Recurrent scale and adaptation: Falandays and SORN

Findings and surprises

Across these tasks, the value of local adaptation depends on recurrent strength, the adaptation mechanism and the task. Synaptic and intrinsic adaptation have distinct effects, and their combination can improve or impair task performance.

The comparison extends the initialisation–plasticity question to two models under shared Tracking and Pong worlds. Each condition has 32 independent wiring blocks, with four full-reset trials per block and six connections per sensory/output port.

The fresh Falandays Pong results retain the original frozen-synapse advantage across all five sampled scales. Tracking retains a scale-dependent reversal: its largest frozen-synapse benefit appears at scale 0.3, while synaptic adaptation helps at 1 and 3. The two smallest Tracking scales have intervals spanning zero.

SORN exposes a different interaction on Pong. Intrinsic threshold adaptation lowers hit rate at the two smallest recurrent budgets and raises it at budget 1. Its synaptic-freezing contrasts are smaller than Falandays', with an interval above zero only at budget 1 on Pong. SORN's Tracking contrasts all include zero. At budget 3, its adaptation policies produce very similar task scores in both tasks.

All 10,240 trials across 80 cells completed, with no failures. The four execution records preserve the complete exploratory sweep.

Motivation

Marc Bacvanski's observation that smaller initial weights could favour frozen Falandays networks prompted the original investigation. His question separates adaptation that brings a network into a useful operating range from adaptation that helps it continue performing. The original report retains the four-task study and its construction review. This fresh comparison concentrates on Tracking and Pong and adds SORN. See Marc's website.

Tracking

The primary contrast is thresholds adapting minus both adapting. Positive values favour frozen synapses while intrinsic adaptation remains active; negative values favour synaptic adaptation. The Falandays contrast is +0.278 [0.181, 0.375] at scale 0.3, −0.336 [−0.393, −0.278] at 1 and −0.399 [−0.458, −0.335] at 3, in raw track_score.

SORN's frozen-synapse contrast ranges from +0.0002 to +0.0209 across the grid. Every interval includes zero, including an interval whose lower endpoint is exactly zero at budget 3. The intrinsic-adaptation contrasts also span zero. This set supplies little evidence of a consistent adaptation benefit on Tracking at these budgets.

Pong

The Falandays frozen-synapse benefit ranges from +0.141 to +0.187 in hit_rate. Every pointwise 95% interval is above zero. Both conditions with frozen synapses have higher mean hit rates than their synaptically adapting counterparts.

For SORN with synapses adapting, the threshold-adaptation effect is −0.102 [−0.179, −0.029] at budget 0.03 and −0.100 [−0.172, −0.029] at 0.1. It becomes +0.086 [0.050, 0.124] at budget 1. The corresponding effects with synapses frozen have the same directions and intervals excluding zero. SORN's frozen-synapse benefit at budget 1 is +0.027 [0.006, 0.049]; the other four intervals span zero.

The size and direction of the adaptation effects therefore depend on the local mechanism as well as recurrent scale. The condition panels show that a larger SORN row-sum budget also lowers mean Pong performance across all four policies in this set.

What carries across the models and tasks

The clearest shared pattern is a change in the value of adaptation across recurrent strength settings. Falandays Tracking favours frozen synapses at scale 0.3 and adapting synapses at 1 and 3. SORN Pong shows a reversal in threshold adaptation: it reduces hit rate at budgets 0.03 and 0.1 and improves it at 1. The reversals occur in different mechanisms, which makes their separate controls useful.

The mechanisms also interact. On Falandays Tracking at scale 0.3, activity-target adaptation improves raw track_score by +0.525 [0.438, 0.603] with synapses frozen. With synapses adapting, its effect is −0.111 [−0.153, −0.070]. A single plasticity switch would conceal this difference between the two local adaptation processes.

Fixed recurrent networks can support useful task performance. The Falandays Pong advantage with frozen synapses persists across the sampled scales in both the original study and these fresh standardised-port runs. SORN's synaptic-freezing effects are generally much smaller; its clearest Pong effects involve thresholds. Together, the results show why recurrent strength, synaptic adaptation and intrinsic adaptation need to be considered jointly when interpreting a task score.

These observations support Marc's operating-range hypothesis as a direction for mechanism experiments. The present sweeps compare adaptation policies from the first frame. They leave open how much adaptation establishes a useful regime and how much continuing adaptation helps after that regime has formed.

What the SORN comparison measures

Falandays scales its initial recurrent weights. SORN scales its excitatory incoming row-sum budget, which governs initial weights and the strength maintained by later normalisation. Inhibition, input weights and initial thresholds retain their paired draws across the positive scale grid. The SORN synaptic branch includes STDP, permanent pruning and normalisation; its intrinsic branch adapts firing thresholds.

The four functional policies are both adapting, thresholds adapting, synapses adapting and both frozen. A tick-1 freeze acts before the first neural frame. SORN's native learning rates remain fixed across scales, so the relative STDP increment also changes with the row-sum budget. These are model-specific mechanisms within a shared task comparison. An initial-weight-only SORN perturbation with a fixed maintained budget would need a new construction axis.

Construction and uncertainty

All four plans use fresh declared world streams, disjoint from the original study and the empirical calibration trajectories. Within a task, model and block, construction and worlds are paired across scales and adaptation policies. The two models share task-world streams; their connectomes and local equations differ. SORN contains 53 excitatory and 11 inhibitory nodes; Falandays contains 64 nodes. Host preparation checked 2,560 initial states and six connections at every active sensory/output port.

The fixed grid is 0.03, 0.1, 0.3, 1 and 3. There is no adaptive refinement or smoothing. Raw intervals average four reset trials per block and resample the 32 independent blocks, using 5,000 Philox samples and seed 2026101202. The same resampling indices are used across conditions within each task/model. Intervals are pointwise 95% percentile intervals. The outcome tables retain the runner's normalised estimates, independent calibration uncertainty, scoring windows and calibration status. All attempted cells are retained; a failed trial makes its complete cell estimate and dependent contrasts unavailable.

The earlier study used eight blocks and Bernoulli port wiring. This comparison uses fresh seeds, 32 blocks and fixed-degree ports. Differences between the studies reflect that combined change; this study does not separately estimate the port-wiring effect.

Execution and retained data

The plans use CPU64 ordered sums, one CPU thread, batch capacity 128, a 256 MiB managed budget and summary recording. Tracking runs for 2,000 world ticks and Pong for 7,200, with no warm-up and one neural frame per world tick. The physical task signatures match the retained 1,024-trajectory calibrations for this arm64 runtime.

All four original records, trial tables, seeds, summaries, captured source and calibrations are retained in the run-data supplements. Summary recording supplies task outcomes; trajectories of firing, thresholds and weights require a separately specified diagnostic recording. These are exploratory task results.

Next questions

The next direct test of the operating-range hypothesis is timed freezing. Let paired conditions adapt for the same initial period, then compare continued adaptation with freezing synapses, intrinsic adaptation or both. Prespecified switch times and fresh blocks would separate a shared adaptation history from the value of continued updates. Record activity, recurrent strength and thresholds alongside task outcomes to locate the changes that accompany each intervention.

For SORN, an initial-threshold sweep is a supported initial-state assay; a pure initial recurrent-weight perturbation needs a declared additional axis. Mechanism recordings can locate changes in activity and recurrent strength, while fresh confirmation blocks can test the selected contrasts.