Examples are points > - Any weight vector is a hyperplane > - One side corresponds to Y=+1 > - Other corresponds to Y=-1	5524bf3059d244f2b2437e74120d1b2b
Perceptron History	Perceptron Notion	Jun 7, 2023	Alan Jo	Alan Jo	Jun 7, 2023		### 1957 released 1957 The Perceptron (Frank Rosenblatt) Mark I Perceptron The machine was connected to a camera that used 20×20 cadmium sulfide photocells to produce a 400-pixel image. Widrow and Hoff, ~1960: Adaline/Madaline 1969 Perceptrons (Minsky, Papert) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fef3314d0-ca74-4a67-bb3c-eae0eed6371d%2FUntitled.png?table=block&id=a63f0670-5848-468e-8eb6-2caf9025092c&cache=v2) > [Perceptron \| Encyclopedia of Computer Science](https://dl.acm.org/doi/abs/10.5555/1074100.1074686)	24992383cc2c44718e86d1cb53133e41
Universal Approximation Theorem	Perceptron Notion	Apr 27, 2023	Alan Jo	Alan Jo	Jun 13, 2023		### at least 1 hidden layer Any continuous function $f: [0, 1] \rightarrow [0, 1]$ can be approximated arbitrarily well by aneural network with at least 1 hidden layer with a finite number of weights One wide (latent) layer is enough, but it is just a memorizer (Cybenko ’89) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fd15809a9-4c4b-4e05-bc71-c1005175924b%2FUntitled.png?table=block&id=86addc25-55d1-4a93-984f-c728d5375d25&cache=v2)	872ac98a16c94f1f8a29f49496a1dfcd
Hidden State	RNN Notion	Mar 7, 2023	Alan Jo	Alan Jo	Jun 21, 2023	[Latent Space](https://texonom.com/latent-space-d67b6bdef18b4058bfbc3d25f87ec087)	### latent state, internal state The hidden state is computed based on the current input and the previous hidden state 이 프로세스는 순차적이었고 병렬화를 방해 > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/)	7b65c9b953ce4a9b843d2a0b381a5453
Long-term dependency	RNN Notion	Mar 6, 2023	Alan Jo	Alan Jo	Mar 6, 2023	[GRU](https://texonom.com/gru-280a4b083b0b49989e3cc36ef493410c) [LSTM](https://texonom.com/lstm-7946d20b7c394827b02e7b6b297bbadf)	RNN Problem 문장이 길어질 경우 앞 단어의 정보를 잃어버리게 > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리)	0b4e9db8d0cd4fc4889b83e32408e5f8
Recurrent Neurons	RNN Notion	Mar 12, 2021	Alan Jo	Alan Jo	Oct 6, 2021		순환 뉴런	341ce79b6a0648bb8c57fafc806573cc
RWKV	RNN Notion	May 28, 2023	Alan Jo	Alan Jo	May 28, 2023		### parallelizable training ### RWKV Implementation \|Title\| \|:-:\| \|[](https://texonom.com/4aea4d107bb245f280ce2ab0023546ce)\| \|[](https://texonom.com/8b87d4786fc6461b97ccc1250d23a9aa)\| > [RWKV: Reinventing RNNs for the Transformer Era](https://arxiv.org/abs/2305.13048)	639fb6fcbaba43e398a6e2fe5495373a
[RWKV-LM](https://github.com/BlinkDL/RWKV-LM)	RWKV Implementation	May 28, 2023	Alan Jo	Alan Jo	May 28, 2023			4aea4d107bb245f280ce2ab0023546ce
[rwkv.cpp](https://github.com/saharNooby/rwkv.cpp)	RWKV Implementation	May 28, 2023	Alan Jo	Alan Jo	May 28, 2023	[GGML](https://texonom.com/ggml-d3297b7eedc846a1bea05cfa0f3cc24a)		8b87d4786fc6461b97ccc1250d23a9aa
Bidirectional RNN	RNNs	Jun 21, 2023	Alan Jo	Alan Jo	Jun 21, 2023		- forwatd state - backward state	4f44262a3cc3417e8b31f24040cef871
GRU	RNNs	Oct 6, 2021	Alan Jo	Alan Jo	Jul 6, 2023		## Gated recurrent unit 2014년 뉴욕대학교 조경현 교수님이 집필한 논문에서 제안 LSTM와 성능과 구조가 비슷하지만 GRU가 학습할 가중치가 적다 출력 게이트 없음 ### GRU Notion \|Title\| \|:-:\| \|[GRU Update Gate](https://texonom.com/gru-update-gate-42daff55639e4781a5fee0170a13b2a2)\| \|[GRU Reset Gate](https://texonom.com/gru-reset-gate-89417d3e3bd44a5f89d67e199b36f381)\| > [Gated Recurrent Units (GRU)](https://yjjo.tistory.com/18)	280a4b083b0b49989e3cc36ef493410c
LSTM	RNNs	Aug 16, 2021	Alan Jo	Alan Jo	Jul 6, 2023	[Quant Trading](https://texonom.com/quant-trading-024f2adb817a4e33b9fc6905bba8eb0c)	## Long Short Term Memory RNN의 단점인 기울기 소실 문제를 해결하기 위해 gate를 추가한 모델 시계열 시퀀스가 있는 데이터에서 다음을 예측 이름에서 알 수 있는 단기 기억을 가지는 것처럼 행동하는 모델 특히 주식 예측에 많이 사용 ### LSTM Notion \|Title\| \|:-:\| \|[LSTM Cell State](https://texonom.com/lstm-cell-state-128d65c00df5470a98d1603ae1e56bd0)\| \|[LSTM forget gate](https://texonom.com/lstm-forgetgate-ee99b6c3bb5b432295e2f54b1b58a9a1)\| \|[LSTM input gate](https://texonom.com/lstm-inputgate-02d2744fe7344bff9502e5bb111ff759)\| - forget gate - input gate - output gate ### LSTMs \|Title\| \|:-:\| \|[BiLSTM](https://texonom.com/bilstm-cc8242f116e643abb71167e859561991)\| ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F4b058d6d-0010-47fe-95d8-9604f829655b%2FUntitled.png?table=block&id=c557c8ba-af7b-4ac2-929d-3e62c6599b79&cache=v2) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F86665118-813b-4d35-a2d2-2f1923910e5c%2FUntitled.png?table=block&id=382fda06-da43-46d6-bb50-3da9736036d3&cache=v2) > [Long Short-Term Memory (LSTM) 이해하기](https://dgkim5360.tistory.com/entry/understanding-long-short-term-memory-lstm-kr) > [08-02 장단기 메모리(Long Short-Term Memory, LSTM)](https://wikidocs.net/22888) > [삼성전자 주식, 이더리움 시세 인공지능으로 예측해보자 - Python, Deep Learning](https://www.youtube.com/watch?v=sG_WeGbZ9A4&t=241s)	7946d20b7c394827b02e7b6b297bbadf
Vanilla RNN	RNNs	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023		짧은 시퀀스에만 효과 the problem of Long-Term Dependencies ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fa6b722a0-f76c-4dcf-a4dd-374531416413%2FUntitled.png?table=block&id=1c9f7c91-3cb8-4d9a-8fb7-04fa17238677&cache=v2) > [08-02 장단기 메모리(Long Short-Term Memory, LSTM)](https://wikidocs.net/22888)	f4032d5ead424efaaf6e1fb210b546cf
GRU Reset Gate	GRU Notion	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023			89417d3e3bd44a5f89d67e199b36f381
GRU Update Gate	GRU Notion	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023			42daff55639e4781a5fee0170a13b2a2
LSTM Cell State	LSTM Notion	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023		cell state에 뭔가를 더하거나 없앨 수 있는 능력 gate로 제어함	128d65c00df5470a98d1603ae1e56bd0
LSTM **forget gate**	LSTM Notion	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023			ee99b6c3bb5b432295e2f54b1b58a9a1
LSTM **input gate**	LSTM Notion	Jul 6, 2023	Alan Jo	Alan Jo	Jul 6, 2023			02d2744fe7344bff9502e5bb111ff759
BiLSTM	LSTMs	Aug 16, 2021	Alan Jo	Alan Jo	Jul 6, 2023		### Bidirectional LSTM	cc8242f116e643abb71167e859561991
AI Data	AI Usages	May 13, 2021	Alan Jo	Alan Jo	Jun 1, 2023	[Model Generalization](https://texonom.com/model-generalization-0a2de5c4578c47dd9a5dc1b83fbaf72b) [Observation Noise](https://texonom.com/observation-noise-dcf85cf878d740b0ad7dd8ecab260165) [Data Processing](https://texonom.com/data-processing-d1b576ef34304e0fa6325db709860e0b)	### Data enhancement is more effective than code enhancement Don't try to be a hero, data is 80 ### AI Data Usages \|Title\| \|:-:\| \|[AI Data Tool](https://texonom.com/ai-data-tool-3cb903994b5c4c2aad0e3fddb4511a3f)\| \|[Data Labeling](https://texonom.com/data-labeling-dda48db348de44a9bafa1ae69ca18133)\| \|[Data Augmentation](https://texonom.com/data-augmentation-fd2918b94bd54726a8c91e26af129545)\| \|[Data Testing](https://texonom.com/data-testing-bda06f8a3b0349ac8cb78d960f24bf46)\| \|[Experiment Management](https://texonom.com/experiment-management-cbea3983015346a6bd81c98d2d65252d)\| \|[Temporal Sequence](https://texonom.com/temporal-sequence-a2558b4ef95c4c37bee54056c1804c8d)\| \|[AI Dataset](https://texonom.com/ai-dataset-6383f9b5b0c54e598b90dc9ea0c8e5d1)\| > [Introduction to streaming for data scientists](https://huyenchip.com//2022/08/03/stream-processing-for-data-scientists.html)	dcd9611d8a5647ba95e7bc63a7e6db1f
AI Development	AI Usages	Jun 13, 2021	Alan Jo	Alan Jo	Jun 21, 2023	[Statistics Tool](https://texonom.com/statistics-tool-cd8ad9a0f6be46b4979d8e302c6dbd52)	> [Emerging Architectures for LLM Applications \| Andreessen Horowitz](https://a16z.com/2023/06/20/emerging-architectures-for-llm-applications) > [Stack Overflow](https://stackoverflow.co/labs/developer-sentiment-ai-ml) > [Futurepedia - The Largest AI Tools Directory \| Home](https://www.futurepedia.io/)	94a4b3b469b5439aaa03a669f894e3c8
