Eros

An intelligence behind its own work

Eros investigates a model whose own learning becomes the foundation for the work it performs.

We built Eros to investigate how learning, memory and action could develop together. Eros works towards a model able to reason through a task, draw on relevant experience and carry its work forward.

This raised a foundational question. How could we recognise meaningful development within the model, and how could useful information remain available as that development continued?

The early research concentrated on establishing those foundations through measured experiments.

A measured foundation for learning

The first learning experiments recorded an improvement on real material held apart from training.

Across the recorded experiment, Eros's prediction error on that material was lower at the final evaluation than at the first. A further recorded run reproduced the measured result.

This established a repeatable result: training changed how the model responded to information it had not trained on.

Training had changed its response to held-aside information in a repeatable way.

That result provided a foundation for investigating how the model's internal development relates to the responses it produces, and how both change as learning proceeds.

Keeping information available

Alongside learning, we investigated information that could be retained within the model while it worked.

Recorded experiments demonstrated exact preservation of part of its internal information during operation. The information could be reconstructed exactly, while the recorded prediction results remained unchanged.

This established a useful distinction for the research: retaining information and deciding how to use it are questions that can be investigated in their own right.

Information could remain available for further investigation without changing the model's existing predictions.

The result opened a basis for exploring when retained information is useful, which situations call for it and how it might contribute to later learning.

Building on the results

The recorded achievements give Eros a foundation for further investigation into learning and retention.

The next questions concern the usefulness of that development: what can be retained, where it can contribute and how it might support more demanding work.

From learning to useful ability

Eros continues to explore how demonstrated learning and faithful retention can support more demanding work.

Eros develops a model that builds on experience, reasons with what becomes useful and maintains continuity as a task progresses.

Learning from information. Preserving what it makes available. Developing the ability to put it to use.