This investigation studies how scalable gradient methods and biologically plausible local learning rules diverge when training recurrent spiking neural networks on temporal tasks.

SNN learning sits between non-differentiable spike events and the need for long-horizon credit assignment. The work asks where surrogate gradients fail, where local eligibility traces help, and how hybrid strategies behave in online settings.

• Benchmarked SuperSpike, SLAYER, and EXODUS-style surrogate gradients.
• Compared biologically plausible alternatives including e-prop, RTRL variants, and STDP-inspired local updates.
• Measured gradient pathologies introduced by leaky integrate-and-fire dynamics across deep temporal unrollings.
• Tested hybrid learning rules that combine global task signal with local eligibility traces.

• Used temporal classification tasks with controlled spike sparsity and sequence length.
• Tracked vanishing and exploding gradient regimes under recurrent SNN dynamics.
• Compared online update feasibility against full BPTT baselines.
• Logged task accuracy, spike rate, memory cost, and update latency.

• Surrogate gradients remain the strongest baseline for task accuracy but scale poorly under long unrollings.
• Eligibility traces reduce memory pressure but require careful stabilization to compete with BPTT.
• Hybrid strategies produced promising online behavior but underperformed full BPTT on harder temporal dependencies.

• Current tasks are diagnostic rather than large-scale neuromorphic deployments.
• The work has not yet evaluated hardware-specific energy behavior.
• Hybrid update rules need stronger theoretical framing.

• Test on event-camera and streaming audio workloads.
• Measure energy-latency tradeoffs on neuromorphic hardware targets.
• Use representation diagnostics to compare temporal credit localization.