AI Industry	AI Usages	Aug 21, 2021	Alan Jo	Alan Jo	Jul 4, 2023	[Industry](https://texonom.com/industry-989c1bacc8b84fb1b63486f61e3b9052) [AI Object](https://texonom.com/ai-object-e69c83e69fd34103a83c1341b65158be) [Sex Industry](https://texonom.com/sex-industry-e5080bf012e0478f9e6fc996bdc8a2cb) [Copyright](https://texonom.com/copyright-02b119d69a0a475289bbcd13faae44f3)	AI progress is constrained by bottlenecks in essential economic sectors and steps of the innovation process, and that breakthroughs in manipulating the physical world are necessary 현재까지 AI 창작물은 저작권 없다는 걸 이용해야 ### AI Companies \|Title\| \|:-:\| \|[OpenAI](https://texonom.com/openai-501e879eb1ca4ef8b388ba5b11e92e23)\| \|[Deepmind](https://texonom.com/deepmind-5eb171c77b344d4786a9a5b23ae70eca)\| \|[Stability AI](https://texonom.com/stability-ai-7285625019ec4498a7300ac9f0736b2d)\| \|[Anthropic AI](https://texonom.com/anthropic-ai-096e36d827a2421b9b0ee8d1ff3b0305)\| \|[AllenAI](https://texonom.com/allenai-930c776c8fdd4e358032803f037cd6fa)\| \|[Saltlux](https://texonom.com/saltlux-de9945849d70489f99a33f81a1707a46)\| \|[Neural Magic](https://texonom.com/neural-magic-021ab1b8be0d4f95b8ae6278c08e1562)\| \|[VoyagerX](https://texonom.com/voyagerx-bdc1dc8597e44f1aac3ab3ec6256eb87)\| \|[Upstage.ai](https://texonom.com/upstageai-319072815ced47bf98335f3a9c87cec4)\| \|[CarperAI](https://texonom.com/carperai-d69725f3dcfb4a57a9cf0f4bea30bc2a)\| \|[Eleuther AI](https://texonom.com/eleuther-ai-9f89e25571fe4930825ffabb3483c02a)\| \|[Preligens](https://texonom.com/preligens-040e14875d3b463fa71674578756944e)\| \|[inworld](https://texonom.com/inworld-29d82633e51742dabca809cc86d6e989)\| \|[Scatter Lab](https://texonom.com/scatter-lab-9524953c003c4c7d8ec48d387c1a42cf)\| \|[VAIV](https://texonom.com/vaiv-15772f18eca5478290a6ca76068cf077)\| \|[X.AI](https://texonom.com/xai-ccd1c9778b1349c788889cf20c1c9a72)\| \|[Inflection AI](https://texonom.com/inflection-ai-533110523b9f4b8ab79f663a6ba7f6b0)\| \|[MosaicML](https://texonom.com/mosaicml-d531ff0e6a51448c84296ab20f675d67)\| \|[Tiny corp](https://texonom.com/tiny-corp-2b42b01b901d4a93b80698286c5e2b40)\| \|[Cohere](https://texonom.com/cohere-5fc93cb3678841a7b23df3c7d0399aca)\| \|[Mind AI](https://texonom.com/mind-ai-007aca806190459c8437ce7046f5fbcc)\| ### AI Products Diagram > [How to Build AI Products People Want — Reforge](https://www.reforge.com/blog/ai-products-arms-race) ### Industry Landscape > [Why transformative AI is really, really hard to achieve](https://zhengdongwang.com/2023/06/27/why-transformative-ai-is-really-really-hard-to-achieve.html) > [LF AI Landscape](https://landscape.lfai.foundation/) ### Scaling > [3 ways AI is scaling helpful technologies worldwide](https://blog.google/technology/ai/ways-ai-is-scaling-helpful/) > [AI Will Create More Developers, Not Less](https://interconnect.substack.com/p/ai-will-create-more-developers-not)	d8709bd0498145e3a66af6da3f963fa7
AI Object	AI Usages	Jun 13, 2021	Alan Jo	Alan Jo	May 21, 2023	[ai-notes](https://github.com/sw-yx/ai-notes) [applied-ml](https://github.com/eugeneyan/applied-ml)	### AI 특성상 정확성이 필요없는 분야부터 대체될 것이고 그렇게 진행되고 있다 ### AI Objects \|Title\| \|:-:\| \|[AI Planning](https://texonom.com/ai-planning-2173477bde2147d1a91975e30756806a)\| \|[Vision AI](https://texonom.com/vision-ai-ba1da978d6f6488998e7cdc59431b41a)\| \|[Sound AI](https://texonom.com/sound-ai-c3f85fde353c4e7581cca5e3e30324f6)\| \|[IA](https://texonom.com/ia-92af447425bc4fe2b10f3aef85c1871e)\| \|[NLP](https://texonom.com/nlp-e0ae7b40dd23463ea7bc92195d6ec7fd)\| \|[AI Coding](https://texonom.com/ai-coding-f659b252482d4dfb90485c4a0428f442)\| \|[Recommend System](https://texonom.com/recommend-system-e94b364527554e65bfd9d3dc63fb1af7)\| \|[AI Invent](https://texonom.com/ai-invent-2fd3b79c01bd43239990cd547a2f6e41)\| \|[AI Game](https://texonom.com/ai-game-7cb70519f66942b1b59d1d7352ce34e1)\| \|[AI Automation](https://texonom.com/ai-automation-afa72ebcb9e4426799162de10da077f8)\| \|[AI Robot](https://texonom.com/ai-robot-0796068b90b246d8abc6aeed851624cd)\| \|[Multimodal AI](https://texonom.com/multimodal-ai-381b5ebe1218438689de97941a3d9829)\| \|[AI Agent](https://texonom.com/ai-agent-5efeaf8cd77f495988083eb2084d7f01)\| \|[Extract Information](https://texonom.com/extract-information-6988fe9ed1a444e7be50249e5ccf11bc)\|	e69c83e69fd34103a83c1341b65158be
AI Protocol	AI Usages	Apr 11, 2023	Alan Jo	Alan Jo	Apr 11, 2023		### AI Protocols \|Title\| \|:-:\|	be3148d3c1b24b90a4f51020fde30bd6
AI Data Tool	AI Data Usages	Jan 18, 2023	Alan Jo	Alan Jo	Jan 18, 2023		### AI Data Tools \|Title\| \|:-:\| \|[Clearnlab](https://texonom.com/clearnlab-479203d2d2cf4ad2999da2d1fcad12ff)\| \|[Eva db](https://texonom.com/eva-db-17a183f2967d455fbe7631ea89172e49)\|	3cb903994b5c4c2aad0e3fddb4511a3f
AI Dataset	AI Data Usages	Mar 27, 2021	Alan Jo	Alan Jo	Jun 6, 2023	[Text Corpus](https://texonom.com/text-corpus-1befe731759849c48134768343a90703)	We typically say that a dataset is high-dimensional if the number of data points N issmaller than the dimensionality D > - not cheatable > - large degree of intra-class variability > > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F895def36-e573-4dd1-8322-2a360807feb7%2FUntitled.png?table=block&id=fcecf613-dddd-441d-bf75-31e18d3ec61a&cache=v2) ### Vision AI Datasets \|Title\| \|:-:\| \|[ImageNet](https://texonom.com/imagenet-d5542cdb730f4e218bacb583b5e2576b)\| \|[WebLI](https://texonom.com/webli-5b0455624497488ba55432160ad812c6)\| ### NLP Datasets \|Title\| \|:-:\| \|[CommonCrawl](https://texonom.com/commoncrawl-e34ff04c04e14260a1ef19e3baa035c8)\| \|[C4 Data](https://texonom.com/c4-data-5e19c38f6e034c92857f71cb820dae89)\| \|[Code Data](https://texonom.com/code-data-071b31dcf1ee4073ad458f1ac781cd5f)\| \|[RedPajama Data](https://texonom.com/redpajama-data-ab8489ee6448474899c4b8e49b6faaac)\| > [20 Open Datasets for Natural Language Processing](https://odsc.medium.com/20-open-datasets-for-natural-language-processing-538fbfaf8e38)	6383f9b5b0c54e598b90dc9ea0c8e5d1
Data Augmentation	AI Data Usages	Jun 22, 2021	Alan Jo	Alan Jo	Jun 7, 2023		### Flip, Scale, Crop, Rotation Jitter, Shear, Stretch, Lens Distortion ### Data Augmentation Tools \|Title\| \|:-:\| \|[AI Data Generator](https://texonom.com/ai-data-generator-dd7e73ff30db4c0f89580aff9fb7df56)\| \|[AugLy](https://texonom.com/augly-0cbb390ff79642d1846f50f79ea9ed91)\| ### Data Augmentation Notion \|Title\| \|:-:\| \|[color jitter](https://texonom.com/color-jitter-08229f0a960c43acb65cedad8e7fb14a)\|	fd2918b94bd54726a8c91e26af129545
Data Labeling	AI Data Usages	May 13, 2021	Alan Jo	Alan Jo	May 11, 2023		### Data Labelings \|Title\| \|:-:\| \|[Image Labeling](https://texonom.com/image-labeling-dca8afabd8a14dc495545f414883d7b4)\| ### Cheap Africa human power > [Africa: The Hidden Workforce Behind AI](https://www.mantralabsglobal.com/blog/ai-in-africa-artificial-intelligence-africa/)	dda48db348de44a9bafa1ae69ca18133
Data Testing	AI Data Usages	Mar 21, 2023	Alan Jo	Alan Jo	May 11, 2023		### Data Testing Tools \|Title\| \|:-:\| \|[Pandera](https://texonom.com/pandera-632fbdee7cc847359bcc61bc9c67bf01)\|	bda06f8a3b0349ac8cb78d960f24bf46
Experiment Management	AI Data Usages	Jun 13, 2021	Alan Jo	Alan Jo	May 11, 2023		### Experiment Management Tools \|Title\| \|:-:\| \|[Neptune.ai](https://texonom.com/neptuneai-c61383aff215447a979fd6d446be277d)\| \|[Weights & Biases](https://texonom.com/weights-biases-51d1cbfc87324d1f8dd43a8978ea5f68)\| \|[TensorBoard](https://texonom.com/tensorboard-6af2721a16bd4dcdb172affcdf055d91)\|	cbea3983015346a6bd81c98d2d65252d
Temporal Sequence	AI Data Usages	Apr 30, 2023	Alan Jo	Alan Jo	May 11, 2023	[TSDB](https://texonom.com/tsdb-e2b5383c82d040ab981de62b970cc2ea)	### Temporal Sequence Notion \|Title\| \|:-:\| \|[DTW](https://texonom.com/dtw-ee045ba3b84e40b3b4e04d2257ed8814)\|	a2558b4ef95c4c37bee54056c1804c8d
Clearnlab	AI Data Tools	Jan 18, 2023	Alan Jo	Alan Jo	Jan 18, 2023		[cleanlab](https://github.com/cleanlab/cleanlab)	479203d2d2cf4ad2999da2d1fcad12ff
Eva db	AI Data Tools	May 2, 2023	Alan Jo	Alan Jo	May 2, 2023	[eva](https://github.com/georgia-tech-db/eva)	### Eva db Usages \|Title\| \|:-:\| ```typeSELECT id, data FROM TrafficVideo WHERE ['car'] <@ YoloV5(data).labels;``` > [EVA AI-Relational Database System documentation](https://evadb.readthedocs.io/)	17a183f2967d455fbe7631ea89172e49
C4 Data	NLP Datasets	Apr 18, 2023	Alan Jo	Alan Jo	Apr 18, 2023		Common Crawl의 웹 크롤링 코퍼스의 거대하고 깨끗한 버전	5e19c38f6e034c92857f71cb820dae89
Code Data	NLP Datasets	Apr 21, 2023	Alan Jo	Alan Jo	Apr 21, 2023		### Code Datas \|Title\| \|:-:\| \|[BigCode](https://texonom.com/bigcode-7c4e3d53d63d4f9a923910cfd999367b)\| \|[The Stack](https://texonom.com/the-stack-f5e306b910ae43de882a3bcd39981b6a)\|	071b31dcf1ee4073ad458f1ac781cd5f
CommonCrawl	NLP Datasets	Apr 18, 2023	Alan Jo	Alan Jo	Apr 18, 2023			e34ff04c04e14260a1ef19e3baa035c8
RedPajama Data	NLP Datasets	May 1, 2023	Alan Jo	Alan Jo	Jun 13, 2023	[MPT](https://texonom.com/mpt-bf2abacce4b14fe99929c1fb19ea2c71)	[RedPajama-Data](https://github.com/togethercomputer/RedPajama-Data) > [togethercomputer/RedPajama-Data-1T · Datasets at Hugging Face](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T) > [RedPajama, a project to create leading open-source models, starts by reproducing LLaMA training dataset of over 1.2 trillion tokens — TOGETHER](https://www.together.xyz/blog/redpajama)	ab8489ee6448474899c4b8e49b6faaac
BigCode	Code Datas	Apr 21, 2023	Alan Jo	Alan Jo	Apr 21, 2023		> [Open and responsible development of LLMs for code](https://www.bigcode-project.org/) > [The Stack: 3 TB of permissively licensed source code](https://arxiv.org/abs/2211.15533)	7c4e3d53d63d4f9a923910cfd999367b

