Method for operating neural network
AssigneesTaiwan Semiconductor Manufacturing Company, Ltd., National Tsing Hua University
InventorsKea Tiong TANG, Chen-Han HSU
Life of this patent
Estimating the term…
Abstract
A method is provided and includes operations as below: receiving multiple spike signals in an input layer of a spiking neural network during multiple time steps; counting a corresponding number of spikes in the spike signals for each of the time steps; weighting, in response to the corresponding number of spikes in one of plurality of time steps being greater than a predetermined count value, the spike signals with multiple synaptic weight values to generate multiple synaptic signals; generating a membrane potential by accumulating a number N of the synaptic signals according to a weight distribution of the synaptic weight values; and generating an output spike signal according to the membrane potential.
Claims20
A method, comprising:
- receiving a plurality of spike signals in an input layer of a spiking neural network during a plurality of time steps;
- weighting the plurality of spike signals with a plurality of synaptic weight values to generate a plurality of synaptic signals;
- in a first time step of the plurality of time steps, accumulating a number of signals of the plurality of synaptic signals according to a sequence of a weight distribution of the plurality of synaptic weight values to generate a first intermediate membrane potential;
- comparing the first intermediate membrane potential and a threshold value to generating an output spike having a value “1” in an output spike signal corresponding to the first time step of plurality of time steps;
- and
- when a sum of remaining signals in the plurality of synaptic signals is smaller than a difference between the first intermediate membrane potential and the threshold value, generating the output spike having a value “0” in the output spike signal corresponding to the first time step of plurality of time steps.
A non-transitory computer-readable medium for storing computer-executable instructions, the computer-executable instructions when executed by a processor implementing a method comprising:
- receiving input spikes received in an input layer of a neural network during a first time step;
- weighting the input spikes with a plurality of first synaptic weight values to generate a plurality of first synaptic signals;
- and
- generating a first membrane potential, comprising:
- accumulating signals in the plurality of first synaptic signals within a first number of computing cycles in the first time step to generate a first intermediate membrane potential;
- comparing the first intermediate membrane potential and a threshold value to generate a difference between the first intermediate membrane potential and the threshold value in response to the first intermediate membrane potential being smaller than the threshold value;
- calculating a sum of synaptic weight values corresponding to remaining signals in the plurality of first synaptic signals;
- and
- when the difference between the first intermediate membrane potential and the threshold value is greater than a sum of synaptic weight values corresponding to remaining signals, terminating operations in remaining computing cycles in the first time step.
A method, comprising:
- receiving input spikes received in an input layer of a neural network;
- weighting the input spikes with a plurality of synaptic weight values to generate a plurality of synaptic signals;
- in a first time step, performing a number N of computing cycles of accumulation of the plurality of synaptic signals to generate a first membrane potential;
- in a second time step, performing a number M of computing cycles of accumulation of the plurality of synaptic signals to generate a second membrane potential, wherein M is different from N;
- and
- generating an output spike signal according to the first and second membrane potentials.
Dependencies are inferred from the claim text.