Machine Learning 12. Local Models
Introduction Clustering Use neural network for unsupervised learning Using K-means algorithm Iterative algorithm Batch learning Use neural network for unsupervised learning Online learning Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Online K-means Reconstruction Error For batch k-means, center update is Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Online k-means Reconstruction error for single instance Using gradient descent Move closest center to the direction of new sample Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Competitive Learning Based on 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Competitive Learning Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Using Dot product If All centers have the same norm Minimum Euclidian distance Correspond to maximum dot product Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Choosing maximum using NN Using Recurrent Networks Lateral inhibition Positive(excitatory) recurrent connection to itself Negative(inhibitory) recurrent connection to its neighbors Suitable weights and activation function converge to a maximum Singe output is 1 All other 0 Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
NN network Winner-take-all network Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Hebbian learning Update First term: increase weight if input and output are activated together (correlated) Second term: prevent unbounded weight growth Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Adaptive Resonance Theory Incremental; add a new cluster if not covered; vigilance, ρ (Carpenter and Grossberg, 1988) Based on Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Self-Organizing Maps Units have a neighborhood defined; mi is “between” mi-1 and mi+1, and are all updated together One-dim map: (Kohonen, 1990) Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Locally Receptive Units Divide the input space into local regions and learn simple (e.g. constant/linear) models in each patch Radial-basis func, mixture of experts Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Local vs Distributed Representation Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Radial-Basis Functions Locally-tuned units: Based on E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Training RBF Hybrid learning: Fully supervised RBF is differentiable First layer centers and spreads: Unsupervised k-means Second layer weights: Supervised gradient-descent Fully supervised RBF is differentiable Back propagation Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)
Rule-Based Knowledge Incorporation of prior knowledge (before training) Rule extraction (after training) (Tresp et al., 1997) Lecture Notes for E Alpaydın 2004 Introduction to Machine Learning © The MIT Press (V1.1)