StarChat

Huggingface H4

null

> [HuggingFaceH4/starchat-beta · Hugging Face](https://huggingface.co/HuggingFaceH4/starchat-beta) > [TheBloke/starchat-beta-GPTQ at main](https://huggingface.co/TheBloke/starchat-beta-GPTQ/tree/main)

ea316ec509564a51b56ad92b22831220

HuggingFace Docker Space

HuggingFace Space SDK

Jul 13, 2023

Alan Jo

Jul 13, 2023

0f250609ad1f493bb91146813de8d8a6

HuggingFace ChatUI Template

HuggingFace Docker Space Templates

Jul 13, 2023

Alan Jo

Jul 13, 2023

9cd1b4ab300f44b0bc91c87f33edf6bb

fastbook

Deep Learning Tools

Nov 25, 2019

Alan Jo

May 29, 2023

> [fastai/fastbook](https://github.com/fastai/fastbook)

d022baa785a14ce5af5cf1ef59995cb1

Netron

Deep Learning Tools

Oct 6, 2021

Alan Jo

May 29, 2023

### visualize Neural Network > [lutzroeder/netron](https://github.com/lutzroeder/netron)

2e410ef29aaa4a7d9e2a2777ce4dd3ee

Activation Function

Neural Network Components

Apr 27, 2023

Alan Jo

Jun 6, 2023

### Unlike sign function for [Back Propagation](https://texonom.com/back-propagation-18f4493692ad43449d4271f1bb293781), use non zero derivative - sigmoid function - probability - sign function - predict class ### Activation Functions |Title| |:-:| |[Sigmoid Function](https://texonom.com/sigmoid-function-417e17b995cf40c996cec78f699ea59b)| |[GELU](https://texonom.com/gelu-8f18023a80b84a06b5afd648811df6bc)| |[ReLU](https://texonom.com/relu-e582549804da48b893758895e446ffb9)| |[Sign Function](https://texonom.com/sign-function-b95e191cc2dc43c0befd6231f022555f)| |[ELU Function](https://texonom.com/elu-function-c7306a8566014b4e9d0733b52bd399fa)| |[Mish Function](https://texonom.com/mish-function-39eaae692b284adbadde08d3d707f13d)| |[Tanh Function](https://texonom.com/tanh-function-679abf35c4eb4fb0bb0e56140cac3d23)| |[Maxout Function](https://texonom.com/maxout-function-ad1f081fa0c34a6fbee08259c82ea4a1)| |[ELU Function](https://texonom.com/elu-function-e6f30f8fff2c422993a85565f8ce5ee7)| > [0023 Loss & Metric - Deepest Documentation](https://deepestdocs.readthedocs.io/en/latest/002_deep_learning_part_1/0023/)

8e52ee5f83a244d88abeeee3fb9497a8

Forward Forward Algorithm

Neural Network Components

Apr 3, 2023

Alan Jo

May 11, 2023

[Back Propagation](https://texonom.com/back-propagation-18f4493692ad43449d4271f1bb293781) 보다 대뇌피질 작동방식에 유사 시냅스의 처음으로 돌아가기 전에 cortical layer를 지나는 과정에서 *모종의 루프를 형성한다* 네트워크의 레이어 수가 많아질수록 FF가 Backprop보다 메모리를 덜 사용합니다. 하지만 20개 이하의 레이어를 가진 thin network에서는 FF가 backprop보다 훨씬 더 많은 메모리를 사용 ### Forward Forward Algorithm Notion |Title| |:-:| > [[딥러닝] The Forward-Forward Algorithm: Some Preliminary Investigations (FF)](https://velog.io/@nochesita/딥러닝-The-Forward-Forward-Algorithm-Some-Preliminary-Investigations) > [Welcome to JunYoung's blog | 딥러닝의 체계를 바꾼다! The Forward-Forward Algorithm 논문 리뷰 (1)](https://junia3.github.io/blog/ffalgorithm) > [The Forward-Forward Algorithm: Some Preliminary Investigations](https://arxiv.org/abs/2212.13345)

6c989313e382466bba02d15c412fb17f

Neural Network Layer

Neural Network Components

Apr 27, 2023

Alan Jo

Jun 7, 2023

### Do not count input layer as perceptron (no weights & bias) $$ learnable\;parameters = weights(edges) + perceptron(bias)$$ ### Neural Network Layer Notion |Title| |:-:| |[Layer Normalization](https://texonom.com/layer-normalization-c5111411c0594c04847b0eb4d8fa6682)| |[Batch Normalization](https://texonom.com/batch-normalization-0e5e919e98b74b168a248a9fd3e0cb63)| |[Fully Connected Layer](https://texonom.com/fully-connected-layer-3fd5a241e2b84a53a2339c60fbcd957a)| |[Input Layer](https://texonom.com/input-layer-abdd1cda0d124843a34c66dc5ec7eba1)| |[Hidden Layer](https://texonom.com/hidden-layer-a22cafa21026469094d260b473fd4789)| |[Output Layer](https://texonom.com/output-layer-184eb47c3a98437daceba11104377194)| |[Residual Connection](https://texonom.com/residual-connection-f6778b3cfbee4b0cab02f7fad533372f)|

e10ef2afe1954cf6b909f8aa40077393

skip connections

Neural Network Components

May 23, 2023

Alan Jo

May 23, 2023

**help dampen the impact of poor layer**

d7ad187f3487468db6eea278f0236a22

ELU Function

Activation Functions

Apr 27, 2023

Alan Jo

Apr 27, 2023

c7306a8566014b4e9d0733b52bd399fa

ELU Function

Activation Functions

Jun 7, 2023

Alan Jo

Jun 7, 2023

e6f30f8fff2c422993a85565f8ce5ee7

GELU

Activation Functions

Apr 8, 2023

Alan Jo

Apr 27, 2023

8f18023a80b84a06b5afd648811df6bc

Maxout Function

Activation Functions

Jun 7, 2023

Alan Jo

Jun 7, 2023

ad1f081fa0c34a6fbee08259c82ea4a1

Mish Function

Activation Functions

May 18, 2023

Alan Jo

May 18, 2023

39eaae692b284adbadde08d3d707f13d

ReLU

Activation Functions

Mar 7, 2023

Alan Jo

Jun 7, 2023

## Rectified Linear Unit 입력값이 음수인 뉴런은 다시 회생시키기 어렵다는 한계 이러한 문제를 Dying ReLU ### ReLU Variants |Title| |:-:| |[PReLU](https://texonom.com/prelu-67cdafbec19248bfad4225946a4c9fc3)| |[Leaky ReLU](https://texonom.com/leaky-relu-fa25fa8a234044068fd839a93512a762)|

e582549804da48b893758895e446ffb9

Sigmoid Function

Activation Functions

Nov 19, 2020

Alan Jo

Jun 14, 2023

### inverse of the natural logit function logistic은 s자 모양을 의미하는 시그모이드중 대표적 ReLU가 등장하기 이전에 활발하게 사용되었던 activation function hidden 노드 바로 뒤에 부착 ### Sigmoid Functions |Title| |:-:| |[SoftMax Function](https://texonom.com/softmax-function-d8364b5e839145e89df423e0c5365dc1)| |[Logit Function](https://texonom.com/logit-function-104e55a12b234813a020cb09ec51c754)| $$\sigma(x) = \frac{1}{1 + e^{-(wx + b)}}$$ > [어떻게 하면 "부드럽게" 연결할까?](https://youtube.com/watch?v=LS-C8KBvJYY&feature=shares) > [Logistic function](https://en.wikipedia.org/wiki/Logistic_function) > [logit, sigmoid, softmax의 관계 - 한 페이지 머신러닝](https://opentutorials.org/module/3653/22995)

