Donders Institute for Brain, Cognition and Behaviour
Zoek in de site...

Thesis defense Mark Blokpoel (Donders Series 195)

5 November 2015
Promotor: Prof. dr. Ivan Toni; Copromotors: dr. I. van Rooij, dr. P.  Haselager ,dr. T. Wareham
Understanding understanding: A computational-level perspective

Imagine two friends in a crowded club with a band playing loud music. From across the crowd the one friend sees the other is trying to communicate with gestures. She points to herself and then puts the fingertips of her hands together to form the shape of a roof. The observer understands the communicative signal and realizes that his friend is going home. Most cognitive theories model this communicative capacity as a form of abductive inference. For example, generating a signal can be modeled as finding the most probable meaning given how one thinks their interlocutor would understand it. Such models, however, invariably fail to answer two important questions. Firstly, according to these models understanding signals may take millions of years to compute. How can we explain that people do not have this issue? Secondly, where do people find the best meaning, especially when they have never seen or produced the signal before?

I answer the first question as follows: The reason many cognitive models are computationally intractable, is because they overgeneralize the set of instances in which people can quickly perform the modeled capacity. Using parameterized complexity analysis one can identify under which conditions the model can be quickly computed. These conditions may then explain why people can, for example, understand communicative signals quickly. In my thesis, I apply this methodology on computational-level models of action understanding, intention recognition and recipient design. The second question is answered by an appeal to analogical reasoning. To explain where people find the best meaning for communicative signals they have never encountered before, I present a computational-level model that—starting from perceptual input and knowledge—can generate sets of meanings by progressive analogies. Such a process can reconceptualize the information available, enabling the understanding of novel signals.