Charleston, Illinois

NeuroCore

NeuroCore is our neuroscience research program. It studies the kind of signals the brain produces, brief spikes in time, and asks what a machine built on the same principle can learn from them, and whether it can do some kinds of work more efficiently than a conventional network. This paper states the question, the approach, and what we will measure. It claims no results, because there are none we are ready to stand behind in public yet.

Back to the home page · Our research page

The question

The program stands on one research question, written down before any experiment ran:

Can a compact, adaptive spiking substrate deliver useful real-time learning or temporal processing capabilities that conventional approaches cannot deliver as efficiently under equivalent operating constraints?

It is a testable hypothesis, never an assumption. Where a conventional algorithm does the job better, it wins, and the program is charged to say so. We would rather prove a promising idea does not work than defend what measurements have disproven.

The approach

The brain does not compute in a steady stream of numbers. Its neurons sit quiet, gather charge, and fire a spike when a threshold is crossed, and the timing of those spikes carries the information. NeuroCore builds one small network on that principle and tests it hard.

  • One substrate, not three systems. A single recurrent network of leaky integrate-and-fire neurons carries every experiment. Associative memory is that substrate with its learning rules and readout; attention is gating inside the same network. The field's working projects are built this way, and so is ours.
  • Three test regimes. Temporal processing (recognizing patterns that unfold in time), associative memory (learning and holding associations), and attention (prioritizing events in a continuous stream).
  • A fourth, interaction objective. How the regimes affect each other: whether attention gating disturbs memories already established, whether sequential learning produces interference between functions that each performed well alone, and whether the network can adapt without retraining the whole substrate. No result here will be read as evidence about cognition or intelligence.
  • Learning rules. Surrogate gradient training carries the benchmark work. Spike timing dependent plasticity, the learning rule closest to the biological one, runs as a named, instrumented experiment of its own.
  • Scale, kept honest. The baseline substrate is 512 neurons. It grows to 1,024, and then to 2,048, only where a measurement shows the added capacity buys a real improvement. Dense recurrence grows about sixteenfold between 512 and 2,048 neurons, so every increase in complexity must purchase a measurable benefit.
  • The bench. The work runs in software, with PyTorch and the snnTorch library, on an isolated laptop bench set apart for research. Versions are pinned and the environment is frozen, so a run can be repeated exactly.

Program state

November 2026

NeuroCore was approved as a controlled research initiative on November 8, 2026, after an independent assessment inside the company amended its design: one substrate instead of three systems, the energy claim restated honestly, and the outcome criteria registered in advance. The program is internal and in its first phase: the bench environment is frozen, the research design and its criteria are written down, and the baseline experiments are being prepared.

A first small cohort of participants is in preparation for the parts of the program that involve people. Participation will be covered by written acknowledgement agreements; our attorney has advised that simple acknowledgement agreements from participants are enough to satisfy the law.

No integration with any other Eidon Labs system is authorized. Any future proposal of that kind would be judged on its own, under separate review, and nothing in this program touches our companion systems or their records.

What we will measure

  • Temporal pattern recognition accuracy.
  • Learning speed and stability.
  • Retention: what the substrate still holds after time and after new learning.
  • Adaptation to novel sequences.
  • Latency, and memory use, at each scale.
  • Resource efficiency, measured as operations, latency, memory, and actual system power on the bench.
  • Head to head results against conventional networks, given a comparable training opportunity, with size and compute budgets documented.

The controls are registered in advance: repeated independent training runs, never a best run reported alone; evaluation data held back and never used for selection or tuning; averages reported with variation and confidence intervals; and full reproducibility, with seeds, versions, datasets, parameters, and configurations recorded. The outcome classes are also registered in advance: an experimental failure closes a regime, an architectural failure closes the program, and a partial technical success narrows it. A failed experiment is never rescued by moving its criteria, and a successful one is never discarded because a separate hypothesis failed.

Limits and open questions

  • There are no public results yet. Anything you read elsewhere suggesting NeuroCore has demonstrated an advantage is ahead of the record.
  • The famous energy advantage of spiking networks is measured on neuromorphic silicon, where silence is free. Simulated on an ordinary graphics card, every timestep is computed whether a neuron fires or not, and the best proxy evidence we found puts the advantage at about 3 times under tuned conditions, not the 10 to 100 times sometimes quoted. Silicon benefits stay projections until they are measured on representative hardware.
  • At this scale, the published record says spiking networks are competitive, not dominant: on the standard spoken digits benchmark, published spiking results run about 83 to 91 percent against 92.4 percent for a conventional convolutional network. Recurrence is worth a great deal; the encoding of the input is worth more than the architecture.
  • Retention is where spiking networks currently trail. On shared continual learning benchmarks, secondary reports put spiking results well behind conventional ones, and forgetting is a documented weakness of the learning rules themselves. The memory regime is therefore the program's hardest test, not its showcase.
  • Nothing published, by us or anyone, supports claims of open-ended associative memory or cognition at this scale. We make none.

Our research page · Back to the home page