A forecast-based STDP rule suitable for neuromorphic implementation

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering

Background:

• Spiking neural networks (SNNs) offer computational efficiency, particularly for on-chip implementation with neuromorphic hardware.
• Traditional learning rules like Spike Timing Dependent Plasticity (STDP) face challenges in efficient implementation on such hardware.
• Alternative learning rules based on post-synaptic neuron membrane potential are being explored.

Purpose of the Study:

• To introduce and evaluate a novel learning rule, STDP TTS (Time-To-Spike), for spiking neural networks on neuromorphic hardware.
• To demonstrate the rule's efficacy in improving learning efficiency and real-time pattern detection.
• To reduce computational complexity and memory bandwidth requirements for SNNs.

Main Methods:

• Implementation of the STDP TTS learning rule on SpiNNaker neuromorphic hardware.
• Modification of the standard STDP algorithm, using membrane potential to predict time-to-spike for Long Term Potentiation (LTP).
• Utilizing the standard STDP rule for Long Term Depression (LTD).

Main Results:

• The STDP TTS algorithm enables statistical prediction of time-to-spike, triggering LTP effectively.
• Approximations in the rule lead to reduced computational time and memory bandwidth.
• On-chip simulations demonstrated improved reliability and speed in pattern detection by the neural network.

Conclusions:

• STDP TTS offers a computationally efficient and easily implementable learning rule for SNNs on neuromorphic hardware.
• The rule is suitable for real-time learning in time-critical applications.
• This approach provides a promising alternative and complement to standard STDP for neuromorphic systems.