Discerning Quantity: Numerosity in Two Embodied Machine Learning Agents

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Machine Learning

Background:

• Evaluating artificial intelligence (AI) systems using psychologically inspired methods is increasingly relevant as AI capabilities advance.
• Numerosity, the ability to perceive and process quantities, is a fundamental cognitive capability.
• Previous research has not extensively explored AI models' numerosity in embodied, psychologically inspired experimental settings.

Purpose of the Study:

• To evaluate the numerosity capabilities of two machine learning models: A-Learning and Proximal Policy Optimisation.
• To investigate how embodiment in a virtual environment (Animal-AI) affects the expression of cognitive capabilities in AI.
• To identify potential confounding factors influencing AI cognitive performance.

Main Methods:

• Two machine learning models, A-Learning and Proximal Policy Optimisation, were implemented.
• Models were embodied within the Animal-AI three-dimensional virtual environment.
• A psychologically inspired numerosity experiment was conducted to assess model performance.

Main Results:

• Neither A-Learning nor Proximal Policy Optimisation reliably demonstrated numerosity capabilities.
• Both models exhibited a tendency to overfit to the first policy that yielded rewarding feedback.
• The expression of numerosity was complicated by environmental properties and perceptual processes within the virtual setting.

Conclusions:

• Predicting cognitive capabilities of embodied machine learning models is complex and non-trivial.
• Environmental factors and perceptual mechanisms significantly influence the expression of AI cognitive abilities.
• Future research should prioritize understanding the interplay between embodied AI, environment, and cognitive function, particularly numerosity.