HyperENTM: Evolving Scalable Neural Turing Machines through HyperNEAT
This addresses the problem of scaling memory-augmented neural networks for researchers, though it is incremental as it builds on existing Neural Turing Machine and HyperNEAT methods.
The paper tackles the scalability issue of memory-augmented neural networks by introducing HyperENTM, which uses a HyperNEAT-based encoding to train on small memory vectors and then apply that knowledge to larger ones, achieving scaling from size 9 to 1,000 without further training in some cases.
Recent developments within memory-augmented neural networks have solved sequential problems requiring long-term memory, which are intractable for traditional neural networks. However, current approaches still struggle to scale to large memory sizes and sequence lengths. In this paper we show how access to memory can be encoded geometrically through a HyperNEAT-based Neural Turing Machine (HyperENTM). We demonstrate that using the indirect HyperNEAT encoding allows for training on small memory vectors in a bit-vector copy task and then applying the knowledge gained from such training to speed up training on larger size memory vectors. Additionally, we demonstrate that in some instances, networks trained to copy bit-vectors of size 9 can be scaled to sizes of 1,000 without further training. While the task in this paper is simple, these results could open up the problems amendable to networks with external memories to problems with larger memory vectors and theoretically unbounded memory sizes.