Is the Brain Bayesian?

Friday, December 4 – Saturday, December 5, 2015

Kimmel Center and Hemmerdinger Hall, New York University
BayesGears
REGISTER HERE

On December 4-5, 2015, the NYU Center for Mind, Brain, and Consciousness will host a conference on “Is the Brain Bayesian?”.

Bayesian theories have attracted enormous attention in the cognitive sciences in recent years. According to these theories, the mind assigns probabilities to hypotheses and updates them according to standard probabilistic rules of inference. Bayesian theories have been applied to the study of perception, learning, memory, reasoning, language, decision making, and many other domains. Bayesian approaches have also become increasingly popular in neuroscience, and a number of potential neurobiological mechanisms have been proposed.

At the same time, Bayesian theories raise many foundational questions, the answers to which have been controversial: Does the brain actually use Bayesian rules? Or are they merely approximate descriptions of behavior? How well can Bayesian theories accommodate irrationality in cognition? Do they require an implausibly uniform view of the mind? Are Bayesian theories near-trivial due to their many degrees of freedom? What are their implications for the relationship between perception, cognition, rationality, and consciousness?

All of these questions and more will be discussed at the conference. The conference will bring together both scientists and philosophers, and both proponents and opponents of Bayesian approaches, to discuss and debate a number of central issues.

Speakers and panelists will include:

Hilary Barth (Wesleyan, Psychology), Jeffrey Bowers (Bristol, Psychology), David Danks (Carnegie Mellon, Philosophy, Psychology), Ernest Davis (NYU, Computer Science), Karl Friston (University College London, Institute of Neurology), Wei Ji Ma (NYU, Neural Science, Psychology), Laurence Maloney (NYU, Psychology), Eric Mandelbaum (CUNY, Philosophy), Gary Marcus (NYU, Psychology), John Morrison (Barnard/Columbia, Philosophy), Nico Orlandi (UC Santa Cruz, Philosophy), Michael Rescorla (UC Santa Barbara, Philosophy), Laura Schulz (MIT, Brain and Cognitive Sciences), Susanna Siegel (Harvard, Philosophy), Eero Simoncelli (NYU, Neural Science, Mathematics, Psychology), Joshua Tenenbaum (MIT, Brain and Cognitive Sciences) and others

The conference sessions will run from 9:30am to 6pm on Friday and Saturday December 4-5. Friday sessions will be in Kimmel Center 914 (60 Washington Square South) and Saturday sessions will be in Hemmerdinger Hall in the Silver Center (100 Washington Square East). Conference registration and coffee will begin at 9am both days. A full schedule will be circulated closer to the conference date.

Registration is free but required. REGISTER HERE.

 

CONFERENCE PROGRAM:

Friday, December 4: NYU Kimmel Center 914 (60 Washington Square South)

9:00 – 9:30 am: REGISTRATION AND COFFEE

9:30 am: OPENING

9:40 am – 12:00 noonFOUNDATIONS

Joshua Tenenbaum (MIT, Brain and Cognitive Sciences)
What besides Bayes is needed to understand the brain?

Jeffrey Bowers (Bristol, Psychology)
What are Bayesian theories in psychology and neuroscience?

David Danks (Carnegie Mellon, Philosophy, Psychology)
The (multidimensional) commitments and constraints of being “Bayesian”

12:00 – 1:30 pm: LUNCH BREAK

1:30 – 3:45 pm:  RATIONALITY AND OPTIMALITY

Eric Mandelbaum (CUNY, Philosophy)
Why I’m not a Bayesian

Wei Ji Ma (NYU, Neural Science, Psychology)
Stop arguing, start comparing models: the case for a pragmatic, though exhausting, approach to the Bayesian debate

Gary Marcus (NYU, Psychology) and Ernest Davis (NYU, Computer Science)
How strong is the empirical evidence for Bayesian models of cognition?

