Recent developments in memory-augmented neural networks allowed sequential problems requiring long-term memory to be solved, which were 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 an external memory component can be encoded geometrically through a novel HyperNEAT-based Neural Turing Machine (HyperNTM). 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 nine bit vectors can be scaled to sizes of 1,000 without further training. While the task in this paper is simple, the HyperNTM approach could now allow memory-augmented neural networks to scale to problems requiring large memory vectors and sequence lengths.
|Titel||International Conference on the Applications of Evolutionary Computation : EvoApplications 2018|
|Status||Udgivet - 2018|
|Navn||Lecture Notes in Computer Science|