Digitial STDP synapse and LIF neuron-based neuromorphic system
AssigneeInternational Business Machines Corporation
InventorsTakeo Yasuda, Kohji Hosokawa, Yutaka Nakamura, Junka Okazawa, Masatoshi Ishii
Life of this patent
Estimating the term…
Abstract
Described is a neuromorphic system implemented in hardware that implements neuron membrane potential update based on the leaky integrate and fire (LIF) model. The system further models synapse weights update based on the spike time-dependent plasticity (STDP) model. The system includes an artificial neural network in which the update scheme of neuron membrane potential and synapse weight are effectively defined and implemented.
Claims19
A neuromorphic system for generating neuron activity, comprising, in a hardware implementation:
- an axon spike timer to determine the elapsed time between occurrences of axon spikes in synapses between neurons;
- a dendrite spike timer to determine the elapsed time between occurrences of dendrite spikes in synapses between neurons;
- a neuron membrane potential register where current neuron membrane potential values are stored, operatively coupled to a neuron membrane potential updating logic;
- a memory unit that stores current values of synapse weights;
- the neuron membrane potential updating logic that determines neuron membrane potential values based on a leaky integrate and fire model of neuron membrane potential behavior, operatively coupled to the axon timer, the memory unit, the neuron membrane potential register and dendrite output register;
- and
- a dendrite output register that converts serial neuron fire data signals to parallel neuron fire data signals.
A neuromorphic system for generating neuron activity, comprising, in a hardware implementation:
- a neuron and synapse network that generates neuron activity and synapse activity between neurons;
- an axon spike timer that measures the elapsed time between occurrences of axon spikes in synapses between neurons;
- a dendrite spike timer that measures the elapsed time between occurrences of dendrite spikes in synapses between neurons;
- a synapse weight memory unit that stores the current values of synapse weight;
- a neuron membrane potential register that stores the current neuron membrane potential values of neurons;
- a neuron membrane potential updating logic that updates the neuron membrane potential values and sends neuron fire signals to a dendrite output register, the updating based on a leaky integrate and fire model of neuron membrane potential behavior;
- and
- the dendrite output register that converts serial neuron fire data received from the neuron membrane potential updating logic to parallel neuron fire data.
A method for simulating neuron activity with spike time-dependent plasticity modeling and leaky integrate and fire modeling comprising, implementing in hardware:
- generating neuron activity and synapse activity between the axons of pre-neurons and neuron bodies of post-neurons in a wired network;
- timing the elapsed time between occurrences of axon spikes in synapses between neurons in a timer unit;
- timing the elapsed time between occurrences of dendrite spikes in synapses between neurons in a timer unit;
- storing current neuron membrane potentials of neurons in a storage block;
- updating neuron membrane potential values in a logic block that performs the update based on leaky integrate and fire model of neuron membrane potential decay;
- and
- updating synapse weight values of the synapses in a logic block that performs the update determination based on spike time-dependent plasticity rule.
Dependencies are inferred from the claim text.