Local Clustering Algorithm DISCOVIR
Image collection within a client is modeled as a single cluster. Current Situation
Proposed Improvement Multiple clusters exist in the image collection
Group of similar local cluster A
Group of similar local cluster B
Group of similar local cluster C
Clustering Algorithm 3 clustering algorithms are proposed and tested C – set of cluster center X – dataset Goal of clustering : minimize Error
Procedure Randomly pick k x j from {X} and assign them as the set {C} as initial cluster center. Find closest cluster center c and update Error change < threshold iterate a certain step? Input dataset and cluster Randomly pick a point from dataset X N Return cluster center Y
Shifting Mean (SM) Suppose x j is picked and c i is the closest cluster center Let p be number of times c i wins, initially p=1 Update c i by
Competitive Learning (CL) Update c i by t – the current number of iteration so far T – total number of iteration intend to run We choose by 0.5, 0.3, 0.1
Illustration cici xjxj
cici xjxj Winner (move closer)
Rival Penalized Competitive Learning (RPCL) Suppose c l is the second closest cluster center to x j Update c i by Update c l by We choose = 0.05
Illustration cici xjxj Winner (move closer) Rival (move away) unchanged
Final Steps For each x j,find the closest c i and mark x j belongs to c i Calculate error function Carryout experiments by varying # of iteration, learning rate
Results 0.3Error (400)Error (1000)Error (5000) SM CL RPCL Error (0.5)Error (0.3)Error (0.1) SM CL RPCL Fixed Iteration Varying learning rate Fixed Learning rate Varying iteration 0.1Error (400)Error (1000)Error (5000) SM CL RPCL Fixed Learning rate Varying iteration
Screen Capture
Others Other variation i – initial learning rate f – final learning rate Interesting link for competitive learning some competitive learning methods some competitive learning methods