AI (16) 썸네일형 리스트형 CS231n - Lec14. Deep Reinforcement Learning What is reinforce learning? agent: 행동 environment: 보상 , next state Markow Decision Process - RL problem formalism MDP는 Markov property를 만족함 Markow property : 현재 상태만으로 전체 상태를 나타내는 성질 discount factor: 보상받는 시간에 대해 우리가 얼마나 중요하게 생각할지 정책 pi : 각 상태에서 agent가 어떤 행동을 취할지 명시 MDP 정의 => 최적의 정책 pi* 찾아야함 (보상의 합 최대화) 우리가 어떤 상태에 있더라도 그 상황에서 보상을 최대화시킬 수 있는 행동 알려줌 Value Function and Q-Value function : definition 상.. CS231n- Lec13. Generative Models Unsupervised Learning just data, no labels Goal: Learn some underlying hidden structure of the data Examples: Clustering, dimensionality reduction, feature learning, density estimation, etc. Generative Models training data 와 같은 분포에서 새로운 sample을 생성하는 model model's prob(x)가 최대한 data's prob(x)와 가깝게 만드는것이 목표 super resolution, colorization ... PixelRNN and PixelCNN PixelRNN fully visible brief networ.. CS231n - Lec12. Visualizing and Understanding Whats going on inside convnets? what types of things in the image they are looking for analyzing this internals of the network? common criticism of deep learning : to trust network, understand the nature, black box of convnets Visualizing Activations First Layer: Visualize Filters AlexNet conv filter 3x11x11 입력 이미지와 직접 내적 수행하기때문에 이 필터가 이미지에서 무엇을 찾고있는지 알아낼 수 있음 64개의 11x11 filter oriented edges, o.. CS231n - Lec11. Detection and Segmentation Other computer vision tasks Semantic Segmentation output the pixel of segment grass/cat/tree/sky (category label) label each pixel in the image with a category label dont defferentiate instances, only care about pixels idea: sliding window take crops, what is the center of this crop , apply on crop rather than whole image this is computationally expensive idea2 : fully convolutional input 3 x H .. CS231n - Lec10. Recurrent Neural Networks 1 image - fixed size output vector 2 image captioning 3 sentiment classification (positive/negative) 4 translation 5 video classification (each frame) sequential processing of fixed outputs RNN 가변 input 가변 output input - RNN (has internal hidden state) -output (usually wnat to predict a vetctor at some time steps) vanilla recurrent neural network ( a state consists of a single hidden vector h ) .. CS231n - Lec9. CNN Architecture LeNet input => conv & pooling => FC layer conv filter = 5x5 , stride=1 pooling layer = 2x2, stride=2 Case Studies AlexNet 2012 ILSVRC'12 why so much nicer than before: fist deep learning and conv net approach First Large scale CNN ImageNet classification task를 잘 수행함 Input : 227x227x3 FirstLayer(Conv1) : 96 11x11 filters, stride=4 output vol - [55x55x96] parameters - (11x11x3)x96 = 35k SecondLaye.. CS231n-Lec8. Deep Learning Software CPU vs GPU CUDA (NVIDIA only) write C-like code that runs directly on the GPU highter-level APIs: cuDNN, cuBLAS, etc.. OpenCL similar to CUDA, but on anything slow cuDNN much faster than unoptimized CUDA Deep Learning Frameworks Caffe / Caffe2 Theano/TensorFlow Torch/PyTorch The Point of deep learning frameworks (1) Easily build big computational graphs (2) Easily compute gradients in computatio.. CS231n - Lec7. Training Neural Networks 2 Fancier Optimization Stochastic Gradient Descent while True: while True: weight_grad = evaluate_gradient(loss_fun, data, weights) weight += -step_size * weights_grad Problems with SGD (1) Loss function has condition number : ratio of largest to smallest singular value of the Hessian matrix is large (sensitive to one direction and not sensitive to the other direction) very slow progress along sha.. 이전 1 2 다음
