thushv89 / convnets Goto Github PK

Instead of removing filters we will replace them
Because removing seems to cause trouble. Accuracy is not going up.

The accuracy didn't go above 10% even after 6000 iterations

Changing function phi
Changing to make the first epoch fully experimental
Remove distMSE

Add operation
Optimize the added filters + all other layers
Replace operation
Optimizer the replaced filters + all other layers
Finetune
Optimize the full network

When adding feature maps, we should finetune the just the added maps with the pool that contain hard examples.
An easy way of doing that can be masking the already existing feature maps and optimizing only that layer using Grad Descent.

This gives more importance to layer depth when regressing q values

Caused by passing unreshaped 1 sample arrrays

When removed accuracy doesn't suddenly drop.
Accuracy seem to be increasing consistently

Incremental, Non-Stationary, Cifar-10
Incremental, Stationary, Cifar-10
NonIncremental, Non-Stationary, Cifar-10
NonIncremental, Stationary, Cifar-10

Embracing Deep Q learning techniques to the Q learner

Experience Replay, to avoid input correlations
Instead of a single state s1 as input, a collection of (s1,a1,s2,a2,s3,a3)
Introduce Target Network (Simply save the previous model to a variable)