Research Roadmap

We’re developing algorithms for Artificial General Intelligence. Unlike most AI, a general intelligence does not require tailoring to specific problems. We’re using the computational properties of human intelligence as inspiration, and building on the best and latest techniques from machine learning.

Memory and Learning
We seek to build a memory system that’s capable of continuous self-learning through interactions with the world.
Unsupervised learning
An essential criteria for learning very difficult real-world problems, for which “ideal” actions are unknowable.
Continuous learning
As an agent explores the world, changes in action policy will lead to distinct changes in the statistics of observations. Therefore, continuous learning without interference or catastrophic forgetting is essential.

  • Sparse coding: Sparse coding provides more powerful and robust representation by distributing the state amongst many attributes in combination.
  • Predictive coding: Well established in neuroscience, it may provide more optimal representations for sequence data.
Sequence learning
We seek to model sequences and changes in the world rather than looking at, for example, classification of static images.
Action selection
We believe that a key ingredient of AGI is an agent’s interaction with the world. This means that action selection is part of the mix.
Reinforcement learning
Without supervision, the only way to guide the selection of actions is reinforcement learning.
Hierarchical planning
Given a sophisticated, deep, hierarchical representation of the world, we believe this same representation should be used to generate action plans.
Selective memory
We consider selective memory to be the biased or filtered storage of features in short and long term memory such as to optimize other goals by removing distraction and enhancing relevant information.
Working memory
We believe that Working Memory can be viewed as a short-term simulation of the causes and effects in a world and how an agent can manipulate them.
Computational Neuroscience
While practical techniques will draw from Machine Learning, we will look to Computational Neuroscience for guidance and inspiration.