Reductive Hypotheses for a Computational Theory of Mind

Speaker:
Rich Sutton

Abstract:

Certain hypotheses about how to think about mind are deserving of a special status, not unlike that of a null hypothesis in an empirical science, because of their simplicity and generality. Other hypotheses must bear the burden of showing a need for their additional complexity. I present four hypotheses that may be deserving of this special status: 1) that scalar rewards are a sufficient explanation of goals and purposes, 2) that predictions of reward are a sufficient explanation of basic learning phenomena and emotional phenomena, 3) that internal simulation is a sufficient explanation of animal reason and planning, and 4) that predictions about experience are a sufficient explanation of knowledge and understanding. Such hypotheses are rarely proved or disproved; they come to be accepted or rejected by a field over a long period of time. Nevertheless, it is useful to explicitly identify them, particularly when fields come together as they are now in the computational and biological sciences.