Empirical · simulated stock market
Technology
A continual learning architecture, with no pretraining and no retraining cycle.
Inside a deployment
Software that makes repeated decisions on live data usually rests on values someone set once, such as an expected-output curve fixed at commissioning or a capacity figure entered at implementation.
Conditions change, the value stays, and nobody notices the gap.
The CLS learns what such a value should be from what actually happens. The software around it goes on making its decisions, with values that match the asset or person in front of it.
One of our recently published studies tests exactly this. In a simulated stock market, an instance recovered a rule nobody described to it and tracked it as the rule changed underneath it, with no reset and no notice. The report states its method and limitations, including the predictions that failed.
How it learns.
There is no pretraining and no retraining cycle. Nothing is trained centrally and shipped out, so there are no weights to refresh. Each instance carries its own memory and learns at runtime from the entity it is attached to.
In a continual learning architecture, learning is a property of the running system, with no training phase before it or after it. When conditions change, an instance adapts in operation, and nobody schedules a retraining run.
| A trained approach | The CLS | |
|---|---|---|
| Before deployment | A training run | Nothing |
| When conditions change | Scheduled retraining | Adapts at runtime |
| What it keeps | Shared weights | Its own memory, per instance |
| Hardware | Training infrastructure, often GPUs | An ordinary processor |
The same principles run in every environment we have tested. See where.
How it acts.
Perception needs no reward signal, and action needs only a single preference: among the actions available in a situation, prefer the one whose recorded outcomes have been larger.
Learned action selection, built and backtested on equities and tested in simulation. In the tax study, what the system held about each action in each kind of situation was a prediction of what that action yields, and experience corrected it.
In a simulated stock market, an instance recovered a hidden 25 percent tax schedule from the consequences of its own trades. In a pre-registered test, it detected and re-adapted to unannounced rule changes faster than a tuned standard method.
What it needs.
No labeled data. Learning is self-supervised: the next observation grades the previous prediction. A cold start needs a history of a few hundred data points.
The data a deployment already collects is the learning signal. What we need is access to it and someone who knows what the values mean.
Is the data the kind the AI learns from?Where it runs.
On the infrastructure where the data lives, inside a deployment. No data leaves that environment.
The CLS runs on ordinary processors, and the decisions stay with the software around it.
How a deployment is wired: the surrounding software sends operational data to the CLS and gets updated reference values back through an API or webhooks. What the CLS learns stays next to that data. Today, the CLS ships as a Python library with an HTTP runner service.
Talk to our team about a deployment.Wired into existing systems.
Kept where the data lives.
What it costs.
Hundreds of live streams on one CPU server, no GPU.
Each instance runs on ordinary CPU hardware, with no GPU and no training infrastructure to pay for. Instances are independent of one another, so every asset or individual can have one of its own.
FAQs
Continual learning, explained.
What is Wakeline, and what does it research?
Wakeline is an AI research lab. We research continual learning, that is how a system can learn from new observations while it runs. Our Continual Learning System (CLS) was deployed first in energy markets and is now being tested by a select few companies in other industries.
What is continual learning in AI?
Continual learning is an approach to AI in which a system continues to learn as new information becomes available. Wakeline’s research focuses on learning from operational experience, so predictions can be informed by the conditions a system encounters over time.
How does Wakeline’s Continual Learning System learn without labelled data?
Wakeline’s CLS uses self-supervised learning: the next observation grades the previous prediction. The operational data therefore provides the feedback needed for learning, without requiring people to label a training dataset.
How is a deployment of the CLS wired into existing software?
A deployment feeds the operational data stream to the CLS and takes updated reference values back through an API or webhooks, into the existing decision logic. The CLS runs on the infrastructure where the data lives, so what it learns stays alongside that data.
How does Wakeline choose where to deploy next?
We work with a select few companies per industry. In a first conversation we find out whether there is a learning task the CLS fits and what data exists.