Competition 2007: Predicting Single-Neuron Behavior

The announcement for the 2007 competition for predicting single-neuron behavior. 

Submission for this competition is over. This entry is here for reference only.

Challenge

  • Is it possible to predict the timing of every spike that a neuron emits with a precision of 2ms?
  • Is it possible to predict the subthreshold membrane potential with a precision of 2mV for arbitrary input?

A first set of annotated training data and test stimuli have been published by March 16 2007 and the results of the first round of submissions was available for the workshop on Quantitative Neuron Modeling, June 25/26 in Lausanne on the Campus of the EPFL . A new data set for challenges A and B will become available by October 2007.

Important dates (challenge of spring 2007)

  • First data set available by March 16, 2007.
  • Participants must submit their prediction by June 1st, 2007.
  • Winner announced around June 10 .
  • Winning results will be presented at the workshop June 25/26, 2007

Competition and Prizes

The competition is run in several categories, called A,B,C,D. Participants may run in one or several categories
  • 1st prize : given to participants having scored best in at least two categories, for example, category A and C. Winner receives
  • 2nd prize: given to participants having scored best in one of the categories, for example, category D.
Note that there may be 0,1, or 2 first prizes.

Workshop and Competition participation

Results of the workshop will be presented at the Quantitative Neuron Modeling workshop June 25/26, 2007 . However, everybody is welcome to participate in the competition without enrolling in the workshop. The competition is an invitation to compare your results and methods with those of other people in the field.

Poster presentation at the workshop, requires submission of results to the competition in at least one category. Workshop registration is mandatory for those wishing to present a poster.

Data format
Data sets will become available by March 15 2007

Data will be available for parameter optimization (training data, input and output given) and prediction (test data, only input given).

We expect to have at least 4 categories:

  • A: Single-electrode data from cortical neurons under random current injection.
    Stimulation is done with currents of different means and fluctation amplitudes. For each inpu, 4 repetitions of the same stimulus are available so as to estimate the intrinsic reliability of neurons. Several stimulation sets are set apart for the prediction.
    Details of Challenge A
    Link to Data for Challenge A
    Please submit challenge A results to renaud.jolivet@epfl.ch
    matlab code for coincidence factor Gamma
  • B: Single-electrode data from cortical neurons under various current injection paradimgs.
    Stimulation is done with current steps of different amplitudes and duration.
    Several input samples (step of different duration and amplitude as well as repeated short current pulses) will given as training data together with the electrophysiological measurements. Several stimulation sets with different current step profiles are set apart for the prediction as well as a slow ramp stimulus. For these the input is given, but not the measured voltage trace.
    Details of Challenge B
    Link to Data for Challenge B
    Please submit challenge B results to renaud.jolivet@epfl.ch
  • C: Two electrode data from cortical neurons under random current injection at soma and dendrite.
    Stimulation is done with currents of different means and fluctuation amplitudes. Some traces (input current plus measured voltage) are given for training. Several stimulation sets are set apart for the prediction.
    Details of Challenge C .
    Link to Data for Challenge C (version 1.1/March 27)
    Please submit challenge C results to arnd.roth@ucl.ac.uk
  • D: More multi-electrode data with a focus on subthreshold stimulation. The membrane potential of a layer V pyramida neuron was always simultaneously recorded at three locations: the soma (S1), a distal dendritic location (D1) and a proximal dendritic location (D2). Challenge D has opened on April 3.
    More on the data
    Details of Challenge D
    Link to Data for Challenge D (version 1.0/April 3)
    Please submit challenge D results to arnd.roth@ucl.ac.uk 

Methods and Models

The only aspect that counts for us is the quality of the prediction on the test set. In terms of methods, anything goes. In particular we welcome contributions using

  • Machine learning methods (SVM, Adaboost, ANN, kNN, ...)
  • Systems idenfication methods and dynamical systems (Wiener kernel, nonlinear dynamics, cascade model etc)
  • Simplified neuron models (integrate-and-fire, FitzHugh-Nagumo)
  • Conductance-based neuron models (Hodgkin-Huxley-type models, multicompartment-multichannel models ...)
  • Deterministic or stochastic models

Miscellaneous. For those interested in prediction and neural challenges, please have also a look at The Neural Prediction Challenge . In contrast to this challenge, our Lausanne competitions are based on challenge on direct neuronal stimulation and not sensory stimulation.

 

This document is a literal copy of http://icwww.epfl.ch/~gerstner/QuantNeuronMod2007/challenge.html

 

Document Actions
Send this Print this Add to favorites