Applied research

Problems that fit

• Exact procedures from few examples. Tasks where the answer must stay correct outside the range of the data, and a small rule is plausible.
• Control with very small models. Controllers with tens to hundreds of parameters, where memory, power or interpretability matter more than peak performance.
• Optimising non-differentiable systems. Simulators, discrete structures or long recurrent loops where gradients are unavailable or unreliable, and evolutionary or other black-box search is the realistic option.
• Parallel simulation. Moving cellular-automaton or population-based workloads onto GPUs.

How we work

1. State a question that the project can answer either way.
2. Run a conventional baseline first. If a standard method solves the problem, that is the result.
3. Publish code and report negative results.

Our current methods are research tools, not products. They are slower to train than standard neural networks and their advantages are still being measured; see the limitations in Emergent Models.

Contact

Describe the problem, the data, and what a successful outcome would look like: info@emergentcomputing.org.