Integrating a memory layer in a neural network for one-shot learning

Abstract

A method for machine learning includes extracting features from a training set of inputs, wherein each input generates a feature set and each the feature set forms a neural network key. The method includes arranging the keys in an in-memory computational layer such that the distance between any pair of keys corresponding to similar inputs is as close as possible while keys for a pair of dissimilar inputs have differing values as far apart as possible, wherein each of the keys has a fixed size. The method also includes searching through the dataset using an in-memory K-nearest neighbor unit to find K keys similar to a query key, the searching occurring in a constant amount of time as a function of the fixed size and irrespective of a size of the dataset.