Neuromorphic STDP Replay Eliminates Backward Connectivity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neuromorphic hardware faces challenges in efficiently implementing long-term potentiation (LTP) due to the need for backward connectivity and increased overhead, which reduces SNN capacity and increases power consumption, especially in sparsely connected or recursive networks.
Innovation Solution
The proposed solution involves replaying each neuron spike after a maximum STDP time interval, allowing LTP to be performed without backward connectivity, by calculating the time difference between the replay spike and the POST spike to determine the synaptic weight update, thus eliminating the need for backward connectivity and reducing energy and memory overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If backward connectivity is implemented to perform LTP, then LTP functionality is achieved, but chip area overhead and energy consumption increase
Solution Approach 1:
Instead of implementing backward connectivity to send signals from post-synaptic to pre-synaptic neurons, the patent inverts the approach by having pre-synaptic neurons send forward connectivity signals that include their identity information. This allows the post-synaptic neuron to perform LTP calculations without requiring physical backward connections, thereby reducing chip area overhead while maintaining LTP functionality.
Solution Approach 2:
The patent extracts the essential information needed for LTP (pre-synaptic neuron identity and timing) from the backward connectivity requirement. By including this information in forward connectivity spike packets, the system separates the information transfer function from the physical connection direction, eliminating the need for dedicated backward connectivity hardware.
2Adaptability or versatility
If backward connectivity is implemented to perform LTP, then LTP functionality is achieved, but energy consumption increases
Solution Approach 1:
The patent inverts the traditional LTP implementation by eliminating backward connectivity and instead using forward connectivity with embedded identity information. This reduces energy consumption by removing the need for separate backward signal transmission paths while preserving the ability to calculate and execute synaptic weight updates based on spike timing differences.
Solution Approach 2:
The patent extracts the timing and identity information from backward connectivity requirements and incorporates it into forward connectivity packets. This allows the system to perform LTP using only forward connections, reducing the energy overhead associated with maintaining and utilizing backward connectivity infrastructure.
3Measurement precision
If maximum STDP time interval is increased to improve learning accuracy, then learning precision is improved, but latency increases
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing the maximum STDP time interval value in each pre-synaptic neuron. When a spike is generated, this pre-stored value is immediately included in the spike packet without requiring real-time calculation or lookup, thereby enabling long time intervals for learning accuracy without adding computational latency to the spike transmission and processing path.
Data Source
AI summary
System and techniques for spike timing dependent plasticity (STDP) in neuromorphic hardware are described herein. A first spike may be received, at a first neuron at a first time, from a second neuron. The first neuron may produce a second spike at a second time after the first time. At a third time after the second time, the first neuron may receive a third spike from the second neuron. Here, the third spike is a replay of the first spike with a defined time offset. The first neuron may then perform long term potentiation (LTP) for the first spike using the third spike.