[Paper Review] Innateness, AlphaZero, and Artificial Intelligence

Gary Marcus|arXiv (Cornell University)|Jan 17, 2018

Reinforcement Learning in Robotics32 references98 citations

TL;DR

The paper critiques the claim that AlphaGo/AlphaZero achieve superhuman performance from tabula rasa learning and argues for greater consideration of innateness in AI, outlining potential notions of innate structure.

ABSTRACT

The concept of innateness is rarely discussed in the context of artificial intelligence. When it is discussed, or hinted at, it is often the context of trying to reduce the amount of innate machinery in a given system. In this paper, I consider as a test case a recent series of papers by Silver et al (Silver et al., 2017a) on AlphaGo and its successors that have been presented as an argument that a "even in the most challenging of domains: it is possible to train to superhuman level, without human examples or guidance", "starting tabula rasa." I argue that these claims are overstated, for multiple reasons. I close by arguing that artificial intelligence needs greater attention to innateness, and I point to some proposals about what that innateness might look like.

Motivation & Objective

• Assess how innateness is discussed in AI research and why it matters for general intelligence.
• Evaluate Silver et al.'s AlphaGo/AlphaZero claims of learning without human data or guidance.
• Argue for incorporating innate structures or priors in AI design and outline potential approaches.

Proposed method

• Critical analysis of AlphaGo/AlphaZero literature and the notion of tabula rasa learning.
• Logical appraisal of the limits of current systems in achieving robust general intelligence.
• propositions outlining what innateness could look like in AI systems.

Experimental results

Research questions

• RQ1 Do AlphaZero-style systems genuinely learn from a tabula rasa starting point in the strongest sense claimed by Silver et al.?
• RQ2 What are the limitations of current deep reinforcement learning approaches that rely on minimal or no human guidance?
• RQ3 How might innate priors or structures be defined and incorporated into AI to enhance learning and generalization?
• RQ4 What concrete proposals exist for introducing innateness into artificial intelligence?

Key findings

• The claims that AlphaGo/AlphaZero reach superhuman performance without human examples or guidance are overstated according to the author.
• There is a need for greater attention to innateness in AI research and development.
• The author outlines proposals about what innateness might look like in artificial systems.