The Architecture of Neuro-Symbolic Computers

Objectives

Neuro-Symbolic Abstract Machines (NSAMs)

Through the introduction of Neuro-Symbolic Abstract Machines (NSAMs), we aim to create neuro-symbolic artificial intelligence systems that can match the human ability to:

• Reason symbolically.
• Learn composable skills.
• Satisfy safety constraints.

NSAMs are neural networks that are structurally equivalent to programming language interpreters. We expect that NSAMs will improve sample efficiency, interpretability, and task performance.

At present, we focus on creating NSAM architectures within the logic/relational and functional programming paradigms.

Programming Language of Thought

As programming language theorists and machine learning researchers, we take up the task of formally specifying the semantics of, and implementing, the programming language of thought.

One imagines a hierarchy of “executive” programs which function to analyze macrotasks into microtasks. Such programs may “call” both one another and lower-level problem-solving routines, though the extent of such cross-referencing is limited by the ingenuity of the program and, of course, the overall computational capacity of the machine.

—Jerry Fodor in The Language of Thought (1975)

Collaborators

• Matthew Retchin (Harvard University)
• Rafaello Sanna (Harvard University)
• Will Byrd (University of Alabama-Birmingham)
• Nada Amin (Harvard University)