International Business Machines Corporation

Lorenz K Muller

Pascal Stark

Stefan Abel

Bert Jan Offrein

Folkert Horst

Roger F Dangel

A method of processing data and related apparatuses. The method relies on an optical finite impulse response (FIR) filter. This optical FIR filter comprises several delay stages having weights set in accordance with parameters of a transformation to be applied by the optical FIR filter. Each of the delay stages is configured to impose a delay matched to a given input data period corresponding to a given input sample rate. According to the method, an optical signal is coupled into the optical FIR filter. The optical signal carries a data stream of input samples encoded at the given input sample rate; the data stream represents the data to be processed. Next, output samples are collected from an output data stream carried by an output optical signal obtained in output of the optical FIR filter. A set of output samples are obtained, which are representative of processed data.

Convolutions with optical finite impulse response filters

A method of operating a neural network. The input layer of the network may have n input nodes connected to output nodes via a hidden layer. The hidden layer may include m hidden nodes. The n input nodes may connect to a subset of k nodes of the m hidden nodes via respective synaptic connections, to which training weights are associated, which form an n×k input matrix W, whereas a subset of m−k nodes of the hidden layer are not connected by any node of the input layer. Running the network may include performing a first matrix vector multiplication between the input matrix Wand a vector of values obtained in output of the input nodes and a second matrix vector multiplication between a fixed matrix Wof fixed weights and a vector of values obtained in output of the m nodes of the hidden layer.

Fixed, random, recurrent matrices for increased dimensionality in neural networks

A method and computer program product for obtaining values are run using a neural network according to a machine learning algorithm. One embodiment may comprise accessing one or more datafiles of input data, where the input data is representable in a d-dimensional space, with d>1. The method may explore N distinct paths of the input data in the d-dimensional space, where N≥1, and collects data along the N distinct paths explored to respectively form N sequences of M objects each, with M≥2. For one or more sequences of the N sequences formed, values obtained from the M objects of each sequence may be coupled into one or more input nodes of a neural network, which is then run according to the machine learning algorithm to obtain L output values from, L≥1.

Growing community of inventors

Dubendorf, Switzerland

Lorenz K Muller