Abstract by Tyler Etchart
Auxiliary Memory Models for Deep Neural Networks
3M. There has been strong premilinary work in the field of auxiliary memory models for deep neural networks. Examples include the Neural Turing Machine and the Differential Neural Computer by Google Deepmind. While these models have proven successful, they are fundamentally limited by their need for differentiable memory structures. We present a new method for using neural networks with non-differentiable memory structures by removing the non-differentiable operation from the computation graph. This is significant because it frees us to utilize all kinds of memory structures, such as arrays, maps, and more.