Non-Linearity. CS 188: Artificial Intelligence. Non-Linear Separators. Non-Linear Separators. Deep Learning I

Transcription

1 Non-Linearity CS 188: Artificial Intelligence Deep Learning I Instructors: Pieter Abbeel & Anca Dragan --- University of California, Berkeley [These slides were created by Dan Klein, Pieter Abbeel, Anca Dragan for CS188 Intro to AI at UC Berkeley. All CS188 materials are available at ai.berkeley.edu.] Non-Linear Separators Data that is linearly separable works out great for linear decision rules: 0 x But what are we going to do if the dataset is just too hard? Non-Linear Separators General idea: the original feature space can always be mapped to some higherdimensional feature space where the training set is separable: Φ: x φ(x) 0 x How about mapping data to a higher-dimensional space: x 2 0 x This and next slide adapted from Ray Mooney, UT

2 Feature Selection Computer Vision To choose between two feature sets: For feature set 1: train perceptron on training data -> Classifier 1 For feature set 2: train perceptron on training data -> Classifier 2 Evaluate performance of Classifier 1 and Classifier 2 on hold-out data Select the one performing best on the hold-out data Object Detection Manual Feature Design

3 Features and Generalization Features and Generalization Image HoG [Dalal and Triggs, 2005] Manual Feature Design! Deep Learning Perceptron Manual feature design requires: Domain-specific expertise Domain-specific effort What if we could learn the features, too? -> Deep Learning w 1 w 2 w 3

4 Two-Layer Perceptron Network N-Layer Perceptron Network w 11 w 21 w 31 w 1 w 12 w 22 w 32 w 2 w 13 w 3 w 23 w 33 Performance Speech Recognition AlexNet graph credit Matt Zeiler, Clarifai graph credit Matt Zeiler, Clarifai

5 N-Layer Perceptron Network Local Search Simple, general idea: Start wherever Repeat: move to the best neighboring state If no neighbors better than current, quit Neighbors = small perturbations of w Properties Plateaus and local optima How to escape plateaus and find a good local optimum? How to deal with very large parameter vectors? E.g., Perceptron Soft-Max Score for y=1: Score for y=-1: w 1 w 2 w 3 Probability of label: (predict) Objective: Classification Accuracy Objective: Issue: many plateaus! how to measure incremental progress? Log:

6 Two-Layer Neural Network N-Layer Neural Network w 11 w 21 w 31 w 1 w 12 w 22 w 32 w 2 w 13 w 3 w 23 w 33 Our Status 1-d optimization Our objective Changes smoothly with changes in w Doesn t suffer from the same plateaus as the perceptron network Challenge: how to find a good w? Could evaluate Then step in best direction and Equivalently: Or, evaluate derivative: Which tells which direction to step into

7 2-D Optimization Steepest Descent Idea: Start somewhere Repeat: Take a step in the steepest descent direction Source: Thomas Jungblut s Blog Figure source: Mathworks Generally, Steepest Direction Steepest Direction = direction of the gradient Optimization Procedure 1: Gradient Descent Init: For i = 1, 2, : learning rate --- tweaking parameter that needs to be chosen carefully How? Try multiple choices Crude rule of thumb: update changes about %

8 Try Many Learning Rates Suppose loss function is steep vertically but shallow horizontally: Q: What is the trajectory along which we converge towards the minimum with SGD? 29 Fei-Fei Li & Andrej Karpathy & Justin Johnson Lecture Jan 2016 Suppose loss function is steep vertically but shallow horizontally: Suppose loss function is steep vertically but shallow horizontally: Q: What is the trajectory along which we converge towards the minimum with SGD? Fei-Fei Li & Andrej Karpathy & Justin Johnson Lecture Jan 2016 Q: What is the trajectory along which we converge towards the minimum with Gradient Descent? very slow progress along flat direction, jitter along steep one Fei-Fei Li & Andrej Karpathy & Justin Johnson Lecture Jan 2016

9 Optimization Procedure 2: Momentum Suppose loss function is steep vertically but shallow horizontally: Gradient Descent Init: For i = 1, 2, Momentum Init: For i = 1, 2, Q: What is the trajectory along which we converge towards the minimum with Momentum? - Physical interpretation as ball rolling down the loss function + friction (mu coefficient). - mu = usually ~0.5, 0.9, or 0.99 (Sometimes annealed over time, e.g. from 0.5 -> 0.99) Fei-Fei Li & Andrej Karpathy & Justin Johnson Lecture Jan 2016 Some More Advanced Optimization Methods* Remaining Pieces Nesterov Momentum Update AdaGrad RMSProp ADAM L-BFGS Optimizing machine learning objectives: Stochastic Descent Mini-batches Improving generalization Drop-out Activation functions Initialization Renormalization Computing the gradient Backprop Gradient checking