3:45 – 4:15 pm: COFFEE BREAK

4:15 – 6:00 pm: PREDICTIVE CODING

Karl Friston (University College London, Institute of Neurology)
I am therefore I think

Michael Rescorla (UC Santa Barbara, Philosophy)
Bayesian modeling and predictive coding

6:00 pm – 7:00 pm: RECEPTION

 

Saturday December 5: Hemmerdinger Hall, NYU Silver Center (100 Washington Square East)

9:00 – 9:30 am: REGISTRATION AND COFFEE

9:30 am – 12:00 noonPERCEPTION

Eero Simoncelli (NYU, Neural Science, Mathematics, Psychology)
Representing priors in sensory systems

Nico Orlandi (UC Santa Cruz, Philosophy)
Bayesian perception is ecological perception

John Morrison (Barnard/Columbia, Philosophy)
Perceptual confidence

Comment by Susanna Siegel (Harvard, Philosophy)

12:00 – 1:30 pm: LUNCH BREAK

1:30 – 3:45 pm: DEVELOPMENT AND DECISION

Laura Schulz (MIT, Brain and Cognitive Sciences)
Beyond belief: thinking and valuing in early childhood

Hilary Barth (Wesleyan, Psychology)
Do Bayesian processes shape our biased estimates?

Laurence Maloney (NYU, Psychology)
Mixing memory and desire: Bayesian decision theory as a framework for perception and action

3:45 – 4:15 pm: COFFEE BREAK

4:15 – 6:00 pm:  PANEL DISCUSSION

Panelists TBA

 

CONFERENCE ABSTRACTS

FOUNDATIONS

Joshua Tenenbaum (MIT, Brain and Cognitive Sciences)
What besides Bayes is needed to understand the brain? 

Bayesian principles of inference have laid the foundations for many recent successes in modeling human cognition.  And yet, they are surely far from sufficient to explain how the mind or the brain works.  What else is needed?  I will argue that Bayesian thinking about the mind and brain can have its greatest impact when integrated with sophisticated thinking about *representations* and *algorithms* – specifically, with structured symbolic representations that support compositional, causal mental models, and efficient stochastic algorithms for approximate inference and search.  This integration is the basis for new Bayesian cognitive models based on *probabilistic programs*.  I will give examples of probabilistic programs for modeling scene perception, physical scene understanding, theory of mind, and visual concept learning, and briefly discuss the prospects for understanding how these models might be implemented in neural circuits, real or artificial.

Jeffrey Bowers (Bristol, Psychology)
What are Bayesian theories in psychology and neuroscience?

I will criticize Bayesian theories in psychology and neuroscience in two general ways.   First, I will argue that Bayesian theories are often constructed in a post-hoc manner, what I have called “Bayesian just-so stories”.  Second, I will argue that the theoretical motivation of Bayesian theories is often unclear. According to some theorists, Bayesian theories make claims about how the brain implements near optimal statistical inference. According to others, Bayesian theories make no claims about optimality, and Bayesian theories make no claims about how the brain computes.  The goal of Bayesian theorizing needs to be clarified.

David Danks (Carnegie Mellon, Philosophy, Psychology)
The (multidimensional) commitments and constraints of being “Bayesian”

Most debates about Bayesian models are framed in terms of the three levels introduced by David Marr (computational, algorithmic, and implementation). In this talk, I will first argue that Marr-levels are inadequate to capture the different types of theoretical commitments that one can have about a neural or cognitive theory. Instead of “levels,” we should instead focus on multidimensional commitments and intertheoretic constraints, which lead to a richer, more nuanced understanding of the ways that theories can be related to one another. This perspective reveals that “Bayesian” does not refer to a single set of commitments, and so the import of a theory — or the brain — being “Bayesian” is underdetermined. In particular, a theory being “Bayesian” does not constrain other theories (even those that implement it) to be Bayesian themselves. Thus, there is no univocal answer about whether cognition or the brain, as wholes, are Bayesian. Throughout, I use a concrete case study in which we developed a non-Bayesian theory of causal learning that is nonetheless consistent (in all relevant ways) with a more behavioral Bayesian theory.

