The idea behind the neural prediction challenge
Motivation
One important goal of any field of science is to develop a theory (or model) that predicts future outcomes. In the particular case of computational and systems neuroscience, we seek a model that can predict the activity of neural systems engaged in sensory processing or behavioral control. The accuracy of a such a model would serve as a benchmark for our understanding of the brain, as well as a tool for revealing principles of neural function.
Clearly, an ideal model of the brain would accurately predict the responses of neurons under natural conditions. But computational neuroscience has yet to produce models that can describe neural responses to natural stimuli. Most current neural models have been constructed based on data collected in classical reductionist experiments optimized to provide powerful tests of specific hypotheses. Although such experiments are critical for model construction, real-world model performance can only be assessed by examining how well a model predicts neural responses under natural conditions. This fundamental problem compels the Neural Predection Challenge. The aim of the Challenge is to accelerate the development of predictive models, and to provide computational neuroscientists an opportunity to test their models.
The aim of the Neural Prediction Challenge is to accelerate the development of predictive models and to provide computational neuroscientists an opportunity to test their models objectively.
The Challenge
The challenge is really quite simple: We will give you some (visual and/or auditory) stimuli and corresponding neural responses, and you must try to predict responses to other stimuli. Each data set will be divided in to two subsets: a fit set (90% of the data) that includes both the stimuli and the corresponding neuronal responses; and a validation set (10% of the data) that includes only stimuli (no responses). Your job is to use the fit set to fit your model and then to generate predicted responses based on the stimuli provided in the validation set. Once you have the predictions you should return them to us. We will compare your predicted responses to the responses actually observed in the validation set.
Data
Current data consist of recordings from visual and auditory neurons during naturalistic stimulation. (In the future we hope to make available representative data sets from many different sensory systems.) Data are provided in simple ascii files that are easily readable in Matlab (or by any other modern programming language). Details on data formatting are provided with each data set.
Predictions
Predictions will be evaluated continuously as they are received and results will be posted in aggregate form. Individuals' names, prediction scores and models will not be posted without prior permission (though we may contact participants directly, see official rules). Please note that this is an academic research project, it is not a traditional contest. There is no real ending date, and there is nothing to win.
Who Can Participate?
To ensure that we will be able to give each entry serious consideration, the challenge will be limited to those working in computational neuroscience. For this reason these data will only be distributed to qualified personnel at academic and research institutions.