Hybrid Computing using a Neural Network with Dynamic External Memory

Hybrid computing using a neural network with dynamic external memory

Paper is an extension of NTM architecture
External memory block is used to surpass short term memory issue in deep neural networks
Three different attention mechanisms are defined for reading/writing/replacing in external memory

Proposed method of using external differentiable memory block is inspired from human memory theory mainly focusing on attribute theory, similarity measures, search of associative memory, temporal context model and serial recall
Proposed soft attention mechanisms, content-lookup for reading from the memory block, attention of temporal links for replacing memory and attention for writing in memory matrix
Method is an end to end differentiable extension of Neural Turing Machine

Entire architecture is divided into 4 blocks, controller (neural network), read-write heads, memory block, and temporal weightage block
Controller Block: mapping function
Read-Write Block: variable weight vectors which encodes the position in memory block where to write or read from memory
Memory Block: Differentiable block which stores past signals, which will be used for further learning
Temporal Block: Differentiable weight block which encodes the similarity between rows of memory block
The output of the controller is not only governed by input but also on one/many feature vectors in memory block

The experiments were conducted on graph problems on family trees and london underground, along with copying task
The results outperforms pre-existing LSTM, NTM techniques
The experiments on reinforcement learning shows the promising results in extending the research direction

The method provides strong direction for memory augmented methods which can be used for multi-domain association problems
They hold the state information for a long amount of time, providing sufficient time for self correction
Some ideas for multi-agent, multi-domain adaptations can be drawn from this paper
If interested in exploring any of the above ideas, message me: koriavinash1@gmail.com