Inference of other’s internal neural models from active observation

Abstract

Recently, there have been several attempts to replicate theory of mind, which explains how humans infer the mental states of other people using multiple sensory input, with artificial systems. One example of this is a robot that observes the behavior of other artificial systems and infers their internal models, mapping sensory inputs to the actuator’s control signals. In this paper, we present the internal model as an artificial neural network, similar to biological systems. During inference, an observer can use an active incremental learning algorithm to guess an actor’s internal neural model. This could significantly reduce the effort needed to guess other people’s internal models. We apply an algorithm to the actor–observer robot scenarios with/without prior knowledge of the internal models. To validate our approach, we use a physics-based simulator with virtual robots. A series of experiments reveal that the observer robot can construct an “other’s self-model”, validating the possibility that a neural-based approach can be used as a platform for learning cognitive functions.

Introduction

Robots can represent a simplified model of human behavior, whereby the robot senses its environment and reacts to various input signals. The robot’s ‘brain’ controls its body in response to the input signals using artificial neural networks. The topology and weights of the neural network characterize the behavioral properties of the robot. Recently, several investigations have used robots in order to gain insight into human cognition by creating a simplified analogous problem (Bongard et al., 2006, Webb, 2001, Floreano and Keller, 2010). Bongard et al. built a starfish robot; however, it was unaware of its own body shape (Bongard et al., 2006). Using an estimation–exploration algorithm (EEA) (Bongard and Lipson, 2007), the robot was able to successfully create a self-model of its body shape using an iterative estimation and exploration procedure. In the estimation step, the robot searched multiple candidates to determine its body shape. Subsequently, in the exploration step, the algorithm determined the actions that most strongly agreed with the multiple candidate body shapes.

Unlike self-modeling, however, theory of mind (ToM) is a high-level cognitive function that models the mental states (beliefs, intents, desires, imagination, knowledge, etc.) of another entity. In robotic studies, robots have demonstrated the ability to mimic the behavior of humans or to decode the intentions of a third party (both human and robot). For example, Scassellati implemented Baron-Cohen’s ToM model for the humanoid robot COG (Scassellati, 2002). Breazeal et al. demonstrated that an animal-like robot could pass the false-belief test widely used to test ToM in young children (Breazeal et al., 2005). Furthermore, Buchsbaum et al. carried out simulations in which one agent attempted to determine another agent’s behavior using rat-like characters (Buchsbaum et al., 2005). In this particular study, the observer exploited his own behavior tree to infer others’ intentions.

However, few reports have described the representation of another entity’s mind as a neural circuit. Revealing an internal neural model based on observations is a challenging task. However, there is great potential for using neural networks as internal models, because it would mimic the underlying mechanisms of human representations in the form of neural connections. Many different definitions of the self and other’s self-representations exist, ranging from symbolic states to complex neural models. For example, Bongard et al. (Bongard et al., 2006) used the morphological structure of a robot as a self-model. The robot had no physical model of itself on which to base an understanding, and attempted to construct models of its body using iterative estimation–exploration steps. Kim and Lipson used a simple feed-forward network to represent the minds of other (Kim and Lipson, 2009a, Kim and Lipson, 2009b, Kim and Lipson, 2009b).

In this paper, we propose the use of active incremental learning to infer the internal neural models of other entities both with and without prior knowledge (Fig. 1). We used two robots, referred to as the actor and the observer. The actor used a neural controller (implemented as an artificial neural network) to control its behavior based on sensory information. The observer monitored the behaviors of the actor and attempted to infer the actor’s internal model from these observations. The observer used the inferred self-model of the actor to predict the actor’s future behavior. In this approach, instead of programming the other’s internal model manually, the observer attempted to predict the other’s self-model interactively. The observer robot started from a single actor trajectory and invited the actor robot to demonstrate additional trajectories, which were then used to infer information about the actor’s self-model using the EEA method (Bongard and Lipson, 2007).

In particular, we tested the impact that prior knowledge had on the actor’s internal model. Initially, we assumed that the actor and observer were the same species and that the observer could use his self-model (neural topology). Therefore, the ToM problem is formulated as the inference of the connection weights given the shared structure. We subsequently assumed that the two robots are different species and that the actor could not use his self-model for the ToM. As a result, the observer needs to search for the architecture of the neural network and the weights simultaneously to infer the other’s self-model. We used a physics-based simulation to run the ToM experiments, which show the potential of this approach given the two experimental conditions.

The rest of this paper is organized as follows. In Section 2 we describe related research, including the research on ToM in robots. In Section 3 we apply the estimation–exploration algorithm for the robotic ToM. Finally, in Section 4, we present our experimental results.