Spiking Neural Network Calculator with Maximum Aggregation Layer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing hardware implementations of convolutional neural networks, particularly in embedded systems, face challenges due to their high complexity and resource requirements, especially in performing the MaxPooling function, which limits their efficiency in image classification tasks.
Innovation Solution
A calculator based on a pulse neural network architecture with a maximum aggregation layer that uses a device to activate neurons by comparing synapse addresses with reference addresses and hardness values, allowing for efficient calculation of output pulses without the need for extensive resource allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If classic convolutional neural network classifiers are implemented in embedded systems, then image classification performance is improved, but device complexity and resource requirements increase
Solution Approach 1:
The patent replaces traditional digital memory-based synapse implementations with analog resistive memory components. This substitution enables the synapses to dynamically adjust their conductance based on voltage, eliminating the need for separate digital memory storage and reducing overall hardware complexity while maintaining classification performance
Solution Approach 2:
The patent changes the implementation parameter of synapses from digital memory-based to analog resistive-based. This parameter change allows the system to reduce hardware resources by utilizing the voltage-conductance relationship in resistive memory to achieve both storage and computation functions
2Reliability
If MaxPooling function is implemented with traditional methods, then classification performance is maintained, but memory and computational resources are excessive
Solution Approach 1:
The patent merges the MaxPooling operation with the convolutional layer processing by integrating the pooling function into the existing neural network architecture. This consolidation eliminates the need for separate MaxPooling layers and reduces overall computational and memory resources required while preserving classification performance
Solution Approach 2:
The patent makes the resistive memory components serve multiple functions: they act as both synapse weight storage and as part of the MaxPooling computation mechanism. This multi-functionality reduces the total hardware resources needed by eliminating dedicated MaxPooling hardware
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention proposes a computer (10) based on a spiking neural network, comprising at least one maximum aggregation layer. In response to an input spike received by a neuron of the maximum aggregation layer, the device is configured to receive the address of the activated synapse. The device comprises an address comparator configured to compare the address of the activated synapse with a set of reference addresses. Each reference address is associated with a hardness value and with an aggregation neuron. The device activates a neuron of the maximum aggregation layer if the address of the activated synapse is equal to one of the reference addresses and if the hardness value associated with this reference address has the highest value among the hardness values associated with the other reference addresses of the set.