31 17th March, Saturday

This week I've been working on setting up a system to actually stop memoization.

The system I have designed takes 10K of the 60K of MNIST training images as a seed. It is then fed the remaining 50K images, along with the same 50K images but with random labels.

My results have show that the distribution network I have trained accepts 99.0% of the original 3125 batches and accepts 0.8% of the batches with images with random labels.

The only issue currently that I have is that I require the original and random data to be put into batches of 16, which can be inconvenient.