> That gave me the gut feeling that most people doing the research were lacking the necessary mathematical background. I don't find this to be the case with mo

> That gave me the gut feeling that most people doing the research were lacking the necessary mathematical background. I don't find this to be the case with most ML researchers. Is it possible you have misunderstood some of these papers? It is, after all, hard to jump straight into a new field. > The second paper converted float to bool and then tried to use the gradient for training. This sounds like binarized neural networks. If that's the case, they keep the activation before binarization to use for backpropegation. > The third paper only used a 3x3 pixel neighborhood for learning long-distance moves. A single layer of 3x3 convolutions would not be able to model long-distance moves. But I have not read a single paper where they have only used one layer. Is it possible they stacked multiple conv + pooling layers? The receptive field of each unit higher up in the stack grows pretty large in the end.