Efficient regularization methods for classification - Githubissues

accosmin / nano

C++ library [machine learning & numerical optimization] - superseeded by libnano

MIT License

1 stars 0 forks source link

Efficient regularization methods for classification #130

Closed accosmin closed 7 years ago

accosmin commented 8 years ago

Regularizing deep models takes too much time because several regularization factors have to be tried out. So remove the criterion_t part.

The idea would be to regularize in one shot:

generate random samples with label {-1}+ (aka background images) and check if adding them to the training dataset improves performance
randomly warp training images to extend the training set

Update benchmark_iterator:

benchmark the impact of using different samplers (e.g. none, warp, nrand)