417e17b995cf40c996cec78f699ea59b

Sign Function

Activation Functions

Apr 27, 2023

Alan Jo

Apr 27, 2023

b95e191cc2dc43c0befd6231f022555f

Tanh Function

Activation Functions

May 23, 2023

Alan Jo

May 23, 2023

679abf35c4eb4fb0bb0e56140cac3d23

Leaky ReLU

ReLU Variants

Apr 27, 2023

Alan Jo

May 23, 2023

음수일 때 출력값을 0이 아닌 0.001과 같은 매우 작은 값을 출력하도록 설정

fa25fa8a234044068fd839a93512a762

PReLU

ReLU Variants

May 18, 2023

Alan Jo

May 18, 2023

67cdafbec19248bfad4225946a4c9fc3

Logit Function

Sigmoid Functions

Nov 1, 2020

Alan Jo

Apr 27, 2023

$$g(z) = \frac{1}{1 + e^{-z}}$$ ### This is important $$g'(z) = g(z)(1 - g(z))$$ [Logit](https://texonom.com/logit-84e304d3bf104e3fbc6029bcc7eea6c2)

104e55a12b234813a020cb09ec51c754

SoftMax Function

Sigmoid Functions

Nov 1, 2020

Alan Jo

Apr 27, 2023

[Cross Entropy](https://texonom.com/cross-entropy-fefc71fd930a41a7842f39bccb3abcc9)

### softmax함수는 sigmoid의 일반형 - 2개 클래스를 대상으로 정의하던 logit을 K개의 클래스를 대상으로 일반화하면 softmax함수가 유도(derived) - softmax함수에서 K를 2로 놓으면 sigmoid함수로 환원 - sigmoid함수를 K개의 클래스를 대상으로 일반화하면 softmax함수가 유도 ### classification softmax함수는 인공신경망이 내놓은 K개의 클래스 구분 결과를 확률처럼 해석하도록 sigmoid는 activation에, softmax는 classification에 사용되지만 수학적으로 같다 > 다루는 클래스가 2개냐 K개냐로 차이가 있을 뿐 > [logit, sigmoid, softmax의 관계 - 한 페이지 머신러닝](https://opentutorials.org/module/3653/22995)

d8364b5e839145e89df423e0c5365dc1

Logit

Logit Function

null

‘logistic’ +‎ ‘probit’ > [logit, sigmoid, softmax의 관계 - 한 페이지 머신러닝](https://opentutorials.org/module/3653/22995)

84e304d3bf104e3fbc6029bcc7eea6c2

**Batch Normalization**

Neural Network Layer Notion

Mar 7, 2023

Alan Jo

Jul 6, 2023

[Layer Normalization](https://texonom.com/layer-normalization-c5111411c0594c04847b0eb4d8fa6682)

레이어마다 Normalization을 하는 레이어를 두어, 변형된 분포가 나오지 않도록 하는 것 ### Limitation 1. dependent to mini batch size 2. hard to apply to RNN ### **Batch Normalization Methods** |Title| |:-:| |[Internal Covariate Shift](https://texonom.com/internalcovariate-shift-827b94e336a141b4a2ca100fc42dab66)| |[Whitening](https://texonom.com/whitening-5bfaed3572904c208125cc9df6c664b0)| > [배치 정규화(Batch Normalization)](https://gaussian37.github.io/dl-concept-batchnorm/) > [Batch normalization](https://en.wikipedia.org/wiki/Batch_normalization)

0e5e919e98b74b168a248a9fd3e0cb63

Fully Connected Layer

Neural Network Layer Notion

Apr 27, 2023

Alan Jo

Jun 1, 2023

## FC Layer input node is connected to each input node is connected to each output node number of parameter is $m\times n$

3fd5a241e2b84a53a2339c60fbcd957a

Hidden Layer

Neural Network Layer Notion

May 11, 2023

Alan Jo

May 11, 2023

a22cafa21026469094d260b473fd4789

Input Layer

Neural Network Layer Notion

May 11, 2023

Alan Jo

May 11, 2023

abdd1cda0d124843a34c66dc5ec7eba1

Layer Normalization

Neural Network Layer Notion

Nov 19, 2019

Alan Jo

Jul 6, 2023

[Batch Normalization](https://texonom.com/batch-normalization-0e5e919e98b74b168a248a9fd3e0cb63)

BN normalizes the activations of each batch, while LN normalizes the activations of each layer

c5111411c0594c04847b0eb4d8fa6682

Output Layer

Neural Network Layer Notion

May 11, 2023

Alan Jo

May 11, 2023

184eb47c3a98437daceba11104377194

Residual Connection

Neural Network Layer Notion

Mar 7, 2023

Alan Jo

May 11, 2023

### **Skip connection** **하위 층에서 학습된 정보가 데이터 처리 과정에서 손실되는 것을 방지하기 위한 방법** > [[DL] Exploding & Vanishing Gradient 문제와 Residual Connection](https://heeya-stupidbutstudying.tistory.com/entry/DL-Exploding-Vanishing-gradient-문제와-Residual-Connection잔차연결)

f6778b3cfbee4b0cab02f7fad533372f

**Internal Covariate Shift**

Batch Normalization Methods

Jul 5, 2023

Alan Jo

Jul 6, 2023

Internal distribution could have bias or variance even if input data distribution is normalized > [[PyTorch로 시작하는 딥러닝 기초] 09-4 Batch-Normalization](https://wegonnamakeit.tistory.com/47)

827b94e336a141b4a2ca100fc42dab66

Whitening

Batch Normalization Methods

Mar 7, 2023

Alan Jo

Jul 6, 2023

들어오는 입력값의 특징들을 uncorrelated 하게 만들어주고, 각각의 분산을 1로 만들어주는 작업 **covariance matrix의 계산과 inverse의 계산이 필요하기 때문에 계산량이 많을 뿐더러, Whitening은 일부 파라미터들의 영향이 무시된다** **특정 파라미터가 계속 커지는 상태로 유지될수 있다는 문제도** > [[Deep Learning] Batch Normalization (배치 정규화)](https://eehoeskrap.tistory.com/430)

5bfaed3572904c208125cc9df6c664b0

ANN

Neural Networks

Mar 5, 2023

Alan Jo

Mar 6, 2023

### Artificial Neural Network 기계학습과 인지과학에서 생물학의 신경망에서 영감을 얻은 통계학적 학습 알고리즘

d4232205ecf9463c95a911d179c87a84

CNN

Neural Networks

Nov 5, 2019

Alan Jo

Seong-lae Cho

Jul 4, 2023

[Capsule Network](https://texonom.com/capsule-network-a14fc7e569864154aae1ef44106e8991) [BCI](https://texonom.com/bci-e7b6d2a55d19435b9d3f8990441cab9e)

## Convolutional Neural Network Actually we use [Correlation](https://texonom.com/correlation-d75ea902ec294bf19e1180b6bfd22137) not the [Convolution](https://texonom.com/convolution-54def9d34de64e199ea46393675af2ce) Convolutional Layer와 Pooling Layer를 번갈아 가며 적용하여 입력 이미지의 특징을 추출하고 (Fully Connected) Layer를 통해 분류 작업을 수행 ### Hyperparameters - Convolution layer’s filter type, size, stride which result each activation map - Placements and types of convolution layer, pooling layer and activation layer So the typical architecture look like this where N is usually up to ~5, M is large, 0 <= K <= 2. but recent advances challenge this paradigm ```type[(CONV-RELU)*N-POOL?]*M-(FC-RELU)*K,SOFTMAX``` ### CNN Notion |Title| |:-:| |[CNN History](https://texonom.com/cnn-history-76ffc514584846348136ea794f37ea9a)| |[Pooling Layer](https://texonom.com/pooling-layer-c36550932e8c4436b4b6da0d4d2fc946)| |[Convolutional Layer](https://texonom.com/convolutional-layer-0880b6c2089b4b20bb9a6232b6cb4b19)| |[CNN bottleneck layer](https://texonom.com/cnn-bottleneck-layer-c67427171be5400db198dd6e7ef644ce)| |[Training CNN](https://texonom.com/training-cnn-dc4f33cdf45544668d6f9cb6d059f431)| |[Local Contrast Normalization](https://texonom.com/local-contrast-normalization-036406a996c94edab249e7ff8a32b2bc)| |[ctivation map](https://texonom.com/ctivation-map-d0117ee2531b44dcbb0d19d33dff23ac)| ### CNN Models |Title| |:-:| |[ResNet](https://texonom.com/resnet-d8c7707445684e848ebef939500672d4)| |[ALexNet](https://texonom.com/alexnet-e33d746e06a540d4b8900be7ccede18c)| |[ConvNets](https://texonom.com/convnets-1846cf4fac8b4feea6fa8cf4ac083df9)| |[SuperVision](https://texonom.com/supervision-9d75ddb2bccd4dc9857f4332ae096298)| |[GoogLeNet](https://texonom.com/googlenet-64b5ad51aafc48d492ba489f47ffa58a)| |[VGG](https://texonom.com/vgg-c59b523e5e894bf2842474c9022dceec)| |[MSRA](https://texonom.com/msra-9024cb9533e54bdc9ada6d30601f847b)|

002bf81a77bc40d1858740d26b61d97b

FFNN

Neural Networks

Mar 5, 2023

Alan Jo

Mar 7, 2023

[Tree Data](https://texonom.com/tree-data-45f2e08b1f184e60b2e47c96e4cb535f) [RNN](https://texonom.com/rnn-f7aad56acb5542b2ac26c2908be4ce16)

### ***Feed-Forward Neural Network*** 고안된 인공 신경망의 최초의 가장 단순한 형태 ### FFNN Usages |Title| |:-:| |[Position-wise FFNN](https://texonom.com/position-wise-ffnn-6fe30c96aa9245d898c73ec34625377d)|

89ecee87d8b7482e86995950db90eb31

GNN

Neural Networks

Apr 30, 2023

Alan Jo

Apr 30, 2023

### **Graph neural network** 그래프로 나타낼 수 있는 데이터를 처리하기 위한 인공 신경망 기존의 특정 신경망 아키텍처는 적절하게 정의된 그래프에서 작동하는 GNN으로 해석가능 ### GNN Notion |Title| |:-:| > [Learnable Structural Semantic Readout for Graph Classification](https://arxiv.org/abs/2111.11523)

