Machine Learning and Adaptive Systems. Lectures 3 & 4

Transcription

1 ECE656- Lectures 3 & 4, Professor Department of Electrical and Computer Engineering Colorado State University Fall 2015

2 What is Learning? General Definition of Learning: Any change in the behavior or performance brought about by the experience or contact with the environment Learning in humans is an inferred process, ie we cannot see it happening directly but observe the resultant changes in behavior Learning in ANN is a more direct process and we can capture each learning step in a cause-effect relationship Designing a machine learning algorithm is based upon learning (supervised) a relationship that transforms inputs to outputs from a set of input-output examples drawn from Experience E Given a training set consisting of pairs of inputs-outputs {, d(k)} K k=1 from E : k th training sample (pattern) vector of dimension N 1 d(k) : desired output (response) for input pattern of dimension M 1 where typically M N

3 The problem is viewed as a task T of approximating a continuous multi-variate mapping function d(k) = f() by a function h(w, ) as closely as possible (measured by some performance metric ρ() wrt E) with w being the parameter vector f() d(k) : Desired ANN h() - o(k) : Actual If we define a performance measure ρ(h(w, ), f()), then for the optimal parameter vector w, ρ(h(w, ), f())) ρ(h(w, ), f())), k [1, K] where w is any other vector Note: Such learning rules are referred to as Performance Learning Examples of ρ() are MSE, classification error, mutual information, etc

4 1 Supervised Learning Requires an external teacher Teacher provides training sample pairs ie and d(k) Training is based upon error learning Training is typically iterative More accurate than unsupervised ones Used for detection/classification, prediction/regression, and function approximation Environment ANN h() Adjustment Mechanism o(k) () d(k) Teacher Issues 1 Suffer from speed-accuracy tradeoff, ie high speed of convergence less accuracy in parameter estimation 2 Incremental learning requires old data 3 Overtraining leads to performance degradation

5 2 Unsupervised Learning Uses unlabeled data ie training samples have no desired outcomes No external teacher ie based upon self-organization Form clusters to discover interesting features from data Clusters are formed based upon underlying statistical properties of the data (unconditional density estimator vs conditional for supervised) Simpler algorithms than those of supervised Used for data clustering, compression, and dimensionality reduction Environment ANN h() o(k) Adjustment Mechanism Issues 1 Requires more parameter tuning 2 No automatic shut-off system 3 Less accurate than supervised counterparts when used for decision-making

6 3 Reinforcement Learning Originated in psychology with experiments on animal learning No unambiguous outcome as in supervised learning ie no explicit supervision Given training set {, r(k)} K k=1, where r(k) = { 1, +1} is supplied by a critic (not desired signal) as a reinforcement (or appropriateness) signal Reward and Punishment-based actions designed to maximize the expected value of a criterion function Learning should devise a sequence of actions over time to obtain large Cumulative Reward Applied in web-indexing, cell-phone network routing, control theory (eg, robotics, unmanned vehicles), marketing strategy selection, etc Environment Action Primary Reinforcement signal Critic ANN h() Adjustment Mechanism Heuristic Reinforcement signal o(k) In machine learning, environment is typically formulated as a Markov decision process (MDP) = dynamic programming techniques It provides feedback to the environment in forms of actions (ie allows for interaction)

7 Generalized Learning Rule (Amari 1990) Weight vector w i is updated in proportion to the input and a learning signal r(w i (k),, d i (k)) Thus, increment of w i (k) at iteration k is w i (k) = µ r(w i (k),, d i (k)) where µ is learning rate (or step size) Then, the weight vector at iteration k + 1 is adjusted using w i (k + 1) = w i (k) + w i (k) w i (0) s are randomly initialized wi1 wij win Cell i o i(k) Δwi win Adjustment di(k) Mechanism Teacher μ : step size This provides a unified representation for several learning rules that we shall cover In the next few lectures we cover, (a) Performance Learning - used for classification, function approximation (b) Coincidence Learning - used for association (memory learning) (c) Competitive Learning - used for clustering (d) Reinforcement Learning used in controls, game theory, multi-agent systems and their applications in real-world problems

8 a Performance Learning-ADALINE (Widrow-Hopf 1960) These learning methods are based upon minimizing or maximizing a performance measure (cost function) ADALINE (ADAptive LINear Element) is based upon McCulloch-Pitts model with bipolar HL activation w i1 w ij Cell i net i(k) HL o i(k) w in w in Adjustment Mechanism e i(k) - d i(k) + Objective: Find w i such that MSE between net i and the desired (supervised) is minimized That is, given {, d i (k)} K k=1, find w i to minimize, J(w i ) = 1 K (d i (k) net i (k)) 2 = 1 K e 2 i (k) : Time-Averaged SE (1) K K k=1 k=1 where net i (k) = w t i and d i(k) = net i (k) + e i (k)

