Epilogue: what we covered and didn`t cover in this course
Transcription
Epilogue: what we covered and didn`t cover in this course
12/5/13 What I hope you got out of this course Epilogue: what we covered and didn’t cover in this course The machine learning toolbox ʻViagraʼ =0 ʻlotteryʼ =0 ĉ(x) = ham =1 ĉ(x) = spam Formulating a problem as an ML problem Understanding a variety of ML algorithms Running and interpreting ML experiments Understanding what makes ML work – theory and practice =1 ĉ(x) = spam 16 14 12 10 8 6 4 2 0 2 4 6 8 10 12 14 1 Learning scenarios we covered Types of models Classification: discrete/categorical labels Geometric q Ridge-regression, SVM, perceptron w wTx + b > 0 wTx + b < 0 Distance-based Regression: continuous labels q Clustering: no labels Probabilistic 16 q K-nearest-neighbors Naïve-bayes 14 12 P (Y = spam|Viagara, lottery) Logical models: Tree/Rule based 10 q 8 Decision trees ʻViagraʼ =0 ʻlotteryʼ =0 6 4 2 0 2 4 6 8 10 12 14 3 Ensembles ĉ(x) = ham =1 ĉ(x) = spam =1 ĉ(x) = spam 4 1 12/5/13 Structured Prediction Types of learning tasks Handle prediction problems with complex output spaces Supervised learning Unsupervised learning Semi-supervised learning Access to a lot of unlabeled data Reinforcement learning Learn action to maximize payoff Structured outputs: multivariate, correlated, constrained – Payoff is often delayed Multi-label classification Structured output learning Novel, general way to solve many learning problems Examples taken from Ben Taskar’s 07 NIPS tutorial Handwriting Recognition x Natural Language Parsing y brace Sequential structure x x y The dog chased the cat y S NP VP NP Det N V Det N Recursive structure 2 12/5/13 Bilingual Word Alignment x y What is the anticipated cost of collecting fees under the new proposal? En vertu des nouvelles propositions, quel est le coût prévu de perception des droits? What is the anticipated cost of collecting fees under the new proposal ? Combinatorial structure Local vs. Global En vertu de les nouvelles propositions , quel est le coût prévu de perception de les droits ? Global classification takes advantage of correlations and satisfies the constraints in the problem brace GO term prediction Types of learning tasks Supervised learning Unsupervised learning Semi-supervised learning Access to a lot of unlabeled data Reinforcement learning Learn action to maximize payoff – Payoff Hierarchical structure is often delayed Multi-label classification Structured output learning Multi-task learning 3 12/5/13 Other techniques Probabilistic modeling Bayesian learning 13 4