Neuromorphic Computing Platform for Sparse Coding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current deep machine learning approaches face limitations in generalizing well to novel situations with limited training examples, suffer from information loss in lower processing stages, and result in brittle representations due to pooling operations, requiring millions of examples to learn and being prone to false alarms from new distractors.

Innovation Solution

A neuromorphic computing platform with a multicore architecture inspired by the brain, featuring hardware-implemented spiking neural networks (SNNs) that support multiple modes of learning, including unsupervised, supervised, and reinforcement learning, using spike timing and asynchronous computation, and integrating with general computing hardware for dynamic pattern learning and sparse coding.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep machine learning approaches are used, then pattern recognition capability is improved, but generalization to novel situations deteriorates

Engineering Contradiction:
Improvepattern recognition capabilityVSAvoidgeneralization to novel situations
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent segments the neural network into specialized functional modules: spiking neural networks for temporal pattern recognition, neuromorphic cores for local computation, and hierarchical processing stages. This modular segmentation enables better generalization by allowing each segment to specialize in specific tasks while maintaining adaptability through configurable parameters and architectures.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements dynamic neural networks with adaptive spiking rates, time-varying synaptic weights, and flexible activation thresholds. The spiking neural networks dynamically adjust their firing rates based on input intensity and temporal patterns, enabling adaptive response to novel situations while maintaining robust pattern recognition through learned dynamics.

Inventive Principle:
Principle #15Dynamics

2Productivity

If pooling operations are applied, then computational efficiency is improved, but representation robustness deteriorates

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidrepresentation robustness
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent introduces spiking neural networks as an intermediary layer between feature extraction and classification stages. These spiking intermediaries preserve temporal information and distribute representations across multiple time steps, avoiding the information loss inherent in traditional pooling operations while maintaining computational efficiency through event-driven processing.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements continuous spiking representations that maintain information flow across time without discrete pooling interruptions. The spiking neural networks continuously integrate inputs over time through leaky integrators, preserving temporal dynamics and providing continuous robust representations rather than discrete pooled features.

Inventive Principle:
Principle #20Continuity of useful action

3Productivity

If traditional neural networks are used, then learning capability is improved, but energy consumption increases

Engineering Contradiction:
Improvelearning capabilityVSAvoidenergy consumption
Core Design Contradiction:
ProductivityVSUse of energy by moving object

Solution Approach 1:

The patent implements periodic spiking activity rather than continuous activation. Neurons fire spikes periodically when threshold conditions are met, creating sparse, event-driven computation that consumes energy only during actual computation events rather than continuously, significantly reducing overall energy consumption while maintaining learning capability.

Inventive Principle:
Principle #19Periodic action

Solution Approach 2:

The spiking neural networks implement self-regulating dynamics where neurons automatically adjust their firing rates and synaptic weights through learned mechanisms without requiring continuous external energy input for normalization or activation. The intrinsic leaky integrator dynamics and adaptive threshold mechanisms provide self-service regulation that reduces computational overhead and energy consumption.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3340124B1Sparse coding using neuromorphic computing
Publication Date: 2021.05.05 INTEL CORP
  • EP3340124B1 patent drawingFigure 1
  • EP3340124B1 patent drawingFigure 2A
  • EP3340124B1 patent drawingFigure 2B~2C

AI summary

A spiking neural network (SNN) includes artificial neurons interconnected by artificial synapses, where the spiking neural network is defined to correspond to one or more numerical matrices, and neurons of the SNN include attributes to inhibit accumulation of potential at the respective neuron responsive to spike messages. Synapses of the SNN have weight values corresponding to one or more numerical matrices. Inputs are provided to the SNN corresponding to a numerical vector. Steady state spiking rates are determined for at least a subset of the neurons and a sparse basis vector is determined based on the steady state spiking rate values.