9 For an ergodic process the time-averaged MSE can be rewritten in terms of ensemble average MSE, ie J(w i ) = E[(d i (k) w t i )2 ] = E[d 2 i (k)] + wt i E[xt (k)] w i 2w t i E[d i(k)] = σ 2 d + wt i Rxx w i 2w t i R xd where E[]: Expectation Operation, σ 2 d = E[d2 i (k)]: variance of d i(k), R xx = E[x t (k)]: correlation matrix of data; and R xd = E[d i ]: cross-correlation vector between input data and desired signal As can be seen J(w i ) is quadratic and has a bowl-shaped surface as a function of w i which is also referred to as Error Performance Surface J(wi) Error Surface This surface has a unique global minimum at which, J(w i ) w i = 0 2R xxw i 2R xd = 0 Solving for w i gives, wi1 wi2 Absolute Minimum Jmin w i = R 1 xx R xd which is the Wiener-Hopf solution (or Minimum Mean Squared Error solution (MMSE))

10 Example: Use an ADALINE to estimate weights of a 2nd order AR based linear predictor = a 1 x(k 1) + a 2 x(k 2) + u(k), u(k) N (0, σ 2 u) Here, net(k) = w(k) t, where = [x(k 1) x(k 2)] t and w i (k) = [a 1 (k) a 2 (k)] t Also, we choose d(k) = Solving for w yields a 1 and a 2 Remarks z -1 z -1 w1(k) w2(k) net(k) 1 Using the Wiener-Hopf solution the minimum MSE is J min (w i ) = σ2 d w t i R xd 2 For stationary processes the error surface has a fixed shape and orientation; whereas for non-stationary processes the bottom of the error surface moves continually and orientation and curvature may be changing too Thus, the - solution should not only seek the bottom but also track it 3 Wiener-Hopf solution requires computing R xx, R 1 xx, and R xd which makes it unsuitable for online learning 4 For ergodic processes, solving the original time-averaged SE, which corresponds to Least Squares (LS) solution, approaches the Wiener-Hopf (ie MMSE) solution for large K This is shown next + e(k)

11 Least Squares Solution & Geometric Interpretation Let us reconsider the time-averaged SE in (1) Now, if we define d i = [d i (1),, d i (K)] t : Desired vector e i = [e i (1),, e i (K)] t : Error vector x t (1) X = : Data matrix x t (K) K N Then we can write d i = Xw i + e i = net i + e i Now, the time-averaged SE is J(w i ) = 1 K e2 i = 1 K (d i Xw i )t (d i Xw i ) Setting J(w i ) = 0 yields w i 2X t (d i Xŵ ils ) = 0 = X t Xŵ ils = X t d i which gives the LS solution, ŵ ils = (X t X) 1 X t d i = X d i where X = (X t X) 1 X t is the pseudo inverse of X

12 Recall from LS solution ŵ ils = (X t X) 1 X t d i = x t (1) X t X = [x(1) x(k)] x t (K) Matrix X t d i = [x(1) x(k)] Cross-correlation Vector Clearly, d i (1) d i (K) R xx, R xd : Time-averaged R xx, R xd : Ensemble averaged R 1 xx R xd where = K k=1 t = R xx: Sample Correlation = K k=1 d i(k) = R xd : Sample If data is stationary and ergodic, the LS solution asymptotically approaches the Wiener-Hopf solution for large K, ie lim K Rxx R xx lim K Rxd R xd then ŵ ils w i MMSE

13 Now, define Orthogonal Projection Matrix, P X, P X = X(X t X) 1 X t which is idempotent ie P 2 X = P X or P n X = P X, n and Symmetrical ie P t x = P X Then, ˆ net i = Xŵ ils = P X d i ie projection of d i onto a space spanned by columns of X Using this result the error vector at the LS solution is, ê i = d i ˆ net i = d i P X d i = P X d i where P X = I P X is the orthogonal complement of P X Note that e i = ê i + ( ˆ net i net i ) The geometrical representation of these is shown in the figure d i e i e i PX d i nˆ et i P X d i net i nˆ et i neti

14 Homework 1-Due September 15th, 2015 Problems: 1 Using the Wiener-Hopf solution show that the estimation error e o(k) = d(k) w t is orthogonal to the data, ie E[e o(k)] = E[(d(k) w t )] = 0 Explain the importance of this property 2 Is the same orthogonality property valid for the LS solution? If it does, drive it Compare the results in Problems 1 and 2 and their interpretations 3 To guarantee uniqueness and improve stability of LS solution, typically a regularization term is added to the cost function, ie J(w i ) = (d i Xw i ) t (d i Xw i ) + λ w i 2 Find the solution of this regularized (or constrained) LS and investigate its relation to the LS solution 4 Solve Problem 23 in your textbook

15 b Performance Learning-Steepest Descent An alternative procedure using steepest (or gradient) descent method can be devised to find the weights in ADALINE without requiring to find Rxx 1 This method involves following steps 1 Start with an initial guess for w i (0) 2 Use this guess to compute the gradient vector J(w i (k)) J(w i (k)) w i (k) at iteration (training sample) k 3 Update weight in a direction opposite to the gradient w i (k + 1) = w i (k) 1 2 µ J(w i (k)) where µ: Step Size 4 Go back to step (2) and repeat until convergence conditions met From the previous result J(w i (k)) = 2R xd + 2R xxw i (k) Thus, w i (k + 1) = w i (k) + µ[r xd R xxw i (k)] ie no matrix inversion is required, but still needs to compute R xx and R xd Step size µ controls the size of incremental corrections For convergence, 0 < µ < 2 λ max, where λ max: largest eigenvalue of R xx