 

RATIONALITY AND OPTIMALITY

Eric Mandelbaum (CUNY, Philosophy)
Why I’m not a Bayesian

A Bayesian mind is, at its core, a rational mind. Bayesianism is thus well-suited to predict and explain mental processes that best exemplify our ability to be rational. However, evidence from belief acquisition and change appears to show that we don’t acquire and update information in a Bayesian way. Instead, the principles of belief acquisition and updating seem grounded in maintaining a psychological immune system rather than in approximating a Bayesian processor.

Wei Ji Ma (NYU, Neural Science, Psychology)
Stop arguing, start comparing models: the case for a pragmatic, though exhausting, approach to the Bayesian debate

Many arguments about whether the brain is Bayesian arise solely from a confusion of the terms Bayesian, optimal (relative or absolute), and probabilistic. I will try to define and distinguish these terms (see also Ma, 2012; Ma and Jazayeri, 2014; Maloney and Mamassian, 2009). Claiming that the brain performs Bayesian, optimal, or probabilistic computation carries entirely different burdens of proof. I will present new results from a deceptively simple visual search task for which we can claim with some confidence that human decisions are Bayesian, probabilistic, and close to relatively optimal. I will also demonstrate that such computation can be learned by a neural network through a simple error-based learning rule.

Gary Marcus (NYU, Psychology) and Ernest Davis (NYU, Computer Science)
How strong is the empirical evidence for Bayesian models of cognition?

Should the mind should be viewed as a “near optimal” or “rational” engine of probabilistic inference? Bayesian inference is a reasonable normative theory in situations where the choice of probabilistic model is clear-cut. However, a large body of empirical evidence, due to Kahneman and Tverksy and many others, demonstrates that high-level human cognition often does not follow this normative theory. Moreover, many of the experiments that have been used as support for Bayesian theories of mind involve situations where the choice of probabilistic model is highly indeterminate. In such situations, Bayesianism is inadequate even as a normative theory, and not convincingly supported as a cognitive theory. In this talk, we draw on our 2013 critique in Psychological Science and point to some important additional problems in the way in which Bayesian models have been applied to empirical data.

 

PREDICTIVE CODING

Karl Friston (University College London, Institute of Neurology)
I am therefore I think

This presentation offers an account of embodied exchange with the world that associates unconscious and conscious operations with actively inferring the causes of our sensations. Its agenda is to link formal (mathematical) descriptions of dynamical systems to a description of perception and behaviour in terms of beliefs and goals. The argument has two parts: the first calls on the lawful dynamics of any dynamic system – from a single cell organism to a human brain. These lawful dynamics suggest that the (internal) states of the brain can be interpreted as modelling or predicting the (external) causes of sensory inputs. In other words, if a system exists, its internal states must encode probabilistic beliefs about external states. Heuristically, this means that if I exist (am) then I must have beliefs (think). The second part of the argument is that the only tenable beliefs I can entertain about myself are that I exist. This may seem rather obvious; however, it transpires that this is equivalent to believing that the world – and the way it is sampled – will resolve uncertainty about the causes of our sensations. We will conclude by looking at the epistemic behaviour that emerges under these beliefs, using simulations of how we sample our visual scene with eye movements.

Michael Rescorla (UC Santa Barbara, Philosophy)
Bayesian modeling and predictive coding

Bayesian models of perception and motor control have achieved great explanatory success. Clark, Friston, Hohwy, and others develop the Bayesian approach by arguing that perception and action serve to minimize prediction error. I critically examine the prediction error minimization account. Regarding perception, I question whether prediction error minimization theories have achieved empirical success proportionate to the philosophical attention they are currently receiving. Regarding motor control, I argue that prediction error minimization theories are less explanatorily successful than “optimal control” versions of Bayesian modeling.

