Generalized Eligibility Traces for Spiking Neural Network Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing implementations of spiking neural networks lack a common framework for implementing eligibility traces, requiring reformulation of learning rules for each type of learning, such as supervised, unsupervised, and reinforcement learning, which complicates the process of state-dependent learning.
Innovation Solution
A generalized probabilistic learning system that determines eligibility traces based on the time history of input spikes and their rate of change, allowing for simultaneous handling of various learning rule combinations and updates in spiking neuron networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If separate learning rules are implemented for each type of learning (supervised, unsupervised, reinforcement), then each learning type can be handled specifically, but the system complexity increases and a common framework is lacking
Solution Approach 1:
The patent implements a universal eligibility trace mechanism that serves multiple learning types (supervised, unsupervised, and reinforcement learning) through a single common framework. The eligibility trace e_ij(t) acts as a state-dependent modifier for synaptic updates across different learning paradigms, eliminating the need for separate specialized rules for each learning type while maintaining their specific characteristics.
Solution Approach 2:
The patent introduces eligibility traces as dynamic parameters that modify synaptic update rules based on neuronal state. By changing the parameter set to include state-dependent eligibility traces e_ij(t) alongside traditional pre- and post-synaptic variables, the system achieves a unified framework that adapts to different learning types through parameter modulation rather than structural complexity.
2Device complexity
If eligibility traces are implemented without state-dependency, then the implementation is simpler, but the learning accuracy and adaptability to neuronal dynamics are reduced
Solution Approach 1:
The patent implements dynamic eligibility traces e_ij(t) that evolve over time based on neuronal state variables. The eligibility trace is not static but continuously updated according to the differential equation de_ij/dt = -e_ij/τ + f(pre-synaptic, post-synaptic), allowing it to adapt to changing neuronal dynamics while maintaining a computationally tractable form through exponential decay with time constant τ.
3Manufacturing precision
If computational updates are performed frequently and comprehensively, then learning accuracy improves, but computational cost and processing time increase
Solution Approach 1:
The patent computes eligibility traces continuously in the background according to pre-defined differential equations, preparing the state-dependent modifiers before they are needed for synaptic updates. This preliminary computation of e_ij(t) allows the actual learning updates to proceed efficiently by simply applying the pre-computed eligibility traces to the synaptic change rules, rather than computing everything from scratch at each update step.
Data Source
AI summary
Generalized state-dependent learning framework in artificial neuron networks may be implemented. A framework may be used to describe plasticity updates of neuron connections based on connection state term and neuron state term. The state connections within the network may be updated based on inputs and outputs to/from neurons. The input connections of a neuron may be updated using connection traces comprising a time-history of inputs provided via the connections. Weights of the connections may be updated and connection state may be time varying. The updated weights may be determined using a rate of change of the trace and a term comprising a product of a per-neuron contribution and a per-connection contribution configured to account for the state time-dependency. Using event-dependent connection change components, connection updates may be executed on per neuron basis, as opposed to per-connection basis.