Initial weight scale and homeostatic plasticity

Findings and surprises

Initial recurrent weight scale changed whether weight adaptation helped in these 64-node Falandays-type networks. With activity-target adaptation enabled, Tracking favoured frozen recurrent weights at scale 0.1 and adapting weights at scales 1 and 3. Pong favoured frozen recurrent weights at all eleven sampled scales.

The clearest surprise was how far the Pong advantage extended: it appeared at zero initial recurrent weights as well as at small and large scales. Frozen networks still receive sensory input, retain membrane state and use fixed output readouts. This result makes the sensory and readout pathways worth investigating alongside recurrent adaptation.

The curves were also noisier than we expected. Each cell contains eight independent wiring blocks with four fully reset trials per block. The 5,632 trials describe the total sweep workload; the independent wiring count determines precision here. Refinement added sampled scales using those same blocks. Small bends and selected transition brackets therefore remain exploratory.

A construction review confirmed the intended pairing and identified variation in sensory and output connection counts. The platform now provides fixed-degree ports as its basic setup. The results below retain the original Bernoulli wiring; how the new policy affects performance and variability awaits a new experiment.

Motivation

Personal correspondence with Marc Bacvanski prompted this investigation. He described observing that smaller initial weights could favour frozen Falandays networks. This raises two questions: which starting scales support useful behaviour, and does adaptation remain useful after reaching that range?

We began with a scale sweep and separate controls for weight adaptation and local activity-target adaptation. The study uses declared BrainlessLab presets. Comparing Marc's exact configuration and scoring protocol would help establish how closely the two investigations align.

What we compared

The network integrates sensory input and previous spikes, then fires at a threshold set by a local activity target. Its post-spike membrane error drives weight and target updates. The target also controls the firing threshold, so weight adaptation and target adaptation can have different effects.

The primary comparison is targets adapting minus both adapting. It measures the benefit of freezing recurrent weights while activity targets continue to adapt. Positive values favour frozen weights; zero indicates no mean difference in the recorded task outcome. The supporting panels show target effects and the interaction between the two adaptation loops.

Scale 1 denotes each task preset's initial recurrent matrix. Other multipliers change its initial mean and spread together. Wiring, sensory projections and world streams stay paired across scales and conditions within each task.

Main findings

The study completed on 12 October 2026. Twelve scientific records retain 176 cells and 5,632 completed trials, with zero failed trials. Each task has eleven sampled scales and four conditions. Intervals are pointwise 95% paired block-bootstrap intervals; adaptively selected peaks and crossings require fresh-seed confirmation.

Tracking: the contribution of adaptation reversed

At scale 0.1, frozen weights improved the mean tracking score by 0.138 [0.057, 0.215]. At scale 1, adapting weights improved it by 0.282 [0.154, 0.402]; at scale 3, the improvement was 0.269 [0.170, 0.343].

The sampled mean contrast changed sign between scales 0.5477 and 0.7401. Both endpoint intervals include zero. The bracket describes the observed mean curve; it gives an exploratory region to examine with independent wirings.

The four condition panels show how absolute performance changes. At scale 0.1, enabling target adaptation while weights adapt changes the mean score by −0.103 [−0.200, −0.011]. The weight-by-target interaction is −0.130 [−0.243, −0.020]. Separating the two adaptation loops makes this interaction visible.

Pong: frozen weights helped throughout the sampled range

The mean advantage from freezing recurrent weights ranged from 0.132 to 0.267 in hit rate. Every pointwise interval had a positive lower endpoint. The advantage appeared at scale 1 and at zero initial recurrent weights, extending beyond the small-scale region that motivated the study.

The both-frozen condition follows a similar high-scoring curve. Together, these outcomes suggest examining what the fixed sensory projections, membrane dynamics and output selections already contribute under this Pong preset. Summary recording does not retain the dynamic histories needed to resolve that mechanism.

Other tasks

Delayed cue: a candidate for independent checking

Most primary contrasts have intervals spanning zero. At the adaptively selected scale 0.4054, targets-adapting accuracy is 0.656 [0.531, 0.781], with a paired advantage of 0.281 [0.031, 0.531] over both-adapting.

At scale 0.1, all four conditions have identical recorded outcomes and zero paired differences. This analysis combines delays of 8, 32 and 128 ticks. A follow-up should increase trials and separate delay lengths.

CartPole: relative effects within short-lived behaviour

At scale 3, continuing weight adaptation improves the mean by 6.16 fitness units [2.91, 9.25]. Most other primary contrast intervals span zero. At scale 1, target adaptation with frozen weights adds 4.19 units [1.16, 8.59].

Absolute performance remained short-lived: scoring windows span 7–236 world steps, the highest cell mean is 23.53, and the matching random-action calibration mean is 21.96. These outcomes motivate focusing the next main phase on Tracking and Pong. The CartPole records, curves and tables remain part of this study.

What the construction review changes

The original comparisons use identical wiring within each paired block, and every trial restores its initial weights, targets, membranes, spikes and learning flags. Between blocks, Bernoulli port sampling varies how many nodes receive each sensory channel and how many contribute to each output.

One Tracking block had a disconnected receptor. Another receptor encodes the same bearing and connects to six nodes, so that bearing still reaches the network. Omitting the affected block in a descriptive sensitivity check preserves the main Tracking reversal: the primary contrast changes from 0.138 to approximately 0.156 at scale 0.1, and from −0.282 to approximately −0.264 at scale 1.

