Make network.reset_state_variables() actually reach the learning rules

[Hananel-Hazan]

Supersedes #777. Co-authored with @saachigoyall, who found and diagnosed the bug.

What is wrong

AbstractFeature.reset_state_variables forwards the reset to self.learning_rule, and it is the only place that does. Every concrete feature overrode it with a bare pass and none called super(), so that forwarding line was unreachable.

The result: network.reset_state_variables() never reached any MulticompartmentConnection learning rule. Driving an MSTDPET connection for 250 steps and then resetting the network left every state variable untouched, including the two that MSTDPET.reset_state_variables already cleared.

On top of that, the rules themselves were incomplete. MSTDPET cleared 2 of its 6 variables (#777). MSTDP and PostPre cleared nothing at all.

Changes

bindsnet/network/topology_features.py

• Drop the seven bare-pass overrides (Probability, Mask, MeanField, Weight, Bias, Intensity, Degradation) so they inherit the base implementation.
• Drop @abstractmethod from AbstractFeature.reset_state_variables, which is why those overrides existed in the first place. This only relaxes the contract: a subclass may still override.
• AdaptationBaseSynapsHistory and AdaptationBaseOtherSynaps keep their own reset logic and now chain to super() first.

bindsnet/learning/MCC_learning.py

• MSTDPET now also clears p_plus, p_minus and the moving-average buffer (@saachigoyall's change).
• MSTDP now clears eligibility, p_plus, p_minus, the moving-average buffer, and the fast path's one-step spike lag. That lag is newer than Fix MSTDPET.reset_state_variables to clear p_plus, p_minus, and moving-average buffer #777, and it matters: without clearing it the first step of an episode pairs with the last step of the previous one, which is exactly the cross-episode contamination Fix MSTDPET.reset_state_variables to clear p_plus, p_minus, and moving-average buffer #777 set out to fix.
• PostPre now clears both averaging buffers and their indices.
• Hebbian holds no state between steps; its no-op is now documented as deliberate rather than looking like an oversight.
• MSTDP's lazily-built state (_prev_source_s, _prev_target_s, eligibility) is declared None in __init__ and guarded with is None instead of hasattr, so reset has something well-defined to restore.

Tests

Eight new cases in TestLearningRuleReset. Seven of the eight fail against master, so they genuinely pin the bug. The strongest one runs two identical episodes with a reset between them and requires identical resulting weights.

This replaces the test from #777, which passed tc_plus and average_update to Weight. Weight.__init__ has a fixed signature that never accepted them, so that test raised TypeError rather than running, on master and on its own branch alike.

Full suite: 91 passed (83 before). black --check and isort clean on the changed files.

Performance

Reset runs once per episode, not per time step, so the added clearing is off the hot path. Two edits do touch the per-step update path (hasattr to is None), so I measured: three interleaved A/B rounds against master, MSTDP fast path, MSTDPET, and MSTDP dense path, at n=64 and n=256. All differences under 1% with the sign varying between rounds, i.e. no measurable change.

🤖 Generated with Claude Code