58adb81cb2b649d9af19019182960bb2

Perceptron

Neural Networks

Nov 5, 2019

Alan Jo

Seong-lae Cho

Jun 7, 2023

[Discriminant Function](https://texonom.com/discriminant-function-ed98e4c8d7284c34824f332770ccaf52) [Neuron](https://texonom.com/neuron-35ebe3514dc64647a74a0c3bad47b51b)

Can mimic OR function but cannot separate XOR by manipulating weights (flexibility) 입력값에 대한 이진 분류를 수행하거나 선형 예측기로 사용 based on only mistakes gradient descent update weight vector ### Perceptron Notion |Title| |:-:| |[Perceptron Criterion](https://texonom.com/perceptron-criterion-5524bf3059d244f2b2437e74120d1b2b)| |[Optimization for Perceptron Criterion](https://texonom.com/optimization-for-perceptron-criterion-e64193867ff34deebc0b319d13be3545)| |[Perceptron convergence theorem](https://texonom.com/perceptron-convergence-theorem-7837263544974463b5bcc6a812fc6dce)| |[Universal Approximation Theorem](https://texonom.com/universal-approximation-theorem-872ac98a16c94f1f8a29f49496a1dfcd)| |[Multi Layer Perceptron](https://texonom.com/multi-layer-perceptron-9de4da2c68f24d76824e792b0c949ab8)| |[Perceptron History](https://texonom.com/perceptron-history-24992383cc2c44718e86d1cb53133e41)|

1deb66f486d54c93bb928d8afba1864c

RNN

Neural Networks

Mar 12, 2021

Alan Jo

Jul 30, 2023

[Sequential Data](https://texonom.com/sequential-data-784908b963384da3b3b603ebf1068e4d) [Tanh Function](https://texonom.com/tanh-function-679abf35c4eb4fb0bb0e56140cac3d23) [Vanishing Gradient](https://texonom.com/vanishing-gradient-6dc21442840841b3879d6502c5efd98a)

## Recurrent Neural Networ sequential 하거나 time series data 에 적합한 모델 RNN기반 모델인 LSTM과 GRU는 구조적으로 고정 길이의 hidden state 벡터에 모든 단어의 의미를 담아야 하기 때문에 문장이 길어지면 모든 단어 정보를 고정 길이 벡터에 담기 어렵다는 문제 Input Seq’s information lost so 보정하기 위해 [Attention Mechanism](https://texonom.com/attention-mechanism-762711860abb45f59904f1ac4e4af285) 사용 어텐션을 RNN의 보정을 위한 용도로서 사용하는 것이 아니라 인코더 디코더에도 사용하게 ### RNN Notion |Title| |:-:| |[Recurrent Neurons](https://texonom.com/recurrent-neurons-341ce79b6a0648bb8c57fafc806573cc)| |[Long-term dependency](https://texonom.com/long-term-dependency-0b4e9db8d0cd4fc4889b83e32408e5f8)| |[RWKV](https://texonom.com/rwkv-639fb6fcbaba43e398a6e2fe5495373a)| |[Hidden State](https://texonom.com/hidden-state-7b65c9b953ce4a9b843d2a0b381a5453)| ### RNNs |Title| |:-:| |[LSTM](https://texonom.com/lstm-7946d20b7c394827b02e7b6b297bbadf)| |[GRU](https://texonom.com/gru-280a4b083b0b49989e3cc36ef493410c)| |[Bidirectional RNN](https://texonom.com/bidirectional-rnn-4f44262a3cc3417e8b31f24040cef871)| |[Vanilla RNN](https://texonom.com/vanilla-rnn-f4032d5ead424efaaf6e1fb210b546cf)| > [07-1.순환 신경망(RNN, Recurrent Neural Network) - (1)](https://excelsior-cjh.tistory.com/183)

f7aad56acb5542b2ac26c2908be4ce16

SNN

Neural Networks

May 30, 2022

Alan Jo

Mar 6, 2023

### **Spiking neural network** **Closely mimic natural neural networks** 자연 신경망을 더 가깝게 모방한 인공 신경망. 뉴런 및 시냅스 상태 외에도 SNN은 작동 모델에 시간 개념을 통합 ### SNN Notion |Title| |:-:| > [Untitled](http://www.koreascience.kr/article/JAKO202025465017052.pdf)

6cf69239a87e4df9a32b3494862374e4

ALexNet

CNN Models

Sep 14, 2020

Alan Jo

May 31, 2023

[ImageNet](https://texonom.com/imagenet-d5542cdb730f4e218bacb583b5e2576b)

### GPU CNN 구조를 기반으로 한 ALexNet, ILSVRC 오차율이 10%로 줄어든 혁신적인 알고리즘 > [CS231n 1강 요약](https://taeyoung96.github.io/cs231n/CS231n_1/)

e33d746e06a540d4b8900be7ccede18c

ConvNets

CNN Models

May 23, 2023

Alan Jo

May 23, 2023

CNN ⇒ [Local Contrast Normalization](https://texonom.com/local-contrast-normalization-6357cfe8ce294c69ae102053dcb5aaff) ⇒ [Pooling Layer](https://texonom.com/pooling-layer-c36550932e8c4436b4b6da0d4d2fc946)

1846cf4fac8b4feea6fa8cf4ac083df9

GoogLeNet

CNN Models

May 23, 2023

Alan Jo

Jun 7, 2023

2014

64b5ad51aafc48d492ba489f47ffa58a

MSRA

CNN Models

May 23, 2023

Alan Jo

Jun 7, 2023

2015

9024cb9533e54bdc9ada6d30601f847b

ResNet

CNN Models

May 30, 2023

Alan Jo

Jun 6, 2023

[CNN bottleneck layer](https://texonom.com/cnn-bottleneck-layer-c67427171be5400db198dd6e7ef644ce)

### Microsoft 2015 [COCO](https://texonom.com/coco-c7524d6e059444c09a5067e07caf94e6) > [ResNet - Azure Machine Learning](https://learn.microsoft.com/en-us/azure/machine-learning/component-reference/resnet?view=azureml-api-2) > [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) > [Pytorch로 ResNet 구현, torch summary 살펴보기](https://velog.io/@gibonki77/ResNetwithPyTorch)

d8c7707445684e848ebef939500672d4

SuperVision

CNN Models

May 23, 2023

Alan Jo

Jun 7, 2023

2012

9d75ddb2bccd4dc9857f4332ae096298

VGG

CNN Models

May 23, 2023

Alan Jo

Jun 7, 2023

2014

c59b523e5e894bf2842474c9022dceec

CNN bottleneck layer

CNN Notion

May 30, 2023

Alan Jo

Jun 7, 2023

## inception module 핵심은 1x1 Convolution 연산량을 최소화 연산량과 정보손실은 서로 tradeoff 관계라 연산량과 정보 손실의 관계를 잘 보면서 Error가 최소가 되는 지점 ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F59552677-4c80-46c9-b3fc-9678673bf88f%2FUntitled.png?table=block&id=4344af8f-0f0c-4470-bc0e-482e0f837efe&cache=v2) [ResNet](https://texonom.com/resnet-d8c7707445684e848ebef939500672d4) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Facc5b086-6258-4b9c-9aba-d9e0bcd5a220%2FUntitled.png?table=block&id=559e2275-d78b-4266-91ce-47404eab13c1&cache=v2) ### [GoogLeNet](https://texonom.com/googlenet-64b5ad51aafc48d492ba489f47ffa58a) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F75fa731a-374c-4ea7-b572-2086555c79ca%2FUntitled.png?table=block&id=8c337e84-43f0-47cf-acf7-9b683cebfc88&cache=v2) > [CNN의 Bottleneck에 대한 이해](https://velog.io/@lighthouse97/CNN의-Bottleneck에-대한-이해)

c67427171be5400db198dd6e7ef644ce

CNN History

CNN Notion

Sep 14, 2020

Alan Jo

May 30, 2023

1998년에 이미 CNN 알고리즘이 제안되었다.하지만 여러 문제들에 부딫혀 그 동안 발전하지 못했다. > 데이터의 양이 너무 적었다.연산 속도가 굉장히 느렸다. ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F12554e84-bfcd-4125-8bed-f31d172c4bf4%2FUntitled.png?table=block&id=4d668c12-9d63-4e37-a8b5-0b23fdeee53c&cache=v2) ### 1980 Fukushima “sandwich” architecture (SCSCSC…) > [Untitled](https://www.rctn.org/bruno/public/papers/Fukushima1980.pdf) ### 1998 [Yann LeCun](https://texonom.com/yann-lecun-7866102480474d939d45d97ddd318c3a), Bottou, Bengio, Haffner Gradient-based learning applied to document recognition > [Untitled](https://ieeexplore.ieee.org/document/726791) ### 2012 [ALexNet](https://texonom.com/alexnet-e33d746e06a540d4b8900be7ccede18c) [Geoffrey Hinton](https://texonom.com/geoffrey-hinton-441d5ce2b78146d0935454042b4f06d9), [Ilya Sutskever](https://texonom.com/ilya-sutskever-f38707abee824d21b2ec9824a1706342) ImageNet Classification with Deep Convolutional Neural Networks > [Untitled](https://proceedings.neurips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf)

76ffc514584846348136ea794f37ea9a

Convolutional Layer

CNN Notion

May 18, 2023

Alan Jo

Jun 7, 2023

### Share the same parameters across different locations (assume input stationary) In practice usually zero pad the border parameter 가 fully connected 나 locally connected 보다 훨씬 작아서 There are weight variables which are same as the size of filter including all that channels Using small filter is better [CNN Stride](https://texonom.com/cnn-stride-b84e6cf4abd347289dc8dbe6ffc1097d) 1x1 is also valid because we can reduce or increase the Output Channel of feature map with different scaling 채널별 필터 총합이 최종필터 **Commonly number of filters K is powers of 2** K가 늘면 지수적으로 연관 픽셀이 늘어나는게 아니라 선형적으로 가까운 거리로 늘어남 ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F2b27a771-bf6c-4c73-bb6c-6dc888a9d56e%2FUntitled.png?table=block&id=04353e66-16fb-48a3-a472-44bf24a7fe77&cache=v2) ### Multiple small sized filter layer’s parameters have much smaller than one large sized filter layer (also more nonlinearity, less compute) so we use mostly 3x3 after [VGG](https://texonom.com/vgg-c59b523e5e894bf2842474c9022dceec) ([ALexNet](https://texonom.com/alexnet-e33d746e06a540d4b8900be7ccede18c) used 11x11 or 7x7 convolution layer) 1x1 으로 더 줄일 수 있는데 그게 [CNN bottleneck layer](https://texonom.com/cnn-bottleneck-layer-c67427171be5400db198dd6e7ef644ce) ### total parameter $$filter \; size ^2 +1\;for\;bias$$

