Think of the people you know. How do you know them? Not “how did you meet them” I mean: how is it that you can know (understand, anticipate) them. If you were to describe how you know your mother, father, boyfriend, aunt how would you do it; think about the kinds of words you would use. You might describe how you met them and how long you’ve shared a relationship. But this doesn’t get us any farther along to how it is that you have come to understand them. Think about their motivations, drives, how is it that you can anticipate their feelings and reactions. How is it that you know their mental state?
Researchers at Google are thinking about it. The title of this post refers to a recently published article by Rabinowitz, Perbert, Song et al. whose goals are ambitious from the first sentence:
Theory of mind (ToM; Premack & Woodruff, 1978) broadly refers to humans’ ability to represent the mental states of others, including their desires, beliefs, and intentions. We propose to train a machine to build such models too.
They aren’t building models to understand humans though, not yet. In this paper, their purpose is to understand other machine learning agents. It’s an important problem because before we allow machine learning agents to go off and take care of important tasks (driving, prescription filling, surgery) we should know a little bit about how they think. And right now we can’t explain exactly what’s going on inside these black boxes. The rhetorical question is: How can we release these models into the world, responsibly (legally), when we can’t determine why they made a decision. Of course, we can release these models to start making life or death decisions but we shouldn’t.
Back to the article. These researchers focus on the problem of how an observer (another model) could learn how the model of interest is making decisions:
We design a Theory of Mind neural network — a ToMnet — which uses meta-learning to build models of the agents it encounters, from observations of their behaviour alone. Through this process, it acquires a strong prior model for agents’ behaviour, as well as the ability to bootstrap to richer predictions about agents’ characteristics and mental states using only a small number of behavioural observations. We apply the ToMnet to agents behaving in simple gridworld environments, showing that it learns to model random, algorithmic, and deep reinforcement learning agents from varied populations, and that it passes classic ToM tasks such as the “Sally-Anne” test (Wimmer & Perner, 1983; Baron-Cohen et al., 1985) of recognising that others can hold false beliefs about the world. We argue that this system — which autonomously learns how to model other agents in its world — is an important step forward for developing multi-agent AI systems, for building intermediating technology for machine-human interaction, and for advancing the progress on interpretable AI.
So what do you think? Here’s another perspective on a theory of mind from Reddit: