Ok so in case you havent read it I uploaded a post about what I call Seed Recall here.
Basically you train a Neural Network on, in this case, a reinforcement learning task and it quickly finds a solution. Now if you create a random seed that is used to change the activation functions in that neural network then you can recall a solution found unique to that seed.
Ie. I put seed 0.52 into the NN, train it on the problem. Reload the weights and the seed and I get the same solution I trained for. And you might say - so what?
Well the beauty of it is if I provide another seed without retraining I get another unique solution based around the form of the first solution - a varient of that solution if you like. This is the essense of 'Seed Recall'
Now here is the good bit. It turns out there are actually three seeds here that are useful. One is the Training seed which you can choose before you train and will always train to the same result guided by the seed.
The second is the Recall seed. This is the seed that is also set at training time and is then used to recall the solution.
The third is the Discovery seed this is set after Training and Recall is done and you want to explore the solution space around the Recall solution.
Three seeds to rule them all :
1. Training Seed
2. Recall Seed
3. Discovery Seed
Stay tuned for a github page for this one. Email me for more info.
No comments:
Post a Comment