0880b6c2089b4b20bb9a6232b6cb4b19

ctivation map

CNN Notion

Jun 7, 2023

Alan Jo

Jun 18, 2023

입력 이미지의 각 위치에서 활성화된 뉴런의 출력값을 시각화한 것 모델이 어떤 부분에서 입력 이미지를 보고 판단하는지 이해하는 데 도움을 줌

d0117ee2531b44dcbb0d19d33dff23ac

Local Contrast Normalization

CNN Notion

Jun 7, 2023

Alan Jo

Jun 7, 2023

after convolution layer, before pooling layer ConvNet

036406a996c94edab249e7ff8a32b2bc

Pooling Layer

CNN Notion

Apr 30, 2023

Alan Jo

Jun 7, 2023

### [Image Downsampling](https://texonom.com/image-downsampling-97df12a6eb2949aa808c6f03dac101a6) feature map의 크기를 줄이는 데 사용 ### CNN Pooling Types |Title| |:-:| |[CNN Spatial Pooling](https://texonom.com/cnn-spatial-pooling-d191e60632c84b54b0266c2a73f54c51)| |[CNN Static Pooling](https://texonom.com/cnn-static-pooling-48534a9ed35e465c94694903dd2c80cb)| |[CNN Dynamic Pooling](https://texonom.com/cnn-dynamic-pooling-3412ef080ffb4e949bebfed80a17cf8d)| |[CNN Global Pooling](https://texonom.com/cnn-global-pooling-fe9913d4c6674a808ef6c899e778dcd4)|

c36550932e8c4436b4b6da0d4d2fc946

Training CNN

CNN Notion

Jun 1, 2023

Alan Jo

Jun 1, 2023

![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F0b17e758-45f3-40f9-986e-592b434eb581%2FUntitled.png?table=block&id=07c57c5d-4ab9-43d3-aab3-dc7fb384178c&cache=v2)

dc4f33cdf45544668d6f9cb6d059f431

CNN Stride

Convolutional Layer

null

How many pixels to skip $$Count = (N + 2P - F) / stride + 1$$ P is padding size, N is input size, F is filter size and count is output size+

b84e6cf4abd347289dc8dbe6ffc1097d

CNN Dynamic Pooling

CNN Pooling Types

Apr 30, 2023

Alan Jo

Apr 30, 2023

segment-level에서 feature를 집계 입력 데이터 내에서 고수준 특징들의 시간 정보를 더 잘 보존 > [Learnable Dynamic Temporal Pooling for Time Series Classification](https://arxiv.org/abs/2104.02577)

3412ef080ffb4e949bebfed80a17cf8d

CNN Global Pooling

CNN Pooling Types

Apr 30, 2023

Alan Jo

Apr 30, 2023

고수준 특징들의 버려질 수 있다 feature map 내의 모든 값을 집계하여 하나의 숫자로

fe9913d4c6674a808ef6c899e778dcd4

CNN Spatial Pooling

CNN Pooling Types

May 23, 2023

Alan Jo

Jun 7, 2023

Sum or max - max pooling - average pooling - min pooling - L2 pooling - L2 pooling over features

d191e60632c84b54b0266c2a73f54c51

CNN Static Pooling

CNN Pooling Types

Apr 30, 2023

Alan Jo

Apr 30, 2023

고정된 크기의 영역을 선택하여 그 영역 내의 값을 집계

48534a9ed35e465c94694903dd2c80cb

****Position-wise FFNN****

FFNN Usages

Mar 5, 2023

Alan Jo

Mar 7, 2023

## Fully-connected FFNN [Seq2Seq](https://texonom.com/seq2seq-01a9854dffa6417c87d92c11a607250c) 인코더와 디코더에서 공통적으로 가지고 있는 서브층 > [16-01 트랜스포머(Transformer)](https://wikidocs.net/31379)

6fe30c96aa9245d898c73ec34625377d

Multi Layer Perceptron

Perceptron Notion

Mar 5, 2023

Alan Jo

May 11, 2023

[Neural Network](https://texonom.com/neural-network-86f54f9f1de848c1a29c56c24f7d5094)

## MLP Can have non-linear decision boundary using 3 Perceptron, now available to separate XOR ### Two layer convex open or closed region ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Ffb96ba60-f3ee-44af-9457-8d166f25df99%2FUntitled.png?table=block&id=83c05706-0b14-4c08-a841-60ce8ac8a182&cache=v2) each line means perceptron ### Three layer arbitrary (complexity limited by number of neurons) ### Hidden layer except final layer

9de4da2c68f24d76824e792b0c949ab8

Optimization for Perceptron Criterion

Perceptron Notion

Apr 27, 2023

Alan Jo

Apr 27, 2023

### Maximize set of misclassified samples under w only update until convergence ### Weight Updates Start with weight 0 **For each training instance** > - If correct (i.e., y=y*), no change > - If wrong: adjust the weight vector - adding or subtracting the feature vector (f / y* exactly > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F96bde13f-ce87-46fb-942e-c9964ec7123f%2FUntitled.png?table=block&id=a35c861a-d054-4419-b50c-c672b7a6ae41&cache=v2) > > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F3e523c5e-9337-4c60-8453-1396f889d8a9%2FUntitled.png?table=block&id=fdbb578d-a453-4351-a61b-5efbfd34f979&cache=v2) > > x0 bias ican be tan add to f ### Error-Driven Linear Classification > Binary case - compare features to weight vector > Learning - figure out weight vector from examples

e64193867ff34deebc0b319d13be3545

Perceptron convergence theorem

Perceptron Notion

Apr 27, 2023

Alan Jo

Apr 27, 2023

Convergence is guaranteed for linearly separable data Cannot classify non-linearly separable data

7837263544974463b5bcc6a812fc6dce

Perceptron Criterion

Perceptron Notion

Apr 27, 2023

Alan Jo

Apr 27, 2023

### Inequality ### Linear Classifiers - Inputs **data **are feature values (**feature vector**) - (x0, x1, x2, ... , xn) - Each feature has a weight - Sum is the activation > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F97f9d462-0b77-464a-8d69-45b5d1a8d92b%2FUntitled.png?table=block&id=18b03cdd-2d54-42ea-8372-c57a7ff78e2a&cache=v2) - output - if activation is positive → 1 negative → -1 (classification +/-) ### Decision Rule binary decision rule One side corresponds to Y=+1 Other corresponds to Y=-1 ### Learning - binary perceptron 1. Start with weights = 0 > 1. For each training instance >> - Classify with current weights >> - If correct (i.e., y=y*), no change >> - If wrong: adjust the weight vector by adding or subtracting the feature vector. Subtract if y* is -1. > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fcc561d3b-0dae-4866-bf99-e3d2593dba1f%2FUntitled.png?table=block&id=81c5098b-4384-4058-9615-2678d672d6e7&cache=v2) ## Multiclass dicision rule ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F9af73bdd-20cd-4421-a317-ddd37127ae63%2FUntitled.png?table=block&id=49e83b0c-2442-4df3-acbb-88eefa82cb3f&cache=v2) lower count of wrong class** (for this instance), **raise count of right class **(for this instance)** ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fc30c5970-b6a9-4c98-a062-562a93d94adb%2FUntitled.png?table=block&id=d15b9c2d-8e80-4688-9a70-675be4443d7f&cache=v2) **black box (kernel) K** that told us the dot product of two examples x and x’ ### Solution 0. set true value to +/- 2. get data 3. choose true value 4. classification by hypothesis → compare to true value 5. update hypothesis to all data try this until no revise to all data redo these to data set ### multiclass 다른거 똑같은데 분류라서 argmax xf = y 틀리면 그놈은 값 빼주고 맞는놈은 올려주기 f만큼 듀얼이면 argmax alpha K(k is sigma ff) - at sigma only one f change by w if wrong 틀린놈은 알파에서 1 빼고 맞는놈들은 1만큼 올려줌 맞는 클래스 틀린 클래스 ### Feature Vectors ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fd1bbd38a-1fa3-43b3-96de-b4161118d84f%2FUntitled.png?table=block&id=0d54acc4-2dc6-4fdf-ba33-4f798cdf9949&cache=v2) Some (Simplified) Biology - Inputs are feature values - Each feature has a weight - Sum is the activation ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fd20f5cf0-dbee-4d69-b541-77430ef0677e%2FUntitled.png?table=block&id=ad108811-66aa-45fa-a82f-89af6936407a&cache=v2) If the activation is: Positive, output +1 Negative, output -1 ### Weights - Binary case: compare features to a weight vector - Learning: figure out the weight vector from examples ### Decision Rules Binary Decision Rule in the space of feature vectors > - Examples are points > - Any weight vector is a hyperplane > - One side corresponds to Y=+1 > - Other corresponds to Y=-1

5524bf3059d244f2b2437e74120d1b2b

Perceptron History

Perceptron Notion

Jun 7, 2023

Alan Jo

Jun 7, 2023

### 1957 released 1957 The Perceptron (Frank Rosenblatt) Mark I Perceptron The machine was connected to a camera that used 20×20 cadmium sulfide photocells to produce a 400-pixel image. Widrow and Hoff, ~1960: Adaline/Madaline 1969 Perceptrons (Minsky, Papert) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fef3314d0-ca74-4a67-bb3c-eae0eed6371d%2FUntitled.png?table=block&id=a63f0670-5848-468e-8eb6-2caf9025092c&cache=v2) > [Perceptron | Encyclopedia of Computer Science](https://dl.acm.org/doi/abs/10.5555/1074100.1074686)

24992383cc2c44718e86d1cb53133e41

Universal Approximation Theorem

Perceptron Notion

Apr 27, 2023

Alan Jo

Jun 13, 2023

### at least 1 hidden layer Any continuous function $f: [0, 1] \rightarrow [0, 1]$ can be approximated arbitrarily well by aneural network with at least 1 hidden layer with a finite number of weights One wide (latent) layer is enough, but it is just a memorizer (Cybenko ’89) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fd15809a9-4c4b-4e05-bc71-c1005175924b%2FUntitled.png?table=block&id=86addc25-55d1-4a93-984f-c728d5375d25&cache=v2)

