Output g f after looking at data ( x n , y n ). 2. Can We do it? E - - PowerPoint PPT Presentation

SLIDE 1

Learning From Data Lecture 15 Reflecting on Our Path - Epilogue to Part I

What We Did The Machine Learning Zoo Moving Forward

• M. Magdon-Ismail

CSCI 4100/6100 recap: Three Learning Principles Occam’s razor: simpler is better; falsifiable. Scientist 1 Scientist 2 Scientist 3 temperature T resistivity ρ temperature T resistivity ρ temperature T resistivity ρ not falsifiable falsifiable Sampling bias: ensure that training and test distributions are the same, or else acknowl- edge/account for it. You cannot sample from one bin and use your estimates for another bin. Data snooping: you are charged for every choice influenced by D. Choose the learning process (usually H) before looking at D. We know the price of choosing g from H. ? h ∈ H ? ? g Data D your choices

− → g

c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 2 /11 Zen Moment− →

• Zen Moment

c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 3 /11 Our Plan − →

Our Plan

• 1. What is Learning?

Output g ≈ f after looking at data (xn, yn).

• 2. Can We do it?

Ein ≈ Eout simple H, finite dvc, large N Ein ≈ 0 good H, algorithms

• 3. How to do it?

Linear models, nonlinear transforms Algorithms: PLA, pseudoinverse, gradient descent

• 4. How to do it well?

Overfitting: stochastic & deterministic noise Cures: regularization, validation.

• 5. General principles?

Occams razor, sampling bias, data snooping

• 6. Advanced techniques.
• 7. Other Learning Paradigms.

concepts theory practice c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 4 /11 LFD Jungle − →

SLIDE 2

Learning From Data: It’s A Jungle Out There

• verfitting

stochastic noise K-means stochastic gradient descent exploration reinforcement exploitation augmented error ill-posed Gaussian processes bootstrapping Lloyds algorithm deterministic noise distribution free learning data snooping Q-learning unlabelled dataexpectation-maximization logistic regression Rademacher complexity linear regression CART bagging Bayesian VC dimension transfer learning learning curve gans sampling bias neural networks Markov Chain Monte Carlo (MCMC) nonlinear transformation Mercer’s theorem support vectors Gibbs sampling decision trees adaboost SVM graphical models bioinformatics linear models

• rdinal regression

training versus testing no free lunch extrapolation DEEP LEARNING cross validation HMMs bias-variance tradeoff PAC-learning biometrics error measures MDL multiclass

• ne versus all

active learning types of learning random forests unsupervised weak learning

• nline-learning

RBF is learning feasible? data contamination perceptron learning noisy targets ranking momentum Occam’s razor conjugate gradients Levenberg-Marquardt RKHS kernel methods mixture of experts boosting ensemble methods AIC permutation complexity multi-agent systems classification primal-dual PCALLE kernel-PCA colaborative filtering semi-supervised learning clustering regularization weight decay Big Data Boltzmann machine c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 5 /11 Theory − →

Navigating the Jungle: Theory

THEORY VC-analysis bias-variance complexity Bayesian Rademacher SRM . . . c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 6 /11 Techniques − →

Navigating the Jungle: Techniques

THEORY VC-analysis bias-variance complexity Bayesian Rademacher SRM . . . TECHNIQUES Models Methods c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 7 /11 Models − →

Navigating the Jungle: Models

THEORY VC-analysis bias-variance complexity Bayesian Rademacher SRM . . . TECHNIQUES Models linear neural networks SVM similarity Gaussian processes graphical models bilinear/SVD . . . Methods c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 8 /11 Methods − →

SLIDE 3

Navigating the Jungle: Methods

THEORY VC-analysis bias-variance complexity Bayesian Rademacher SRM . . . TECHNIQUES Models linear neural networks SVM similarity Gaussian processes graphical models bilinear/SVD . . . Methods regularization validation aggregation preprocessing . . . c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 9 /11 Paradigms − →

Navigating the Jungle: Paradigms

THEORY VC-analysis bias-variance complexity Bayesian Rademacher SRM . . . TECHNIQUES Models linear neural networks SVM similarity Gaussian processes graphical models bilinear/SVD . . . Methods regularization validation aggregation preprocessing . . . PARADIGMS supervised unsupervised reinforcement active

• nline

unlabeled transfer learning big data . . . c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 10 /11 Moving Forward − →

Moving Forward

• 1. What is Learning?

Output g ≈ f after looking at data (xn, yn).

• 2. Can We do it?

Ein ≈ Eout simple H, finite dvc, large N Ein ≈ 0 good H, algorithms

• 3. How to do it?

Linear models, nonlinear transforms Algorithms: PLA, pseudoinverse, gradient descent

• 4. How to do it well?

Overfitting: stochastic & deterministic noise Cures: regularization, validation.

• 5. General principles?

Occams razor, sampling bias, data snooping

• 6. Advanced techniques.

Similarity, neural networks, SVMs, preprocessing & aggregation

• 7. Other Learning Paradigms.

Unsupervised, reinforcement concepts theory practice c A M L Creator: Malik Magdon-Ismail Reflecting on Our Path: 11 /11