Jan Drugowitsch, Ph.D.

Center for Brain Science
Department of Neurobiology
Harvard Medical School
Goldenson Bldg., Room 420
220 Longwood Avenue
Boston, MA 02115

tel: (617) 432-1772
email: jan_drugowitsch@hms.harvard.edu
Web: http://neuro.hms.harvard.edu/people/faculty/jan-drugowitsch/

Research Interests:

One of the biggest challenges in neuroscience is how to translate between algorithms and their neural implementation. The realization that every-day decision-making is fundamentally probabilistic in nature has led to algorithms based on Bayesian inference and optimal control that could explain a wide variety of observed behavior and its neural correlates. Though surprisingly successful, the computations underlying these algorithms are often intractable, such that it is fallacious to assume that they have direct correlates in the nervous system.
 
My research aims at investigating the boundary between idealized inference and human/animal performance in real-world tasks, with emphasis on low-level tasks such as perceptual decision-making. Even in these tasks, inference is only tractable in some approximate form. Consequently, if we want to understand how organisms perform these tasks, we need to get a handle on the approximations they use. Identifying these approximations is an essential step in linking algorithms to their neural implementation. The envisaged approach is more powerful than proposing ad-hoc heuristics as it uses optimal task solutions as a basis of investigation. Specifically, it rests on the following three pillars:

  1. identify the possibly intractable computational requirements of optimal task solutions, and draw inspiration from machine learning and neurophysiology on which kind of approximations are tractable and feasible,
  2. investigate approximations and assumed neural mechanisms for when they will perform well and under which circumstances they will fail, and
  3. seek for behavioral and/or neural evidence that allows for the rejection of the various hypotheses that emerge from the two previous steps.

We have successfully applied the above approach to questions of how humans and other animals accumulate perceptual evidence over time, and how they trade off the speed of decisions with their accuracy. In particular, we have for increasingly natural settings shown that optimal decision-making behavior is implementable by diffusion models with time-varying decision boundaries. Based on this, observed behavior of humans and monkeys indicated that they seemed to feature a cost for the accumulation of evidence that is rising over time, and that finds its neural correlate in the previously observed urgency signal. We have furthermore demonstrated that, as soon as time pressure plays a role in decision-making with multiple sources of evidence, standard tests of how information is combined across these sources start to fail. Our novel approach based on Bayesian strategies implemented by diffusion models, in contrast, again revealed that human subjects combining visual and vestibular cues did so optimally across time and cues.

My ongoing and future research pushes the boundaries of our knowledge of decision-making based on uncertain information in various interesting directions. First, evidence accumulation over time is frequently associated with neural activity that rises over time to a threshold, and is theoretically linked to diffusion models. Such models implement optimal decision strategies under simple scenarios, but break down once we move to settings that require tracking of higher-dimensional statistics. My aim is to rigorously describe such setting statistically, to investigate the performance loss if organism were to use then-suboptimal diffusion models to form decisions, and to formulate hypotheses for neural mechanisms implementing optimal higher-dimensional strategies, to be scrutinized in the light of empirical findings. Second, findings of grid and place cells spawned hypothesis of their use in navigation and trajectory planning. Most work assumes them to encode location point estimates without considering the possibility of additional representations of uncertainty about this location, even though such representations are essential for efficient behavior in uncertain and ambiguous environments. In collaboration with Chris Harvey, we aim to investigate if such representations of uncertainty exist and are used, what forms they take, how such forms facilitate computation, and how they could be used for navigation, planning, and task-learning. Third, most work on how the nervous system could perform computations with uncertain quantities has led to predictions of neural activity that is overly structured when compared to that observed in animal cortex. The latter is well described by recurrent neural networks that are trained to perform particular tasks. These networks, however, are not equipped to feature useful representations of uncertainty. Thus, my aim is to combine both approaches, and to investigate how such networks could be endowed with probabilistic representations while being trained on tasks that require such representations.

Selected Publications:

Jan Kneissler, Jan Drugowitsch, Karl Friston, and Martin V Butz (2015). Simultaneous learning and filtering without delusions: a Bayes-optimal derivation of combining predictive inference and adaptive filtering. Frontiers in Computational Neuroscience, 9(47).

Jan Drugowitsch, Rubén Moreno-Bote, and Alexandre Pouget (2014). Relationship between belief and performance in perceptual decision-making. PLoS ONE, 9(5): 2014;e96511.

Jan Drugowitsch, Gregory C DeAngelis, Eliana M Klier, Dora E Angelaki, and Alexandre Pouget (2014). Optimal multisensory decision-making in a reaction-time task. eLife 2014; 3:e03005.

Jan Drugowitsch, Rubén Moreno-Bote, Anne K Churchland, Michael N Shadlen, and Alexandre Pouget (2012). The cost of accumulating evidence in perceptual decision making. The Journal of Neuroscience, 13, 3612-3628.

 

 

Page created and maintained by Xaq Pitkow