Seong-lae Chocase based learning (reasoning)
Kernel trick: we can substitute any*similarity function in place of the dot product
classification by similarity
image 0 < i < 1 vector dot product is similar function
similar by distance
knn - nonditermine, local answer
Trade-offs: Small k gives relevant neighbors(오버피팅), Large k gives smoother functions (적당한 피팅)
- k random center pick
- 가까운 center에이 거기 클러스터 - initial membership
- 3. 거기의 중간으로 이동 (좌표값평균)
- 다시 가꾸은 center에 속하게
- 반복
stop no change
Agglomerative clustering
모든 거 사이에 간격 있다
간격 제일 작은거부터 합치는데
cluster도 점으로 처리
Parametric models - fixed seet of parameter non parametric - often limit - classifier increasees with data
kernelization
weight vectors (the primal representation) from update counts (the dual representation)