Does Progressive Neural Network works?

[facetohard]

I have read the code and I found that when apply gradients, it apply all variable in the share networks which is written in line 343 of "ProgressiveNN_atari_2.py"
self.apply_gradients = master.shared_opt.apply_gradients( zip(self.accum_grads, master.shared_net.get_vars()), global_step=master.global_step)
The paper says that when update the weights, it should frozen the weights of previous learned policies, but I can't find which step frozens the previous weights.

[GuangpingYuan]

I have the same doubt, do you have any answer yet?

[GuangpingYuan]

@yao62995 I have the same doubt, do you have any answer yet?