[Article] 'Consciousness' in the machine learning (deep learning) perspective
**The Consciousness Prior**
Yoshua Bengio
Université de Montréal, MILA
September 26, 2017
**Abstract**
A new prior is proposed for representation learning, which canbe combined with other priors in order to help disentangling abstract factorsfrom each other. It is inspired by the phenomenon of conscious-ness seen as theformation of a low-dimensional combination of a few concepts constituting aconscious thought, i.e., consciousness as awareness at a particular timeinstant. This provides a powerful constraint on the representation in that suchlow-dimensional thought vectors can correspond to statements about realitywhich are either true, highly probable, or very useful for taking decisions.The fact that a few elements of the current state can be combined into such apredictive or useful statement is a strong constraint and deviates considerablyfrom the maximum likelihood approaches to modeling data and how states unfoldin the future based on an agent's actions. Instead of making predictions in thesensory (e.g. pixel) space, the consciousness prior allow the agent to makepredictions in the abstract space, with only a few dimensions of that spacebeing involved in each of these predictions. The consciousness prior also makesit natural to map conscious states to natural language utterances or to expressclassical AI knowledge in the form of facts and rules, although the consciousstates may be richer than what can be expressed easily in the form of asentence, a fact or a rule.