Making sense of transformer success

Area of Science:

• Philosophy of Artificial Intelligence
• Cognitive Science
• Computational Linguistics

Background:

• The central question in AI philosophy has shifted from 'can machines think?' to 'how can machines think?'.
• Neural Language Models (NLMs), particularly those based on the Transformer architecture, present an explanatory gap regarding their linguistic competence.

Purpose of the Study:

• To epistemologically analyze current explanations for NLMs' linguistic competence.
• To compare explanatory strategies for NLMs with those used in cognitive science.
• To investigate the mechanisms behind in-context learning and explore co-simulation approaches.

Main Methods:

• Analysis of existing explanatory strategies for NLMs.
• Examination of experimental studies on theory of mind, discourse entity tracking, and property induction in NLMs.
• Review of functional analysis in cognitive science and mechanistic explanations of in-context learning (copying algorithm, induction head).

Main Results:

• Explanatory strategies for NLMs are comparable to those in cognitive science.
• The 'copying algorithm' and 'induction head' offer mechanistic explanations for in-context learning.
• Pioneering attempts at co-simulation between NLMs and brain activity are emerging.

Conclusions:

• Current explanations for NLMs' capabilities align with established cognitive science methodologies.
• Mechanistic insights into in-context learning are being developed.
• Co-simulation presents a novel approach for understanding language processing in both AI and humans.