1 Midterm Review Machine Learning 10-601 Tom M. Mitchell Machine Learning Department Carnegie Mellon University October 25, 2011 See practice exams on our website Midterm is in class October 27 Midterm is open book, open notes, NO computers, NO internet Covers all material presented up through todayʼs class. Some Topics We’ve Covered Decision trees entropy, mutual info., overfitting Probability basics Bayes rule, MLE, MAP, conditional indep. Naïve Bayes conditional independence, # of parameters to estimate, decision surface Logistic regression form of P(Y|X) generative vs. discriminative Linear Regression minimizing sum sq. error (why?) regularization ~ MAP Sources of Error unavoidable error, bias, variance Overfitting, and Avoiding it priors over H cross validation PAC theory: probabilistic bound on overfitting Bayesian Networks factored representation of joint distribution, conditional independence assumptions, D-separation inference in Bayes nets learning from fully/partly observed data PAC Learning sample complexity probabilistic bounds on errortrain – errortrue VC dimension2 Understanding/Comparing Learning Methods Form of learned model • Inputs: • Outputs: Optimization Objective: Algorithm: Assumptions: Guarantees?: Decision boundary: Generative/Discriminative? Naïve Bayes Form of learned model • Inputs: • Outputs: Optimization Objective: Algorithm: Assumptions: Guarantees?: Decision boundary: Generative/Discriminative? Logistic Regression3 Four Fundamentals for ML 1. Learning is an optimization problem – many algorithms are best understood as optimization algs – what objective do they optimize, and how? Local minima? – gradient descent/ascent as general fallback approach Four Fundamentals for ML 1. Learning is an optimization problem – many algorithms are best understood as optimization algs – what objective do they optimize, and how? 2. Learning is a parameter estimation problem – the more training data, the more accurate the estimates – MLE, MAP, M(Conditional)LE, … – to measure accuracy of learned model, we must use test (not train) data4 Four Fundamentals for ML 1. Learning is an optimization problem – many algorithms are best understood as optimization algs – what objective do they optimize, and how? 2. Learning is a parameter estimation problem – the more training data, the more accurate the estimates – MLE, MAP, M(Conditional)LE, … – to measure accuracy of learned model, we must use test (not train) data 3. Error arises from three sources – unavoidable error, bias, variance – PAC learning theory: probabilistic bound on overfitting: errortrue - errortrain given some estimator Y for some parameter θ, we note Y is a random variable (why?) the bias of estimator Y : the variance of estimator Y : consider when • θ is the probability of “heads” for my coin • Y = proportion of heads observed from 3 flips consider when • θ is the vector of correct parameters for learner • Y = parameters output by learning algorithm Bias and Variance of Estimators5 Four Fundamentals for ML 1. Learning is an optimization problem – many algorithms are best understood as optimization algs – what objective do they optimize, and how? 2. Learning is a parameter estimation problem – the more training data, the more accurate the estimates – MLE, MAP, M(Conditional)LE, … – to measure accuracy of learned model, we must use test (not train) data 3. Error arises from three sources – unavoidable error, bias, variance – PAC learning theory: probabilistic bound on overfitting: errortrue - errortrain 4. Practical learning requires making assumptions – Why? – form of the f:X à Y, or P(Y|X) to be learned – priors on parameters: MAP, regularization – Conditional independence: Naive Bayes, Bayes nets,
View Full Document