AI-ON/TheConsciousnessPrior

AI-ON Consciousness Prior

The Consciousness Prior

• Tagline: Consiousness inspired attentional models in machine learning
• Date: December 2017
• Category: Fundamental Research
• Contact(s): William Fedus, Sherjil Ozair, Yoshua Bengio

Project Status:

• Initial exploration of consciousness prior in a purely observational setting.
• Constructing environments to test efficacy of CP in tracking objects, particularly when the task is complicated with high-entropy observations in the pixel space.

Community Links:

• Google Group
• Slack Link

Problem description:

See introduction of project proposal.

Why this problem matters:

See introduction of project proposal.

How to measure success:

See introduction of project proposal.

Datasets:

• Observational datasets will initially be synthetic. We propose a generalization to the Billiards task introduced in Sutskever et al.(2009)

Extracting Data - Preprocessing details:

• Data will initially be synthetic.

Relevant Work:

Below you will find other relevant material to this project. This short list already assumes a familiarity with common deep neural networks (RNNs, CNNs) and training procedures (BPTT).

• Bengio. (2017). The Consciousness Prior
• Dehaene et al. (2017). What is consciousness, and could machines have it?
• Goodfellow et al. (2016). Deep Learning Esp. Ch 15.
• Bahdanau et al. (2014). Neural Machine Translation by Jointly Learning to Align and Translate
• Goodfellow et al. (2014). Generative Adversarial Networks
• Brakel and Bengio. (2017). Learning Independent Features with Adversarial Nets for Non-linear ICA
• Jaderberg et al. (2016). Reinforcement Learning with Unsupervised Auxiliary Tasks
• Grathwohl et al. (2017). Backpropagation Through the Void

Contribute:

To contribute to this project:

1. Sign up for the Slack Channel and Google Group.
2. Please familiarize yourself with the Relevant Work.
3. Create a new branch and then begin work on an open GitHub issue. Coordinate with others on Slack so that work is not duplicated unnecessarily.

The AI-ON process is experimental, but we will establish a review process for new code pull requests. We will seek to establish unit tests as well as performance tests on tasks as new modules are created.

Observational

We propose first investigating an extension to the Billiards task introduced in Sutskever et al. (2009)

Reinforcement Learning

We may simultaneously investigate the consciousness prior in the context of reinforcement learning, where the agent is in influencing the sensory input via control.