Evidence-Gated Plasticity and Context-Dependent Consolidation for Multi-Goal Learning in Spiking Neural Networks

Background / Introduction: Biologically inspired spiking neural networks offer a mechanistic framework for adaptive behavior, yet learning multiple goals remains challenging when synaptic updates for different targets interfere. We investigated whether multi-timescale plasticity and context-specific credit assignment improve interleaved multi-goal learning in a spiking navigation system inspired by entorhinal-hippocampal circuitry. Methods: We introduce evidence-gated plasticity (EGP), in which reward-modulated spike-timing-dependent plasticity (STDP) generates transient synaptic proposals and need-weighted behavioral evidence modulates their slower consolidation into durable weights. Global EGP pools these proposals and the corresponding evidence across goals. Target-context EGP extends this framework through goal-specific proposal and evidence stores and additional need-dependent scaling of consolidation. We compared the algorithms across paired spiking networks, tested learned weights with plasticity disabled, and examined upstream representations and consolidation components. Results: Global EGP supported multi-goal learning and durable navigation. Target-context EGP improved early reward across 40 paired networks, including an independently specified 30-pair replication. Across eight paired networks trained for 2000 episodes, target-context EGP maintained a reward advantage, increased target-contact success, and reduced expected time to reach the target when unsuccessful blocks were included. Independent fixed-weight evaluation demonstrated durable navigation in both algorithms, with seven of eight networks favoring target-context EGP. A component experiment reproduced the target-context reward advantage and identified additional need-dependent scaling of context-specific consolidation as the only component with adjusted statistical support for improved reward. Conclusions: EGP supports durable multi-goal learning in spiking neural networks, and its target-context extension improves acquisition and conditional target selectivity in this interleaved navigation task. The framework offers a basis for testing context-sensitive, evidence-modulated plasticity in other multi-context learning problems.