Neuromorphic Synapse Segmentation for Asymmetric Conductance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neuromorphic processing architectures, particularly spiking neural networks (SNNs), face challenges in accuracy due to asymmetric conductance responses in memristive devices like PCM cells, which affect performance in complex cognitive tasks such as pattern recognition and classification, especially when dealing with partially overlapping patterns like handwritten digits.
Innovation Solution
The introduction of a preprocessor unit that generates two sets of input spike signals, one encoding the original pattern and the other encoding complementary values, which are processed by a spiking neural network to adjust synaptic weights during a learning mode, enhancing the network's sensitivity for correlation detection and mitigating issues with asymmetric conductance responses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If synapses are implemented using single PCM cell with asymmetric STDP rule, then device complexity is reduced, but manufacturing precision deteriorates due to abrupt cell reset becoming critical for complex tasks
Solution Approach 1:
The synapse is segmented into two separate PCM cells (first and second cells) that operate in parallel. Each cell processes one aspect of the input signal, allowing the system to avoid the abrupt reset problem of single-cell implementations while maintaining relatively simple device structure. This segmentation enables more precise weight adjustment without critical reset issues.
Solution Approach 2:
Different characteristics are assigned to different parts of the synapse implementation. The first PCM cell handles positive weight adjustments while the second cell handles negative weight adjustments, allowing each cell to be optimized for its specific function and reducing the impact of asymmetric conductance responses.
2Manufacturing precision
If synapses use more than one cell in differential configuration, then manufacturing precision improves by mitigating asymmetric conductance response, but device complexity increases and requires cyclic rebalancing
Solution Approach 1:
The synapse is divided into two functional segments (first and second PCM cells) that work together to mitigate asymmetric conductance effects. This segmentation provides the precision benefits of multi-cell implementations while using a standardized differential configuration that reduces overall system complexity.
Solution Approach 2:
Instead of trying to correct asymmetric conductance through complex control mechanisms, the invention inverts the approach by using two cells with opposite polarity adjustments. The first cell potentiates weight while the second cell depresses weight, naturally balancing the asymmetric responses without requiring cyclic rebalancing operations.
3Speed
If conventional single-set input patterns are used, then processing speed is maintained, but measurement precision deteriorates due to inability to distinguish partially overlapping patterns
Solution Approach 1:
The input pattern is segmented into two complementary sets: original input patterns and inverted input patterns. Each set is processed through dedicated synapse cells, allowing the system to distinguish between partially overlapping patterns more effectively while maintaining processing speed through parallel operation.
Solution Approach 2:
The inverted input patterns act as an intermediary representation that provides additional discriminatory information. By processing both original and inverted patterns through the differential synapse configuration, the system gains enhanced pattern recognition capability without significant speed penalty.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach improves the accuracy of SNNs by considering both original and complementary patterns, leading to better performance in distinguishing overlapping patterns and reducing the impact of asymmetric conductance responses, as demonstrated in MNIST classification tasks, with improved power consumption and balanced spike contributions.
Implementation Method 1
The synapse receives spikes from its input neuron and provides post-synaptic signals ('post-synaptic potentials'), to its output neuron. A post-synaptic signal is a graded synaptic potential which depends on conductance (also known as 'synaptic efficacy' or 'weight') of the synapse.
Implementation Method 2
nanoscale synapses can be realized using memristive properties of nanodevices, e.g. resistive memory cells such as phase-change memory (PCM) cells
Implementation Method 3
Resistive memory cells such as PCM cells also exhibit an asymmetric conductance response, whereby the process of reducing the cell resistance differs from that for increasing cell resistance.
Data Source
AI summary
Neuromorphic processing apparatus is provided. The present invention may include a spiking neural network comprising a set of input spiking neurons each connected to each of a set of output spiking neurons via a respective synapse for storing a synaptic weight which is adjusted for that synapse in dependence on network operation in a learning mode of the apparatus, and each synapse is operable to provide a post-synaptic signal, dependent on its synaptic weight, to its respective output neuron. The present invention may further include a pre-processor unit adapted to process input data, defining a pattern of data points, to produce a first set of input spike signals which encode values representing respective data points, and a second set of input spike signals which encode values complementary to respective said values representing data points, and to supply the input spike signals to respective predetermined input neurons of the network.