The revised basic setup gives each active receptor a fixed number of destinations and each effector an equally sized population. At 64 Falandays nodes and probability 0.1, this gives six connections per port. Identities remain random and are shared within paired comparisons. The change removes port-degree variation by construction; a new experiment will measure its effect on task outcomes.

Next experiment and a SORN comparison

A new paired study can test whether the main patterns persist with fixed-degree ports, more independent wiring blocks and fresh world streams. Recording spike activity, weight changes and target trajectories would help explain the operating regimes. A later timed-freezing comparison would begin with a shared adaptation period and score a common subsequent window.

SORN provides a useful comparison because it combines local synaptic updates, synaptic normalisation and intrinsic threshold adaptation. These mechanisms are described in Lazar, Pipa and Triesch's SORN paper. The local implementation has excitatory and inhibitory populations and fixed sensory and inhibitory weights.

Four analogous conditions are available: both mechanisms adapting; synaptic adaptation frozen with thresholds adapting; synaptic adaptation active with thresholds fixed; and both frozen. In SORN, freezing weights also stops synaptic normalisation. Setting the STDP rate to zero alone leaves normalisation active.

A supported first sweep varies the incoming excitatory weight row sum. This changes both initial excitatory strength and the strength maintained by normalisation. It tests a scale–adaptation relationship with a persistent recurrent budget. An initial-weight-only perturbation with a fixed normalisation budget would require a separately specified construction axis.

Prepared follow-up plans pair Tracking and Pong worlds across SORN and new Falandays references. Both use 64 total nodes and six connections per active sensory or output port, with their own declared weight and threshold rules. The candidate grid is 0.03, 0.1, 0.3, 1 and 3 relative to each model's reference setting, with 32 independent blocks and four reset trials per cell. The complete two-model, two-task sweep contains 10,240 reservoir trials. These are planned comparisons; no SORN results are reported here.

For SORN, zero excitatory row sum is excluded from that candidate grid. The current constructor skips weight draws at zero, shifting other random draws. Zero excitatory weights also leave the excitatory–inhibitory feedback pathway in place. A zero-scale control needs its own declared meaning and construction checks.

Initial excitatory thresholds offer a second, distinct SORN initial-state experiment. Their range can vary while recurrent weights remain fixed across conditions. This would ask whether intrinsic adaptation compensates for different starting excitabilities.

Method and source mapping

Scientific protocol

The coarse Falandays grid is 0, 0.01, 0.03, 0.1, 0.3, 1 and 3. Two bounded refinement rounds added two scales per task per round, prioritising primary-contrast crossings, intervals adjoining local extrema and large adjacent changes. Geometric midpoints were selected before execution. The same eight blocks were used throughout.

Each cell has four reset trials per independent block. Construction and world streams are shared across scales and conditions within each task. Root seed is 2026100801; the partition is initialisation-plasticity-exploration-v1. Analysis averages trials within blocks and uses 5,000 shared Philox block-bootstrap resamples, seed 2026100802.

All reservoirs have 64 nodes. Input gain is 8 for CartPole and 1 elsewhere. Leak is 0.25, target floor 1 and connectivity probability 0.1. Enabled weight and target adaptation rates are 1 and 0.01 respectively. All records use the original Bernoulli sensory and output wiring.

Scoring starts without warm-up. Tracking runs 2,000 ticks and Pong 7,200 ticks. CartPole ends on physical failure or at 15,000 world steps, with 24 neural frames per action. Delayed cue has an eight-tick transient cue, delays of 8, 32 or 128, and an eight-tick response window.

Execution uses CPU64, ordered sums, one CPU thread and summary recording. Task outcomes retain raw and normalised values, calibration status, scoring windows and failure counts. All three physical-task calibration signatures matched; their null calibrations contain 1,024 independent trajectories each. No invalid-denominator bootstrap draws were recorded. Delayed cue uses analytic chance 0.5. A separate four-record smoke workflow completed 64 diagnostic trials.

Model lineage and preset differences

The local mechanism follows Falandays, Nguyen and Spivey's 2021 homeostasis-as-prediction model and the resonance paper, published online in 2023 and in Cognitive Neurodynamics 18 in 2024. The latter applies homeostatic reservoirs to Tracking, Pong and wall avoidance. CartPole and delayed cue extend the mechanism to additional task interfaces.

The transcription is retained at revision d0fc756d. It is distinct from the resonance paper's general zero-mean, unit-spread weight description. The retained source-mapping note gives complete distributions and references.

Complete outcomes and refinement records

The following primary contrasts use native outcome units. Entries show the raw difference and its pointwise 95% paired interval at three coarse-grid anchors.

Full tables retain every sampled cell, calibration status and scoring window.

Tracking tables

Pong tables

Delayed cue tables

CartPole tables

Refinement decisions

Retention and reproduction

The workspace retains original attempts, exact plans, seeds, calibration inputs, captured source, environments, compact trial tables and separately executed analysis receipts. Summary recording retains no full trajectories. Export includes original run-data supplements and retained analysis inputs within the publication cap.

The final analysis retains both refinement decisions and their complete dependency closure. Both decisions were replayed against their frozen inputs and matched. Specified simulation reruns and reanalysis use the retained inputs and original source. An independent scientific rerun and numerical comparison have not been executed.

The revised article and figures present the same recorded outcomes and intervals. The fixed-degree setup and the planned SORN comparisons remain distinct from those original execution records.