Speaker: Richard Evans (Google DeepMind)
Title: Kant's Cognitive Architecture
Time: 16:00 - 17:30
Location: ILLC Seminar Room F1.15
In this talk, I will describe a machine implementation of assimilative apperception that is based on Kant’s discussion of the “synthetic unity of apperception” in the Critique of Pure Reason. Imagine a machine, equipped with various sensors. Perhaps it has some light sensors, perhaps a microphone, perhaps some touch sensors. The machine receives a stream of sensory input, and must “make sense” of this stream. – But what does it mean, exactly, for a robot to “make sense” of the sensory stream? In machine learning, this is called the “unsupervised learning problem”, and it is still relatively under-explored and misunderstood. I will argue that the answer to this contemporary question lies, hiding in plain sight, in the first half of the First Critique.
Our computer agent assimilates information by constructing a theory that explains the sensory stream and that satisfies four unity conditions. First, the objects must be unified in space. Second, the predicates must be unified in the “space of reasons” by featuring in mutual incompatibility relations. Third, the various propositions must be unified into a “moment” (a consistent and complete cluster of propositions that together represent a single slice of time). The fourth and final unity condition is that moments must be unified into a sequence by causal rules. Apperceiving a sensory sequence means constructing a theory that explains the sequence and satisfies the four unity conditions. This is a form of program synthesis, but it is unsupervised program synthesis. In this talk, I will not focus on the gory technical details of the unsupervised program synthesis system, but on the general architecture, the underlying Kantian principles that inspired the architecture, and on a diverse range of experiments which show the versatility of the approach.