Neuroplastic Inference
AI’s third
Big Bang
Today’s most powerful AI systems draw heavily on what they learn during training. When a large language model encounters an unfamiliar problem, it can struggle to find a solution that works. It might brute-force its way through many possibilities at a huge cost, but in science and engineering, checking each solution can require a laboratory experiment, a physical prototype, or weeks of observation. Adapting the model usually means gathering more data and retraining it, which is slow and expensive.
The brain works differently
Neuroplasticity is the brain’s ability to rewire itself in response to experience. Evolution has shaped it to learn from unfamiliar situations and adjust as its goals change. Conscium’s researchers are building that capability into training, so that its models can handle highly complex tasks without retraining. We call this Neuroplastic Inference.
Built for goals that change
Real-world problems are rarely static. Goals change, conditions shift, and there may be no examples showing you what to do next. People deal with this constantly. Nobody learns to avoid driving off a cliff by trying it. We use our understanding of the world to rule out outcomes we have never experienced. The challenge for AI is to solve new problems before it has seen any examples of their solutions.
Today’s language models can struggle when there is no familiar answer to draw on, and no practical way to test thousands of possibilities. A prompt can describe a new goal, but that does not ensure that the model can achieve it. Most models are trained for the world as it exists in their data set. Our models are trained for a world that will change. When it does, they adjust at the moment of use, drawing on existing knowledge and whatever new observations are available, without retraining.
Inside the model
Learning how problems are built
Conventional inference maps inputs to outputs. Neuroplastic Inference uncovers a problem’s underlying structure and generates solutions from it, much as people generalise, improvise, and handle rare cases which are not covered by any of the training data. Trained across a wide range of tasks, our models can pursue new objectives from limited evidence, without another training run. The similarity to the human brain lies in this ability to adapt, rather than in imitating neurons.
Structure
How a family of problems works
Objective
A goal that can change
Solution
Worked out from the problem’s structure, not memorised
Problems that cannot be brute-forced
mRNA and finance
We are applying Neuroplastic Inference first in a couple of focused domains, where its strengths are easiest to verify. One domain is mRNA design, and we will have more to say about this soon. Designing an mRNA sequence means balancing properties such as stability and protein output, and the right balance shifts from one therapy to the next. Another domain is finance, where constraints, risk appetite and market regimes change continuously. A model built for one country’s market can be applied to another’s without retraining.
Beyond the training data
Getting a model to perform well on problems it was never trained on is one of the biggest open questions in deep learning. Because our models learn how a family of problems is structured, they handle new members of that family without retraining. The same approach carries over to a new domain with little data, expensive tests and moving goals, from materials and chemistry to hardware design. A new domain needs a new family of problems to train on, not a new kind of model.
The next step change
From two Big Bangs to a third
Deep learning, and transformer-based large language models each marked a step change in what AI can do. Their arrivals in 2012 and 2017 respectively were the first two “Big Bangs” in AI. AI that adapts to problems and goals it has never encountered before could be the third. Neuroplastic Inference is our contribution to getting there.
2012
Deep learning
2017
Large language models
Next
AI that adapts to problems and goals it has never encountered
Get in touch
Interested in Neuroplastic Inference?
If you’d like to discuss our research, explore a collaboration, or see how Neuroplastic Inference could apply to your domain, get in touch with us.