From single cells to cognition in software and hardware

Members: Aleksandrs Ecins, Ashley Kleinhans, Chris Eliasmith, Daniel B. Fasnacht, Francisco Barranco, Gert Cauwenberghs, Garrick Orchard, John Harris, Janelle Szary, Jonathan Tapson, Mounya Elhilali, Michael Pfeiffer, Ryad Benjamin Benosman, Sergio Davies, Shih-Chii Liu, Timmer Horiuchi, Tobi Delbruck, Troy Lau, Terry Stewart, Andre van Schaik

- Organized by Kwabena Boahen & Chris Eliasmith

Kwabena Boahen boahen@… 26-Jun 16-Jul
Chris Eliasmith celiasmith@… 26-Jun 16-Jul
'Eugene Izhikevich' Eugene.Izhikevich@… 6-Jul 8-Jul
'Bill Softky' 10-Jul 16-Jul
Bartlett Mel mel@… 6-Jul 9-Jul
'Eric Shea-Brown' etsb@… 4-Jul 7-Jul
'Matthijs van de Meer' mvdm@… 2-Jul 7-Jul
'Bruce Bobier' bbobier@… 9-Jul 16-Jul
'Yan Wu' yan.wu@… 26-Jun 16-Jul
Emmett McQuinn emmett@… 26-Jun 12-Jul
Francesco Galluppi francesco.galluppi@… 26-Jun 16-Jul

Page contents: Goals | Possible Projects | For Sure Projects | Software

Focus and goals

The purpose of this workshop is to bring together two methods for large-scale neural circuit construction. One is the Neural Engineering Framework (NEF), a formal method for mapping control-theoretic algorithms on to the neural connections between a highly heterogeneous population of spiking neurons. This method is embodied in Nengo, a freely available software package for neural simulation developed at the University of Waterloo. The other is the Neuromorphic Computational Fabrics (NCF), flexible hardware platforms that emulate spiking neural networks using arrays of silicon neurons in field- programmable mixed-analog-digital chips. This method is embodied in Neurogrid (from which the wiki name ng11 derives), a hardware platform with a million silicon neurons developed at Stanford.

The benefit of combining these methods is that the NEF models do not need to be learned, and so can be constructed rapidly and with reasonably predictable behavior, and Neurogrid can run a model with up to a million neurons in real-time, which is not possible with current standard hardware.

This synergy will allow the construction and testing of large-scale spiking networks for a wide range of behaviors. Students will be introduced to both methods, and techniques for combining them. Students will be able to greatly expand their own projects, or choose one of the provided options. These options range from simple networks found in nearly all vertebrates (e.g. the neural integrator for controlling eye position), to sophisticated networks for performing language-based cognitive tasks, such as solving Ravenʼs Progressive Matrices (a test of human fluid intelligence).

This workgroup will bring together experts in the field of engineered spiking neural networks to tutor the students, and to collaborate on extending NEFʼs theoretical foundations and finding new ways of exploiting NCFʼs computational power.

Possible Projects (difficulty in bold)

  1. Effect of diffusion on Neurogrid: Collecting and analyzing Neurogrid data to determine effect of employing diffuser to decrease number of connections. 2
  2. Working memory: Allows the construction of spiking networks that can be compared directly to recorded single cell data from monkeys on working memory tasks. An extension to serial working memory that compares well with human cognitive data will be considered, time permitting. 1
  3. Basal ganglia: Implement a spiking network model that includes the major components of the basal ganglia. Performs simple and complex cognitive or motor action selection, and forms the core of several cognitive models.
  4. Real-time linguistic induction: Rapid symbol-like variable creation is necessary for finding patterns and new linguistic structures, and has been proposed as a major challenge to all neural cognitive systems. This project allows the construction of a large-scale spiking network that meets this challenge. Application to the Ravenʼs Progressive Matrices (a fluid intelligence test) will be explored. 3
  5. Human-sized associative memory: Scaling of associative memory to handle human- sized vocabularies (about 60,000 lexical items) in a spiking network is crucial for many cognitive applications. This project explores the construction and performance of such a network. 1
  6. Hippocampus: Reproduce place cells. Encode episodic memories as well as place in a hippocampal code. Implement a dynamical model that keeps track of position and environment. 2
  7. Object recognition: Use a learned RBM network as the basis for a fully spiking neural model. Reproduce realistic single cell tuning curves and perform object recognition on the MNIST human handwriting database. 1-3
  8. Attention: Construct a hierarchical attention model that routes information through the hierarchy depending on top-down attentional demands. Models the standard visual hierarchy and reproduces psychological, psychophysical, and single cell attentional effects.
  9. Decision making: Use motion camera with integrator decision model to implement a state-of-the-art decision making circuit, that can integrate stimulus information. 3
  10. Multi-modal fusion: Models of recognition based on multimodal input. OR Multimodal input to disambiguate other mode. 2
  11. Tic Tac Toe: Just what it says.

For Sure Projects

If you're project isn't listed here, we don't know you're doing it. Let us know and we'll add it to the list.

  1. Head direction, path integration, hippocampus with a robot [Results]
  2. Nengo networks in Neurogrid hardware [Results]
  3. Nengo networks in SpiNNaker hardware [Results]
  4. Early visual areas driven by motion camera [Results]
  5. Lamprey model driving a robot [Results]
  6. SVMs in neurons [Results]
  7. Auditory visual fusion [Results]
  8. Working memory for decisions [Results]
  9. Tic Tac Toe [Results]
  10. Magic squares [Results]


Software is available on USB keys in the labs. You can also download it here by clicking on the 'Nengo-XXXX' and '(all files in zip)' from  http://ctnsrv.uwaterloo.ca:8080/hudson/job/Nengo/lastSuccessfulBuild/artifact/

There are some Nengo tutorials to practice below. They cover examples of representation, computation, basal ganglia control, etc.

For using scripting in Nengo, see  http://www.arts.uwaterloo.ca/~cnrglab/?q=node/616