Lecture note
the reason why object detection with edge detecting techinques is hard.
- brittle
- with other objecsts, we need to start all over again
SO, Data Driven Approach is needed.
Homework - CIFAR10
KNN for image classification? NO.
training with nearest neighbor can be result in testing process is O(1), and predicting process is O(n). This should be sometwhat backwards.
When using knn, we need to choose which distance metric should be used, how many ks should be needed. these are the hyperparameters and it's about choosing algorithm that we set rather than learn.
setting hyperparameter
is spliting data into test/train just enough? it may not.
test / validation (do all debugging, choose hyperparameter..) / test
knn is never used for image classifier.
- very slow at test time
- distance metrics on pixels are not informative
Linear Classification?
simple y = Wx + b. Can't solve XOR problem.
'데이터분석 > Deep Learning' 카테고리의 다른 글
| dnn sample (0) | 2018.10.09 |
|---|---|
| [Link] 강화학습 관련 링크 (0) | 2017.12.03 |
| Deep Learning 개발 라이브러리 장단점 (0) | 2017.11.14 |
| 평균제곱오차와 교차엔트로피 분류오차 차이 (0) | 2017.11.10 |
| 텐서플로우 입문기 (4) - 딥러닝(CNN)을 통한 손글씨 인식. (0) | 2016.12.24 |