Stochastic & Neuromorphic Computing  /  Explore  /  Labs

Labs
emerging directions

Not everything in a research toolkit is production-ready, and pretending otherwise would break the honesty the rest of this site is built on. These three directions are earlier-stage — grounded in real code, but experimental — and interesting enough to show with that label attached: a reservoir that tunes itself, attention without the quadratic cost, and computing with light.

Auto-critical reservoir computing research

A reservoir (a liquid-state or echo-state machine) computes by letting a fixed random recurrent network transform an input, and reading out from its rich transient. It works best at the edge of chaos — the boundary between order and turbulence — but finding that regime usually means hand-tuning. SC-NeuroCore's reservoir places itself there analytically, from mean-field theory, with no hyperparameters.

$$W_c = \frac{\theta}{2\,\beta\,N}, \qquad \text{at } W=W_c:\ \text{≈ half the neurons active — the critical regime}$$

where $\theta$ is the threshold, $\beta$ the leak and $N$ the neuron count. Construct AutoCriticalReservoir(n_neurons=1000) and it is already critical.

regime  active fraction

Illustrative: below $W_c$ activity dies out (ordered), above it saturates (chaotic), and at $W_c$ it self-sustains near half-active — the regime with the longest memory and richest computation.

Spiking transformers research

Transformers rule sequence modelling, but their self-attention is quadratic in sequence length and multiply-heavy — expensive on an edge budget. Spiking transformers (Spikformer-style blocks, and 3D spatial variants) replace the dense softmax attention with sparse, event-driven attention on the spiking substrate, trading a little expressivity for a large drop in switching.

$$\text{Attention}(Q,K,V) = \operatorname{softmax}\!\Big(\tfrac{QK^\top}{\sqrt{d_k}}\Big)V \;\xrightarrow{\text{spiking}}\; \text{sparse spike-domain attention}$$
Deeper: why attention wants to be sparse
Dense attention pays for every query–key pair whether or not it matters; in the spiking form the interactions are carried by spikes, so the work tracks activity rather than sequence length squared — the same event-driven economy that motivates the rest of the site, applied to the architecture that currently dominates machine learning. It is early, but it is the natural place for neuromorphic efficiency to meet modern models.
Photonic compilation early

The furthest-out direction: computing with light instead of electrons. SC-NeuroCore carries an emerging silicon-photonic path — a photonic compiler and layout export, a bitstream-to-optical conversion, emitter and waveguide models, and an electromagnetic (FDTD) solver for crosstalk. It is genuinely early and code-led rather than documented as a product, but it points at where a stochastic bitstream — already a stream of discrete events — might one day be carried by photons.

Deeper: bitstreams like light
A stochastic bitstream is a sequence of on/off events, which is a comfortable fit for an optical channel that is naturally on/off; a photonic AND or a weighting could in principle be a passive optical element rather than a switching transistor, sidestepping the electrical switching energy entirely. The photonic modules here model that path — compilation, emission, propagation and crosstalk — at a research maturity, flagged as such rather than dressed up.
Bio-hybrid & wetware research

The most exploratory corner of the toolkit reaches past silicon toward biology — storing bits in DNA, letting simulated gene expression modulate a network, and reading real single-cell transcriptomes. These modules are code-led and carried at research maturity; they are here because they run and are grounded in the literature, not because they are products.

DNA data storage. A bitstream is a base-4 number in disguise: two bits pick one of four nucleotides. The encoder maps every pair of bits to a base — 00→A, 01→C, 10→G, 11→T — so any bitstream becomes a DNA string at exactly two bits per base, with an optional per-base mutation model on read-back. Generate a stream and watch it become DNA and come back:

bitstream
DNA (2 bits / base)
decoded back
bases   bit errors after round-trip

Raise the mutation rate and errors appear on read-back — the mismatched bits are highlighted. Real DNA storage adds error-correcting codes on top of exactly this base mapping; this module models the mapping and the noisy channel, not the wet-lab chemistry.

The other bio-hybrid modules

Genetic regulatory layer. A slow, homeostatic modulation loop: neural spikes drive a simulated protein whose level follows a production–decay balance, $\dot P = \alpha\,s - \beta\,P$, and that protein raises the neuron's firing threshold. It is a biologically-motivated way to give a network a slow adaptive variable on top of its fast spiking dynamics.

Neuromodulation. A global chemical context — dopamine, serotonin, noradrenaline — updated by reward and stress signals, which then re-parameterises the neurons: dopamine lowers thresholds (excitation), serotonin damps noise (stabilisation), noradrenaline raises it (exploration). It is the “mood” layer that shifts a whole network's regime without rewiring it.

Transcriptomic interfaces. Beyond metaphor, the toolkit implements the core rank-value encoding used by single-cell foundation models — genes ranked by expression, weighted by inverse corpus frequency — with interfaces mirroring two published models (Geneformer, Theodoris et al. 2023, Nature 619; and scKG-BERT, Li et al. 2025, Genome Biology 26:402). These are core-algorithm implementations and interfaces, not the full pretrained networks.

Honestly labelled

These are shown because they are real and interesting, not because they are finished. Reservoir auto-criticality is the most mature; spiking transformers are an active research architecture; photonic compilation and the bio-hybrid modules are early and code-led. They live under "Labs" precisely so the production parts of the toolkit are not diluted — the boundary is the point.