872ac98a16c94f1f8a29f49496a1dfcd

Hidden State

RNN Notion

Mar 7, 2023

Alan Jo

Jun 21, 2023

[Latent Space](https://texonom.com/latent-space-d67b6bdef18b4058bfbc3d25f87ec087)

### latent state, internal state The hidden state is computed based on the current input and the previous hidden state **이 프로세스는 순차적이었고 병렬화를 방해** > [Some Intuition on Attention and the Transformer](https://eugeneyan.com/writing/attention/)

7b65c9b953ce4a9b843d2a0b381a5453

Long-term dependency

RNN Notion

Mar 6, 2023

Alan Jo

Mar 6, 2023

[GRU](https://texonom.com/gru-280a4b083b0b49989e3cc36ef493410c) [LSTM](https://texonom.com/lstm-7946d20b7c394827b02e7b6b297bbadf)

RNN Problem 문장이 길어질 경우 앞 단어의 정보를 잃어버리게 > [[딥러닝] 언어모델, RNN, GRU, LSTM, Attention, Transformer, GPT, BERT 개념 정리](https://velog.io/@rsj9987/딥러닝-용어정리)

0b4e9db8d0cd4fc4889b83e32408e5f8

Recurrent Neurons

RNN Notion

Mar 12, 2021

Alan Jo

Oct 6, 2021

순환 뉴런

341ce79b6a0648bb8c57fafc806573cc

RWKV

RNN Notion

May 28, 2023

Alan Jo

May 28, 2023

### parallelizable training ### RWKV Implementation |Title| |:-:| |[](https://texonom.com/4aea4d107bb245f280ce2ab0023546ce)| |[](https://texonom.com/8b87d4786fc6461b97ccc1250d23a9aa)| > [RWKV: Reinventing RNNs for the Transformer Era](https://arxiv.org/abs/2305.13048)

639fb6fcbaba43e398a6e2fe5495373a

[RWKV-LM](https://github.com/BlinkDL/RWKV-LM)

RWKV Implementation

May 28, 2023

Alan Jo

May 28, 2023

4aea4d107bb245f280ce2ab0023546ce

[rwkv.cpp](https://github.com/saharNooby/rwkv.cpp)

RWKV Implementation

May 28, 2023

Alan Jo

May 28, 2023

[GGML](https://texonom.com/ggml-d3297b7eedc846a1bea05cfa0f3cc24a)

8b87d4786fc6461b97ccc1250d23a9aa

Bidirectional RNN

RNNs

Jun 21, 2023

Alan Jo

Jun 21, 2023

- forwatd state - backward state

4f44262a3cc3417e8b31f24040cef871

GRU

RNNs

Oct 6, 2021

Alan Jo

Jul 6, 2023

## Gated recurrent unit 2014년 뉴욕대학교 조경현 교수님이 집필한 논문에서 제안 LSTM와 성능과 구조가 비슷하지만 GRU가 학습할 가중치가 적다 출력 게이트 없음 ### GRU Notion |Title| |:-:| |[GRU Update Gate](https://texonom.com/gru-update-gate-42daff55639e4781a5fee0170a13b2a2)| |[GRU Reset Gate](https://texonom.com/gru-reset-gate-89417d3e3bd44a5f89d67e199b36f381)| > [Gated Recurrent Units (GRU)](https://yjjo.tistory.com/18)

280a4b083b0b49989e3cc36ef493410c

LSTM

RNNs

Aug 16, 2021

Alan Jo

Jul 6, 2023

[Quant Trading](https://texonom.com/quant-trading-024f2adb817a4e33b9fc6905bba8eb0c)

## Long Short Term Memory RNN의 단점인 기울기 소실 문제를 해결하기 위해 gate를 추가한 모델 시계열 시퀀스가 있는 데이터에서 다음을 예측 이름에서 알 수 있는 단기 기억을 가지는 것처럼 행동하는 모델 특히 주식 예측에 많이 사용 ### LSTM Notion |Title| |:-:| |[LSTM Cell State](https://texonom.com/lstm-cell-state-128d65c00df5470a98d1603ae1e56bd0)| |[LSTM forget gate](https://texonom.com/lstm-forgetgate-ee99b6c3bb5b432295e2f54b1b58a9a1)| |[LSTM input gate](https://texonom.com/lstm-inputgate-02d2744fe7344bff9502e5bb111ff759)| - forget gate - input gate - output gate ### LSTMs |Title| |:-:| |[BiLSTM](https://texonom.com/bilstm-cc8242f116e643abb71167e859561991)| ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F4b058d6d-0010-47fe-95d8-9604f829655b%2FUntitled.png?table=block&id=c557c8ba-af7b-4ac2-929d-3e62c6599b79&cache=v2) ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F86665118-813b-4d35-a2d2-2f1923910e5c%2FUntitled.png?table=block&id=382fda06-da43-46d6-bb50-3da9736036d3&cache=v2) > [Long Short-Term Memory (LSTM) 이해하기](https://dgkim5360.tistory.com/entry/understanding-long-short-term-memory-lstm-kr) > [08-02 장단기 메모리(Long Short-Term Memory, LSTM)](https://wikidocs.net/22888) > [삼성전자 주식, 이더리움 시세 인공지능으로 예측해보자 - Python, Deep Learning](https://www.youtube.com/watch?v=sG_WeGbZ9A4&t=241s)

7946d20b7c394827b02e7b6b297bbadf

Vanilla RNN

RNNs

Jul 6, 2023

Alan Jo

Jul 6, 2023

짧은 시퀀스에만 효과 **the problem of Long-Term Dependencies** ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fa6b722a0-f76c-4dcf-a4dd-374531416413%2FUntitled.png?table=block&id=1c9f7c91-3cb8-4d9a-8fb7-04fa17238677&cache=v2) > [08-02 장단기 메모리(Long Short-Term Memory, LSTM)](https://wikidocs.net/22888)

f4032d5ead424efaaf6e1fb210b546cf

GRU Reset Gate

GRU Notion

Jul 6, 2023

Alan Jo

Jul 6, 2023

89417d3e3bd44a5f89d67e199b36f381

GRU Update Gate

GRU Notion

Jul 6, 2023

Alan Jo

Jul 6, 2023

42daff55639e4781a5fee0170a13b2a2

LSTM Cell State

LSTM Notion

Jul 6, 2023

Alan Jo

Jul 6, 2023

cell state에 뭔가를 더하거나 없앨 수 있는 능력 gate로 제어함

128d65c00df5470a98d1603ae1e56bd0

LSTM ****forget gate****

LSTM Notion

Jul 6, 2023

Alan Jo

Jul 6, 2023

ee99b6c3bb5b432295e2f54b1b58a9a1

LSTM ****input gate****

LSTM Notion

Jul 6, 2023

Alan Jo

Jul 6, 2023

02d2744fe7344bff9502e5bb111ff759

BiLSTM

LSTMs

Aug 16, 2021

Alan Jo

Jul 6, 2023

### Bidirectional LSTM

cc8242f116e643abb71167e859561991

AI Data

AI Usages

May 13, 2021

Alan Jo

Jun 1, 2023

[Model Generalization](https://texonom.com/model-generalization-0a2de5c4578c47dd9a5dc1b83fbaf72b) [Observation Noise](https://texonom.com/observation-noise-dcf85cf878d740b0ad7dd8ecab260165) [Data Processing](https://texonom.com/data-processing-d1b576ef34304e0fa6325db709860e0b)

### Data enhancement is more effective than code enhancement Don't try to be a hero, data is 80 ### AI Data Usages |Title| |:-:| |[AI Data Tool](https://texonom.com/ai-data-tool-3cb903994b5c4c2aad0e3fddb4511a3f)| |[Data Labeling](https://texonom.com/data-labeling-dda48db348de44a9bafa1ae69ca18133)| |[Data Augmentation](https://texonom.com/data-augmentation-fd2918b94bd54726a8c91e26af129545)| |[Data Testing](https://texonom.com/data-testing-bda06f8a3b0349ac8cb78d960f24bf46)| |[Experiment Management](https://texonom.com/experiment-management-cbea3983015346a6bd81c98d2d65252d)| |[Temporal Sequence](https://texonom.com/temporal-sequence-a2558b4ef95c4c37bee54056c1804c8d)| |[AI Dataset](https://texonom.com/ai-dataset-6383f9b5b0c54e598b90dc9ea0c8e5d1)| > [Introduction to streaming for data scientists](https://huyenchip.com//2022/08/03/stream-processing-for-data-scientists.html)

dcd9611d8a5647ba95e7bc63a7e6db1f

AI Development

AI Usages

Jun 13, 2021

Alan Jo

Jun 21, 2023

[Statistics Tool](https://texonom.com/statistics-tool-cd8ad9a0f6be46b4979d8e302c6dbd52)

> [Emerging Architectures for LLM Applications | Andreessen Horowitz](https://a16z.com/2023/06/20/emerging-architectures-for-llm-applications) > [Stack Overflow](https://stackoverflow.co/labs/developer-sentiment-ai-ml) > [Futurepedia - The Largest AI Tools Directory | Home](https://www.futurepedia.io/)

94a4b3b469b5439aaa03a669f894e3c8

AI Industry

AI Usages

Aug 21, 2021

Alan Jo

Jul 4, 2023

[Industry](https://texonom.com/industry-989c1bacc8b84fb1b63486f61e3b9052) [AI Object](https://texonom.com/ai-object-e69c83e69fd34103a83c1341b65158be) [Sex Industry](https://texonom.com/sex-industry-e5080bf012e0478f9e6fc996bdc8a2cb) [Copyright](https://texonom.com/copyright-02b119d69a0a475289bbcd13faae44f3)

