Download presentation
Presentation is loading. Please wait.
Published byElinor Sharp Modified over 9 years ago
1
Competitive learning College voor cursus Connectionistische modellen M Meeter
2
2 Unsupervised learning To-be-learned patterns not wholly provided by modeller u Hebbian unsupervised learning u Competitive learning
3
3 The basic idea © Rumelhart & Zipser, 1986
4
4 What’s it good for? n discovering structure in the input n discovering categories in the input u Classification networks: ART (Grossberg & Carpenter) CALM (Murre & Phaf) n mapping inputs onto a topographic map u Kohonen maps (Kohonen) u CALM - Maps (Murre & Phaf)
5
5 Features of Competitive learning n Two or more layers (no auto-association) n Competition between output nodes n Two phases: u determining a winner u learning n Weight normalisation
6
6 Two or more layers Input must come from outside the inhibitory clusters © Rumelhart & Zipser, 1986
7
7 Competition between output nodes n At every presentation of an input pattern, a winner is determined n Only winner is activated [activation at learning discrete: (0,1) ] u Hard Winner Take All: Find node with maximum input max. ( w ij a j ) u Inhibition between nodes
8
8 Inhibition between nodes n Example: inhibition in CALM
9
9 Two phases 1.One node wins the competition 2.That node learns, others not n Nodes start off with random weights n No ‘correct’ output connected with inputs: unsupervised learning
10
10 Weight normalisation n Weights of winner node i changed w ij = * a j n Weights add up to constant sum... w ij = 1 rule of Rumelhart & Zipser: w ij = g * a i / n k - g * w ij n …or constant distance: (w ij ) 2 = 1
11
11 Geometric interpretation n Both weights & input patterns can be seen as vectors in a hyper space n Euclidian normalisation [ (w ij ) 2 = 1] u all vectors on a sphere in space of n dimensions (n = number of inputs) u node with weight vector closest to input vector is winner n Linear normalisation [ w ij = 1] u all weights on a plane
12
12 Geometric interpretation II n Weight vectors move towards input in the hyper space w ij = g * a i /n k - g * w ij n Output nodes move towards clusters in inputs © Rumelhart & Zipser, 1986
13
13 Stable / unstable n Output nodes move towards clusters in inputs n If input not clustered......output nodes will continue moving through input space! © Rumelhart & Zipser, 1986
14
14 Statistical equivalents n Sarle (1994): Classification = k-means clustering Kohonen = mapping continuous dimensions onto discrete ones u Statistical techniques usually more efficient... u...because statistical techniques use whole data set
15
15 Importance of competitive learning n Supervised - unsupervised learning n Structure input sets not always given n Natural categories
16
16 Competitive learning in the brain n Lateral inhibition feature of most parts of the brain … Implements winner-take-all ?
17
17 Part II
18
18 Map formation in the brain n Topographic maps omnipresent in the sensory regions of the brain u retinotopic maps: neurons ordered as the locations of their visual field on the retina u tonotopic maps: neurons ordered according to tone for which they are sensitive u maps in somatosensory cortex: neurons ordered according to body part for which they are sensitive u maps in motor cortex: neurons ordered according to muscles they control
19
19 Somatosensory maps © Kandel, Schwartz & Jessell, 1991
20
20 Somatosensory maps II © Kandel, Schwartz & Jessell, 1991
21
21 Speculations n Map formation ubiquitous (also semantic maps?) n How do maps form? u gradients in neurotransmitters u pruning
22
22 Kohonen maps n Teuvo Kohonen first to show how maps can develop n Self Organising Maps (S.O.M.) n Demonstration: the ordering of colours (colours are vectors in a 3-dimensional space of brightness, hue, saturation).
23
23 Kohonen algorithm n Finding the activity bubble n Updating the weights for the nodes in the active bubble
24
24 Finding the activity bubble Lateral inhibition
25
25 Finding activity bubble II n Find the winner n Activate all nodes in the neighbourhood of the winner
26
26 Updating the weights n Move weight vector of winner towards the input vector n Do the same for the active neighbourhood nodes weight vectors of neigbouring nodes will start resembling each other
27
27 Simplest implementation n Weight vectors & input patterns all have length 1 (e.i., (w ij ) 2 = 1 ) n Find node whose weight vector has mimimal distance to the input vector: min. (a j - w ij ) 2 n Activate all nodes in neighbourhood radius N t n Update weights of active nodes by moving weights towards the input vector: w ij = t * ( a j - w ij ) w ij (t+1) = w ij (t) + t * ( a j - w ij (t) )
28
28 Results of Kohonen © Kohonen, 1982
29
29 Influence of neighbourhood radius © Kohonen, 1982 Larger neighbourhood size leads to faster learning
30
30 Results II: the phonological typewriter © Kohonen, 1988
31
31 Phonological typewriter II © Kohonen, 1988
32
32 Kohonen conclusions n Darn elegant n Pruning? n Speech recognition uses Hidden Markov Models
33
33 Summary n Prime example of unsupervised learning n Two phases: u winner node is determined u weights are updated of the winner only n Very good at discovering structure: u discovering categories u mapping the input onto a topographic map n Competitive learning important paradigm in connectionism
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.