The “Bayesian Brain”
In his article “Are Brains Bayesian?,” John Horgan discusses the power and limitations of Bayes’ Rule as a model of human decision-making. In this course, we leveraged Bayes’ Rule to determine conditional probabilities, or the probability of an event, A, occurring, given that another event, B, occurs. Bayes’ rule provides a succinct quantification of decision-making when the decision-maker has received some other private information or has observed certain other decisions. The later case was particularly emphasized in class, wherein Bayes’ Rule was used to support a general model of cascading decisions. In this way, the course content implicitly modeled human cognition with Bayes’ Rule. Horgan’s article explores the appropriateness of this choice, questioning whether it is the reality of the human decision-making process or merely an adequate approximation.
Horgan’s article explores both sides of this debate. First, he presents the position of MIT’s Joshua Tenenbaum, who argues in favor the “Bayesian brain.” The crux of Tenenbaum’s argument comes from the idea that Bayesian analysis allows us to “glean knowledge even from sparse, ambiguous data” (Horgan). This is aligned with the course’s discussion of cascades, in which individuals were often able to make powerful, optimized decisions with knowledge of only a few prior decisions. Bayesian analysis provided a clear indication of whether, given the previous decisions in a sequence, an individual should follow or disregard his/her private information. Tenenbaum argues that, in Bayesian fashion, the human mind demonstrates a profound ability to “jump from particular facts to generalizations” (Horgan). Interestingly, Tenenbaum posits that this is a central component of how infants are able to recognize faces, perceive emotion, and even gauge the stability or instability of a tower of blocks. In an attempt to “reverse engineer” the human mind and replicate its performance through computers, Tenenbaum explains that Bayesian approaches are superior to even the “much touTed ‘deep learning’” method (Horgan). Ultimately, a Bayesian model of the mind “captures our capacity for action, imagination, explanation, and creative generalization” (Horgan).
Horgan also presents the contrasting option of Jeffrey Bowers, who claims that, given the appropriate adjustment of a priori knowledge, Bayesian models “can replicate virtually any cognitive task” (Horgan). Bowers argues that while a Bayesian model may adequately approximate human cognition for computational purposes, this is merely a result of the flexibility of Bayesian analysis, rather than a clear indication that the mind follows a Bayesian approach. Bowers raises the interesting point that the Bayesian claim assumes that the brain “employs highly efficient, … ‘optimal’ methods for carrying out cognitive tasks” (Horgan). However, this contradicts the Darwinian perspective that our minds are designed to be “good enough” rather than optimal in the face of natural selection. Moreover, Bowers counters Tenenbaum’s computational argument by explaining that other information-processing models, such as neural networks, can replicate the results of Bayesian models.
Horgan ultimately sides with Bowers’ “non-Bayesian brain” position, emphasizing that Bayesian models provide a valuable approximation of the results of cognition, but cannot be assumed to reflect the realities of the overall cognitive process. In this way, Horgan’s conclusion provides a more profound understanding of the general cascade model demonstrated in class. While Bayes’ rule provides an adequate model of basic decision-making within the relatively limited decision-space discussed in class, it does not fully capture the highly multidimensional decisions that our minds make throughout the day. Moreover, Horgan’s discussion emphasizes that while these basic network models are highly valuable approximations of decision-making, it is impossible for them to capture the fully complexity of the true decision-making process. However, given the recurring notion of conditional probability in this course and in the study of perception, learning, and reasoning, even an approximate model provides valuable intuition into human cognition.
Link: https://blogs.scientificamerican.com/cross-check/are-brains-bayesian/