Synapse Runtime

Bounded live learning

Keeps learning after deployment without running away, and without overwriting what it already knows.

A kilobyte-scale neural network that updates its own weights on the device, with no cloud, and doesn’t diverge. Edge AI ships inference. Trustworthy on-device learning is the missing piece.

Licensing & pilot inquiries hello@synapseruntime.ai

Don’t just take our word for it. Ask about a pilot and run it on your own hardware.

BOUND PLANT FLIP
Bounded vs. unbounded after the world changes.

The claim

Nearly everything ships frozen for one reason: an updating model can diverge, and divergence can’t be certified or left unattended on a live asset. Take that risk off the table and on-device learning stops being a liability.

  • 01

    Bounded while learning

    Weights, not just activations, stay in envelope under continuous online updates, and behavior returns after the internal state is corrupted.

  • 02

    Holds what it already learned

    The catastrophic-forgetting failure mode doesn’t arise. It doesn’t have to shed old skills to keep learning. The only limit is neuron count: neither an architecture tax nor a price of staying live.

  • 03

    Microcontroller scale, no cloud

    Kilobytes and microseconds per decision on a single CPU thread. No GPU or floating-point unit required. It learns where the data already is, so no data ever leaves the device: nothing to transfer, disclose, or breach.

Demos

Measured results

Beyond the guarantee

  • 04

    Learning is opt-in

    Ship it frozen and switch learning on when you’re ready. Switched on, it learns live and label-free from what happens next on the device, and stays bounded by construction, so it can’t run away.

  • 05

    Re-authorable in the field

    Provision spare inputs at ship time, then bind a new signal or teach a new objective later with no firmware redeploy and the same stability.

  • 06

    Irregular streams welcome

    Can learn from duty-cycled, event-driven sensors where the gap between samples is itself information, a case a time-blind learner can’t handle at all.

What it isn’t

This is not gradient clipping, EWC, replay, or test-time adaptation.

Weights bounded while they learn, on the device, without the continual-learning toolbox that buys stability by giving up plasticity. What this adds is the adaptation layer for when the world changes after deployment.

Available for

  • Licensing

    A horizontal bounded-learning license for edge silicon and MCU toolchains: a learning workload for chips that ship inference-only today.

  • Proof of concept

    Embed the runtime and prove it in whatever you’re building. See it run on your own hardware and in your own software.

  • Strategic partnership

    Longer-term structures with the right counterpart: exclusivity, assignment, or an inventor-attached arrangement. Patent pending.

Read on