AI progress is constrained by bottlenecks in essential economic sectors and steps of the innovation process, and that breakthroughs in manipulating the physical world are necessary 현재까지 AI 창작물은 저작권 없다는 걸 이용해야 ### AI Companies |Title| |:-:| |[OpenAI](https://texonom.com/openai-501e879eb1ca4ef8b388ba5b11e92e23)| |[Deepmind](https://texonom.com/deepmind-5eb171c77b344d4786a9a5b23ae70eca)| |[Stability AI](https://texonom.com/stability-ai-7285625019ec4498a7300ac9f0736b2d)| |[Anthropic AI](https://texonom.com/anthropic-ai-096e36d827a2421b9b0ee8d1ff3b0305)| |[AllenAI](https://texonom.com/allenai-930c776c8fdd4e358032803f037cd6fa)| |[Saltlux](https://texonom.com/saltlux-de9945849d70489f99a33f81a1707a46)| |[Neural Magic](https://texonom.com/neural-magic-021ab1b8be0d4f95b8ae6278c08e1562)| |[VoyagerX](https://texonom.com/voyagerx-bdc1dc8597e44f1aac3ab3ec6256eb87)| |[Upstage.ai](https://texonom.com/upstageai-319072815ced47bf98335f3a9c87cec4)| |[CarperAI](https://texonom.com/carperai-d69725f3dcfb4a57a9cf0f4bea30bc2a)| |[Eleuther AI](https://texonom.com/eleuther-ai-9f89e25571fe4930825ffabb3483c02a)| |[Preligens](https://texonom.com/preligens-040e14875d3b463fa71674578756944e)| |[inworld](https://texonom.com/inworld-29d82633e51742dabca809cc86d6e989)| |[Scatter Lab](https://texonom.com/scatter-lab-9524953c003c4c7d8ec48d387c1a42cf)| |[VAIV](https://texonom.com/vaiv-15772f18eca5478290a6ca76068cf077)| |[X.AI](https://texonom.com/xai-ccd1c9778b1349c788889cf20c1c9a72)| |[Inflection AI](https://texonom.com/inflection-ai-533110523b9f4b8ab79f663a6ba7f6b0)| |[MosaicML](https://texonom.com/mosaicml-d531ff0e6a51448c84296ab20f675d67)| |[Tiny corp](https://texonom.com/tiny-corp-2b42b01b901d4a93b80698286c5e2b40)| |[Cohere](https://texonom.com/cohere-5fc93cb3678841a7b23df3c7d0399aca)| |[Mind AI](https://texonom.com/mind-ai-007aca806190459c8437ce7046f5fbcc)| ### AI Products Diagram > [How to Build AI Products People Want — Reforge](https://www.reforge.com/blog/ai-products-arms-race) ### Industry Landscape > [Why transformative AI is really, really hard to achieve](https://zhengdongwang.com/2023/06/27/why-transformative-ai-is-really-really-hard-to-achieve.html) > [LF AI Landscape](https://landscape.lfai.foundation/) ### Scaling > [3 ways AI is scaling helpful technologies worldwide](https://blog.google/technology/ai/ways-ai-is-scaling-helpful/) > [AI Will Create More Developers, Not Less](https://interconnect.substack.com/p/ai-will-create-more-developers-not)

d8709bd0498145e3a66af6da3f963fa7

AI Object

AI Usages

Jun 13, 2021

Alan Jo

May 21, 2023

[ai-notes](https://github.com/sw-yx/ai-notes) [applied-ml](https://github.com/eugeneyan/applied-ml)

### AI 특성상 정확성이 필요없는 분야부터 대체될 것이고 그렇게 진행되고 있다 ### AI Objects |Title| |:-:| |[AI Planning](https://texonom.com/ai-planning-2173477bde2147d1a91975e30756806a)| |[Vision AI](https://texonom.com/vision-ai-ba1da978d6f6488998e7cdc59431b41a)| |[Sound AI](https://texonom.com/sound-ai-c3f85fde353c4e7581cca5e3e30324f6)| |[IA](https://texonom.com/ia-92af447425bc4fe2b10f3aef85c1871e)| |[NLP](https://texonom.com/nlp-e0ae7b40dd23463ea7bc92195d6ec7fd)| |[AI Coding](https://texonom.com/ai-coding-f659b252482d4dfb90485c4a0428f442)| |[Recommend System](https://texonom.com/recommend-system-e94b364527554e65bfd9d3dc63fb1af7)| |[AI Invent](https://texonom.com/ai-invent-2fd3b79c01bd43239990cd547a2f6e41)| |[AI Game](https://texonom.com/ai-game-7cb70519f66942b1b59d1d7352ce34e1)| |[AI Automation](https://texonom.com/ai-automation-afa72ebcb9e4426799162de10da077f8)| |[AI Robot](https://texonom.com/ai-robot-0796068b90b246d8abc6aeed851624cd)| |[Multimodal AI](https://texonom.com/multimodal-ai-381b5ebe1218438689de97941a3d9829)| |[AI Agent](https://texonom.com/ai-agent-5efeaf8cd77f495988083eb2084d7f01)| |[Extract Information](https://texonom.com/extract-information-6988fe9ed1a444e7be50249e5ccf11bc)|

e69c83e69fd34103a83c1341b65158be

AI Protocol

AI Usages

Apr 11, 2023

Alan Jo

Apr 11, 2023

### AI Protocols |Title| |:-:|

be3148d3c1b24b90a4f51020fde30bd6

AI Data Tool

AI Data Usages

Jan 18, 2023

Alan Jo

Jan 18, 2023

3cb903994b5c4c2aad0e3fddb4511a3f

AI Dataset

AI Data Usages

Mar 27, 2021

Alan Jo

Jun 6, 2023

[Text Corpus](https://texonom.com/text-corpus-1befe731759849c48134768343a90703)

We typically say that a dataset is high-dimensional if the number of data points N issmaller than the dimensionality D > - not cheatable > - large degree of intra-class variability > > ![](https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2F895def36-e573-4dd1-8322-2a360807feb7%2FUntitled.png?table=block&id=fcecf613-dddd-441d-bf75-31e18d3ec61a&cache=v2) ### Vision AI Datasets |Title| |:-:| |[ImageNet](https://texonom.com/imagenet-d5542cdb730f4e218bacb583b5e2576b)| |[WebLI](https://texonom.com/webli-5b0455624497488ba55432160ad812c6)| ### NLP Datasets |Title| |:-:| |[CommonCrawl](https://texonom.com/commoncrawl-e34ff04c04e14260a1ef19e3baa035c8)| |[C4 Data](https://texonom.com/c4-data-5e19c38f6e034c92857f71cb820dae89)| |[Code Data](https://texonom.com/code-data-071b31dcf1ee4073ad458f1ac781cd5f)| |[RedPajama Data](https://texonom.com/redpajama-data-ab8489ee6448474899c4b8e49b6faaac)| > [20 Open Datasets for Natural Language Processing](https://odsc.medium.com/20-open-datasets-for-natural-language-processing-538fbfaf8e38)

6383f9b5b0c54e598b90dc9ea0c8e5d1

Data Augmentation

AI Data Usages

Jun 22, 2021

Alan Jo

Jun 7, 2023

### Flip, Scale, Crop, Rotation Jitter, Shear, Stretch, Lens Distortion ### Data Augmentation Tools |Title| |:-:| |[AI Data Generator](https://texonom.com/ai-data-generator-dd7e73ff30db4c0f89580aff9fb7df56)| |[AugLy](https://texonom.com/augly-0cbb390ff79642d1846f50f79ea9ed91)| ### Data Augmentation Notion |Title| |:-:| |[color jitter](https://texonom.com/color-jitter-08229f0a960c43acb65cedad8e7fb14a)|

fd2918b94bd54726a8c91e26af129545

Data Labeling

AI Data Usages

May 13, 2021

Alan Jo

May 11, 2023

### Data Labelings |Title| |:-:| |[Image Labeling](https://texonom.com/image-labeling-dca8afabd8a14dc495545f414883d7b4)| ### Cheap Africa human power > [Africa: The Hidden Workforce Behind AI](https://www.mantralabsglobal.com/blog/ai-in-africa-artificial-intelligence-africa/)

dda48db348de44a9bafa1ae69ca18133

Data Testing

AI Data Usages

Mar 21, 2023

Alan Jo

May 11, 2023

bda06f8a3b0349ac8cb78d960f24bf46

Experiment Management

AI Data Usages

Jun 13, 2021

Alan Jo

May 11, 2023

### Experiment Management Tools |Title| |:-:| |[Neptune.ai](https://texonom.com/neptuneai-c61383aff215447a979fd6d446be277d)| |[Weights & Biases](https://texonom.com/weights-biases-51d1cbfc87324d1f8dd43a8978ea5f68)| |[TensorBoard](https://texonom.com/tensorboard-6af2721a16bd4dcdb172affcdf055d91)|

cbea3983015346a6bd81c98d2d65252d

Temporal Sequence

AI Data Usages

Apr 30, 2023

Alan Jo

May 11, 2023

[TSDB](https://texonom.com/tsdb-e2b5383c82d040ab981de62b970cc2ea)

a2558b4ef95c4c37bee54056c1804c8d

Clearnlab

AI Data Tools

Jan 18, 2023

Alan Jo

Jan 18, 2023

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

479203d2d2cf4ad2999da2d1fcad12ff

Eva db

AI Data Tools

May 2, 2023

Alan Jo

May 2, 2023

[eva](https://github.com/georgia-tech-db/eva)

### Eva db Usages |Title| |:-:| ```typeSELECT id, data FROM TrafficVideo WHERE ['car'] <@ YoloV5(data).labels;``` > [EVA AI-Relational Database System documentation](https://evadb.readthedocs.io/)

17a183f2967d455fbe7631ea89172e49

C4 Data

NLP Datasets

Apr 18, 2023

Alan Jo

Apr 18, 2023

Common Crawl의 웹 크롤링 코퍼스의 거대하고 깨끗한 버전

5e19c38f6e034c92857f71cb820dae89

Code Data

NLP Datasets

Apr 21, 2023

Alan Jo

Apr 21, 2023

071b31dcf1ee4073ad458f1ac781cd5f

CommonCrawl

NLP Datasets

Apr 18, 2023

Alan Jo

Apr 18, 2023

e34ff04c04e14260a1ef19e3baa035c8

RedPajama Data

NLP Datasets

May 1, 2023

Alan Jo

Jun 13, 2023

[MPT](https://texonom.com/mpt-bf2abacce4b14fe99929c1fb19ea2c71)

[RedPajama-Data](https://github.com/togethercomputer/RedPajama-Data) > [togethercomputer/RedPajama-Data-1T · Datasets at Hugging Face](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T) > [RedPajama, a project to create leading open-source models, starts by reproducing LLaMA training dataset of over 1.2 trillion tokens — TOGETHER](https://www.together.xyz/blog/redpajama)

ab8489ee6448474899c4b8e49b6faaac

BigCode

Code Datas

Apr 21, 2023

Alan Jo

Apr 21, 2023

> [Open and responsible development of LLMs for code](https://www.bigcode-project.org/) > [The Stack: 3 TB of permissively licensed source code](https://arxiv.org/abs/2211.15533)

7c4e3d53d63d4f9a923910cfd999367b