TL;DR
This paper explores Fred Dretske's theory of mental content by applying it to the learning capabilities of a mobile robot. The authors argue that while Dretske's theory is attractive, its strict distinction between representation "authorship" and its categorization of representations are difficult to maintain in the context of autonomous learning systems. They propose that a more nuanced understanding of how learning contributes to content is needed.
What Problem Does This Paper Address?
The paper tackles a fundamental question in the philosophy of mind: how do sensorimotor agents, like human infants or robots, develop "contentful" mental states? Contentful mental states are those that are "about" something in the world. The authors use Fred Dretske's influential theory of mental content as a framework to investigate this question, specifically in the context of a mobile robot that learns from its experiences. Dretske's theory posits that for a representation to be contentful, it must be learned and indicate an external condition. The challenge is to see if this theory holds up when applied to a system that learns autonomously, where the "authorship" of representations becomes ambiguous.
Methods
The core of the paper involves a mobile robot (Pioneer 1) equipped with numerous sensors and a learning system. The robot is programmed with minimal innate structure and learns through unsupervised methods. A key technique is "clustering by dynamics," which groups similar time-series sensor data into "prototypes." These prototypes represent learned "activities" or "experiences" of the robot, such as moving past an object. The authors then attempt to map these learned prototypes to Dretske's criteria for contentful mental states, specifically "Type III learned representational states." This involves analyzing how the robot's learned representations indicate external conditions and how their functions are acquired through learning.
Datasets
Not discussed in the paper. The paper focuses on the robot's internal learning process and its self-generated data rather than external datasets.
Key Insights and Ideas
- Dretske's theory is attractive but faces practical challenges with learning robots: The idea that learned representations are contentful is compelling, but the strict criteria for "authorship" and the clear-cut distinctions Dretske makes are hard to apply when a robot learns autonomously.
- "Authorship" of representations is ambiguous in learning robots: When a robot learns through unsupervised methods, it's difficult to pinpoint who or what is the "author" of a representation. This challenges Dretske's distinction between representations where the function is assigned by an external agent (Type II) and those learned (Type III).
- Not all learned representations are equally contentful: The paper suggests that Dretske's theory doesn't adequately differentiate between various classes of learned representations. Some might be more "contentful" than others, depending on their ability to denote roles and relationships, not just sensory experiences.
- Learning is crucial, but not the sole criterion for content: The authors argue that while learning is essential for developing contentful states, it shouldn't be the only criterion. Content should be explained by how learning contributes to it, rather than simply being a prerequisite.
- The goal is to understand bootstrapping of mental content: The research aims to understand how simple agents can develop complex mental states from basic wants and learned actions, mirroring infant development.
Key Results
The robot, through clustering by dynamics, learned prototypes representing various activities like passing an object, driving towards an object, and bumping into an object. These learned prototypes are argued to be "perceptually-based beliefs" and potentially "Type III learned representational states" according to Dretske's theory. However, the paper highlights difficulties in precisely applying Dretske's criteria, particularly regarding the "authorship" of these learned representations.
Limitations and Future Work
- Ambiguity in "authorship": The authors acknowledge that their role in designing the robot's learning algorithms and setting up experimental trials makes the "authorship" of learned representations unclear, posing a challenge to Dretske's Type III criteria.
- Dretske's strict learning criterion: The paper questions whether Dretske's insistence on learning as the sole criterion for contentful states is too restrictive, potentially excluding genuinely contentful states.
- Future Work: The authors suggest that a better theory of mental content should explain how states become about something, rather than just using learning as a prerequisite. They propose focusing on how agents develop representations that denote roles of participants in activities, leading to richer ontologies and supporting more complex mental activities like language. They also aim to explain how mental content can be "bootstrapped" from a minimal set of innate structures.
Summary updated