Neuron unit, neural network and signal processing method

A signal processing method for processing a plurality of input signals in a neuron unit and for outputting an output signal which is indicative of a result of the processing, said signal processing method comprising the steps of:

• (a) carrying out a forward process including:
• 1) carrying out an operation on each of the input signals using one of weight functions stored in the memory means;
• and
• 2) outputting an operation result as the output signal;
• and
• (b) carrying out a self-learning process including:
• 1) generating new weight functions based on an error signal describing an error between the output signal obtained by the forward process and a teaching signal;
• and
• 2) varying the weight functions used by the forward process to the new weight functions which are generated by renewing contents of the memory means;
• wherein at least the weight function out of a group including the input signals, the weight functions, the output signal, the error signal and the teaching signal is described by a stochastically encoded pulse train.

A signal processing method for processing a plurality of input signals in a neuron unit and for outputting an output signal which is indicative of a result of the processing, said signal processing method comprising the steps of:

• (a) storing input signal data of the input signals in first memory means and storing teaching signal data of a teaching signal in second memory means;
• (b) carrying out a forward process including:
• 1) carrying out an operation on each of the input signals using one of weight functions stored in the memory means;
• and
• 2) outputting an operation result as the output signal;
• and
• (c) carrying out a self-learning process including:
• 1) generating new weight functions based on an error signal describing an error between the output signal obtained by the forward process and the teaching signal;
• and
• 2) varying the weight functions used by the forward process to the new weight functions which are generated;
• wherein said error signal includes:
• 1) a first error signal component which is calculated from a logical product of the output signal output by said step (b) and a logical NOT of the teaching signal;
• and
• 2) a second error signal component which is calculated from a logical product of a logical NOT of the output signal output by said step (b) and the teaching signal.

A signal processing method for processing a plurality of input signals in a neuron unit and for outputting an output signal which is indicative of a result of the processing, said signal processing method comprising the steps of:

• (a) storing input signal data of the input signals in first memory means and storing teaching signal data of a teaching signal in second memory means;
• (b) carrying out a forward process including:
• 1) carrying out an operation on each of the input signals using one of weight functions stored in the memory means;
• and
• 2) outputting an operation result as the output signal;
• and
• (c) carrying out a self-learning process including:
• 1) generating new weight functions based on an error signal describing an error between the output signal obtained by the forward process and the teaching signal;
• and
• 2) varying the weight functions used by the forward process to the new weight functions which are generated;
• wherein at least the weight function out of a group including the input signals, the weight functions, the output signal, the error signal and the teaching signal is described by a stochastically encoded pulse train.

A computer-implemented method of simulating a neuron unit for processing a plurality of input signals and for outputting an output signal which is indicative of a result of the processing, said computer-implemented method comprising the steps of:

• (a) receiving the input signals;
• (b) carrying out a forward process including:
• 1) supplying weight functions and carrying out an operation on each of the input signals using one or the weight functions;
• and
• 2) for outputting an operation result as the output signal;
• and
• (c) carrying out a self-learning process including:
• 1) generating new weight functions based on an error signal which describes an error between the output signal and a teaching signal;
• and
• 2) varying the weight functions to the new weight functions which are generated;
• wherein said error signal includes:
• 1) a first error signal component which is calculated from a logical product of the output signal output by said step (b) and a logical NOT of the teaching signal;
• and
• 2) a second error signal component which is calculated from a logical product of a logical NOT of the output signal output by said step (b) and the teaching signal.

A computer-implemented method of simulating a neuron unit for processing a plurality of input signals and for outputting an output signal which is indicative of a result of the processing, said computer-implemented method comprising the steps of:

• (a) receiving the input signals;
• (b) carrying out a forward process including:
• 1) supplying weight functions and carrying out an operation on each of the input signals using one or the weight functions;
• and
• 2) outputting an operation result as the output signal;
• and
• (c) carrying out a self-learning process including:
• 1) generating new weight functions based on an error signal which describes an error between the output signal and a teaching signal;
• and
• 2) varying the weight functions to the new weight functions which are generated;
• wherein at least the weight function out of a group including the input signals, the weight functions, the output signal, the error signal and the teaching signal is described by a stochastically encoded pulse train.