Deep Learning

[youjun99]

[Deep Learning Papers Reading Roadmap]을 따라 첫 번째 논문 요약합니다! LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. "Deep learning." Nature 521.7553 (2015): 436-444.

Deep Learning -multiple processing layers -learn representations of data through abstraction -->SotA speech recognition, visual object recognition, object detection, drug discovery, genomics, etc. -method: backpropagation algorithm, deep convolutional nets(images/video/speech/audio)(CNN), recurrent nets(text/speech)(RNN)

Machine Learning -limits: processing raw data -->Representation Learning : automatically discover representations for detection/classification *layers of features<--general-purpose learning procedure

Supervised Learning -most common form of ML -CLASSIFICATION--"label" [input-output]; weight vector <-gradient vector, objective function ->simple; *SGD(Stochastic Gradient Descent): AVERAGE of gradients (~stop decreasing)

*Classifier ML; linear classifier! (pixel) ---"Selectivity Invariance Dilemma" DL; NON-linear

Backpropagation <- *Chain Rule "poor local MINIMA"; X big deal

**CNN(Convolutional Neural Network) / ConvNet (2): Convolution Layer(Feature map), Pooling Layer

Distributed Representations and Language Processing 2 advantages; 1.generalization to new combinations / 2. another exponential advantage

RNN(Recurrent Neural Networks) -sequential inputs(speech, language) *"hidden state"

Future of DL Unsupervised Learning!!!! Natural Language Understanding <- *RNN

[minha62]

전반적인 흐름을 볼 수 있어 좋았습니다! 좋은 내용 감사합니다:)