 

PERCEPTION

Eero Simoncelli (NYU, Neural Science, Mathematics, Psychology)
Representing priors in sensory systems

I’m aiming to talk about two things (although depending on the time allotted, I may reduce to one).  The first is a general view of “optimal” computations in the brain, partially covered by this book chapter: Optimal estimation in sensory systems.  The second is our recent work on statistically optimal allocation of neural resources in sensory coding: Efficient sensory encoding and Bayesian inference with heterogeneous neural populations.

Nico Orlandi (UC Santa Cruz, Philosophy)
Bayesian perception is ecological perception

There is a certain excitement in vision science concerning the idea of applying the tools of Bayesian decision theory to explain our perceptual capacities. Bayesian models are thought to be needed to explain how the inverse problem of perception is solved, and to rescue a certain constructivist and Kantian way of understanding the perceptual process. Anticlimactically, I argue both that Bayesian outlooks do not constitute good solutions to the inverse problem, and that they are not constructivist in nature. In explaining how visual systems derive a single percept from underdetermined stimulation, orthodox versions of predictive coding accounts encounter a problem. The problem shows that such accounts need to be grounded in Natural Scene Statistics (NSS), an approach that takes seriously the Gibsonian insight that studying perception involves studying the statistical regularities of the environment in which we are situated. Additionally, I argue that predictive coding frameworks postulate structures that hardly rescue a constructivist way of understanding perception. Except for percepts, the posits of Bayesian theory are not representational in nature. Bayesian perceptual inferences are not genuine inferences. They are biased processes that operate over non-representational states.

John Morrison (Barnard/Columbia, Philosophy)
Perceptual confidence

Perceptual Confidence is the view that perceptual experiences assign degrees of confidence.  After introducing and clarifying the view, I will use introspection to motivate it.  I will then argue that it fills a hole in Bayesian theories of perception.

 

DEVELOPMENT AND DECISION

Laura Schulz (MIT, Brain and Cognitive Sciences)
Beyond belief: thinking and valuing in early childhood

The “Bayesian” research program has contributed greatly to our understanding of learning in early childhood.  However, children learn in ways that go beyond simply updating degrees of belief from data; here I’ll provide two examples.  First, children learn through active exploration and hypothesis generation, not just passive evaluation of existing hypotheses. Second, notions of utility are central, both in understanding how children structure their search for hypotheses as well as their intuitive theories of their own and other minds. I will discuss evidence for each of these claims, and the implications for scientific theories of cognition if our intuitive theories treat humans as rational agents who tend to maximize expected utilities.

Hilary Barth (Wesleyan, Psychology)
Do Bayesian processes shape our biased estimates?

A recent theme in the cognitive sciences is the idea that cognitive or neural systems may combine information sources in a Bayesian manner (or some approximation), tracking source variability and using this to guide the weighting of the sources. If broadly true, this is an important principle of cognitive and neural processing. Recent data suggest that at least some findings commonly taken as evidence of these processes may be better explained in terms of simpler models. For example, Bayesian combination features prominently in literature on the role of categories in cognition, where it’s thought to explain estimation biases in numerous and diverse tasks over development. I will discuss data from both children and adults showing that (1) the observed patterns of bias are not necessarily evidence of the Bayesian combination of information sources, as they can be explained in terms of a considerably simpler psychophysical model, and that (2) individual-level data appear to violate predictions of the dominant Bayesian interpretation.

Laurence Maloney (NYU, Psychology)
Mixing memory and desire: Bayesian decision theory as a framework for perception and action

Bayesian decision theory (BDT) is an elegant mathematical framework that allows us to calculate the action maximizing expected gain by integrating current perceptual knowledge with prior, non-sensory knowledge about the likely state of the environment. Humans in certain simple motor tasks come close to matching the maximum performance possible. In other tasks they fail catastrophically. I will discuss possible explanations for such uneven performance. I will also touch on what might be called the Bayesian paradox: the model of biological function provided by BDT has no place in it for perceptual representations or conscious experience, at least in the ways we tend to think of them.