Sander Bohte

Full Name
Prof.dr. S.M. Bohte
Function(s)
Coordinator Partnerships & Transfer, Professor - Universiteit van Amsterdam, Scientific Staff Member, Chair Works Council
Email
S.M.Bohte@cwi.nl
Telephone
+31 20 592 4074
Room
L135
Department(s)
Machine Learning, Works council
Homepage
https://homepages.cwi.nl/~sbohte/

Biography

I develop computational models to understand information processing in neural networks. I strongly believe that neural networks and computational neuroscience models should "compute"; the challenge is to develop insights from neuroscience into usefully computing neural networks, and to bring machine learning insights into models of how neurons in the brain compute. I particularly focused on continuous-time information processing, where time is an explicit dimension of the problem domain. This includes network of spiking neurons, predictive coding, interactive neural cognition, supervised neural learning, and deep reinforcement learning methods.

Research

  • A key research interest is work on neural adaptation and predictive coding for optimal spiking information processing. An example is the notion of multiplicative adaptation for adaptive spike coding, which allows spiking neurons to efficiently encode analog signals over vastly different and rapidly changing dynamic ranges. Current work focuses framing such adaptation in terms of predictive coding, and applying this paradigm to standard learning theory.
  • We also work on biologically plausible policy gradient reinforcement learning of working memory, for so called Semi-Markov Decision Processes. In AuGMent, we show how synaptic tags combined with integrating neurons allow neural networks to learn sequences of tasks, closely mimicking the way monkeys learn these tasks. Recent work has shown how we can formulate learning and processing in RNNs like AuGMent in continuous-time, and implement this in spiking neural networks.
  • Other research relates to neural models of early vision and audition, where the dynamic properties of real neurons are crucial to understanding the relationship between spiking and neural information processing.
  • Active collaborations involve Shirin Dora and Cyriel Pennartz at SILS, UvA; Catherin Wacogne, Jonathan Williford and Pieter Roelfsema, at NIN Amsterdam; Steven Scholte at B&C, UvA and Margriet van Gendt, at LUMC Leiden.
  • In more applied machine learning efforts, we work on super-resolution for CT-scan reconstruction (with Joost Batenburg in the CWI Computational Imaging group) and the use of deep neural network models in finance (with Kees Oosterlee in the CWI Stochastichs group).

Talented students are always welcome to come and do their MSc-thesis work at CWI on Bio-inspired Deep Learning, either based on their own proposed ideas, or based some ready-made potential MSc thesis projects. In particular for projects that link (spiking) neural networks to biology, and also on work that explores (probabilistic) learning in spiking neural networks (SNNs), for example in connection with reservoir computing, or in projects on efficiently simulating large-scale SNNs, with for example GPUs. Feel free to contact me for more information.

Publications

Current projects with external funding

  • Deep Spiking Vision: Better, Faster, Cheaper (DEVIS)
  • Efficient Deep Learning Platforms (eDLP)

Professional activities

  • Organizer: Systems Neuroscience Modeling Symposium - [SNMS]
  • Editor: モFrontiers in Neuroscience and Robotics and AIヤ, Frontiers
  • Editor: IEEE Trans. on Neural Networks and Learning Systems

Grants

  • NWO NAI ᅠ(PI) grant for appointment of PhD-student on Deep Spiking Vision (), 2014. (2014)
  • NWO NAI ᅠ(co-PI) grant for appointment of PhD student on Reward-basedlearning of Subroutines by Neural networks (2014)
  • NWO NAI ᅠ(advisor) grant for appointment of PhD student on Deep Learning for Robust Robot Control (2014)
  • NWO NAI ᅠ(advisor) grant for appointment of PhD student on Decentralized UAV Control (2014)
  • UvA ABC ᅠ(co-PI) grant for appointment of PhD-student on Plausible (2014)
  • ERCIM - ERCIM ᅠPlausible models of continuous time neural reinforcement (2013)
  • NWO IDEAS ᅠ - NWO IDEAS programme proposal Flexible Robot Brains selected for (2013)
  • Veni Innovational Research Grant NWO - Scalable Reinforcement Learning in Spiking Neural Networks (2004)