5th Bernstein Sparks workshop: Neural models of decision making in natural inference tasks - from theory to experiment

The brain needs to extract behaviorally relevant information from sensory signals which contain only indirect, incomplete and highly variable information about the world. Beginning with the study of the neural basis of random dot motion discrimination in the late eighties, neural decision making has been studied extensively in neuroscience. However, the models used to reason about decision making do not account for the computational complexity of natural inference tasks such as object recognition, speech recognition or alike.

Based on recent advances in machine learning more and more complex artificial neural networks are developed that become increasingly proficient in mimicking perceptual inference abilities of humans and animals. As a side effect of their popularity in technology, the increasing availability and diversity of high-performing neural network models opens a new door for studying the neural mechanisms of robust decision making.

Important differences between these networks are the presence or absence of feedback connections, the presence or absence of stochasticity, and the diversity of different nonlinear mechanisms. The existence of this diversity mirrors important discussions in neuroscience on the role and effect of feedback signals is in the brain [1], whether the brain represents and computes with probabilities [2], whether feedback signals are essential for performing probabilistic inference in hierarchical models [3], and whether neural stochasticity can be interpreted in terms of sampling [4] or regularization such as dropout [5].

As a joint effort between theoreticians and experimentalists, the goal of this Sparks workshop will be to survey and discuss the role of these mechanisms for robust decision making in artificial and real neural networks and to derive discriminative experimental tests and tools that seem most promising to analyze them.

[1] Gilbert & Li, Nat Rev Neuro 2013   
[2] Pouget et al, Nat Rev Neuro 2013   
[3] Lee & Mumford, JOSA 2003
[4] Fiser et al., TICS 2010
[5] Srivastava et al, JMLR 2014

Confirmed Speakers:



  • Wieland Brendel (University of Tuebingen)
  • David Cox (Harvard University)
  • Jaime de la Rocha (Barcelona)
  • Nicolas Heess (Google DeepMind)
  • Mate Lengyel (University of Cambridge)
  • Wolfgang Maass (TU Graz)
  • Hendrikje Nienborg (University of Tuebingen)
  • Andreas Tolias (BCM Houston)
  • Dan Yamins (MIT)
  • Rich Zemel (University of Toronto)


  • Matthias Bethge (BCCN Tübingen) 
  • Ralf Haefner (University of Rochester)
  • Richard Hahnloser (University of Zürich)
  • Bernstein Coordination Site (BCOS)

Poster/ Contributed Talks: Poster submissions are highly encouraged. We also may have a small number of slots for contributed talks. Deadline for abstract submission is May, 31th.

Registration/Fees: Attendance is limited to 70 participants. Seats will be allocated on first-come-first-served basis. Please register on the registration website until May, 31th

Workshop registration only (incl. lunch, dinner and coffee breaks):

  • For members of the Bernstein Network and the Bernstein Association for Computational Neuroscience: 40 Euro
  • For non-members100 Euro

If you want to become a Bernstein member, please click here.

The registration is valid after receipt of money. After register you receive the bank account details per email.

Spemannstr. 36
72076 Tübingen

Accommodation: A few rooms are reserved for workshop participants (10-12th). If you are interested, please contact Stefanie Wanner.
A list of hotels can be found here.


Thursday (June 11th)    


9:00 - 09:30
09:30 - 10:15
10:15 - 11:00
11:00 - 11:45
12:00 - 13:30
13:30 - 14:15
14:15 - 15:00
15:00 - 16:00
16:00 - 16:45
16:45 - 17:30
17:30 - 18:30
19:00 -            

Matthias Bethge (intro: Understanding biological and artificial neural networks)
Andreas Tolias: Structure and function of cortical microcircuits
- Coffee break -
Nicolas Heess: Learning to perceive and act with neural networks
- Lunch break -
David Cox: What can studying artificial deep networks teach us about the nature of object representations in the brain?
Daniel Yamins: Using computational models to predict neural responses in higher sensory cortex
- Coffee break -
Mate Lengyel: A sampling-based representation of uncertainty: theoretical advances andempirical evidence
Laurence Aitchinson: Probabilistic Synapses
Discussion (Richard Hahnloser)
Dinner + poster session

Friday (June 12th)


09:00 - 09:30
09:30 - 10:15
10:15 - 11:00
11:00 - 11:45
12:00 - 13:30
13:30 - 14:15
14:15 - 15:00
15:00 - 16:00
16:00 - 16:30
16:30 - 17:00
17:00 - 17:30
17:30 - 18:30
19:00 -

Ralf Haefner (intro: decision making)
Jaime de la Rocha: Dynamics of evidence integration in network models of perceptual decision making
- Coffee break -
Hendrikje Nienborg: Dissecting decision-related activity of sensory neurons to isolate a top-down component
- Lunch break -
Wolfgang Maass: Analogies and differences between learning in artificial and biological neural networks
Richard Zemel: Learning generative models of images and words
- Coffee break -
Wieland Brendel: Balancing sensory input and feedback implements approximate back propagation with Hebbian-type plasticity
Peter Diehl: Cortical-style learning and Inference
Jakob Jordan: Deterministic neural networks as sources of uncorrelated noise for probabilistic computations
Discussion (Richard Hahnloser)
BBQ + wrap-up

 Expected outcomes of this workshop will consist of:

  • Cross-inspiration between theoretical insights on robust decision making in artificial neural networks and systems neuroscience.
  • New specific ideas on the use of stochasticity and feedback in artificial neural networks.
  • New neurophysiological and psychophysical predictions based on potential algorithms that the brain might employ for robust decision making.
  • Highlight analysis techniques for high - dimensional data that make full use of modern population recordings and genetic tools.

Contact: Stefanie Wanner