|
|
|
 |
|
Michael Berry
Princeton, USA
Wednesday 1 October 2008
16.30
Seminar Room B10 (Basement)
Alexandra House, 17 Queen Square, London, WC1N 3AR
Reading a Correlated Population Code
Throughout the brain, information is represented by large populations of neurons, yet little is known about the population code. A key issue is the pattern and strength of correlations among neurons. Correlations effect both the information represented by a population and the strategies necessary to read out this information. We used the retina as a model system to study these issues. We presented the retina with a set of 36 different shapes and recorded from 162 ganglion cells in multiple retinas. We asked how well the brain could discriminate one shape from all the rest – a task that involves signal correlation. With a binary code for each cell (spike vs. no spike), a decoder that takes correlation into account could reach zero error in 3600 trials for 4 shapes. With a spike latency code for each cell, we could reach zero error for 20 out of 36 shapes. Decoders that ignored correlation (e.g., linear classifier, tempotron) performed much worse, in some cases having error rates >100-fold higher. Decoders that approximated the correlation structure in the population by matching all pairwise correlations (maximum entropy model) were more successful. These results show that both temporal information and correlation among cells can dramatically effect the fidelity of a population code.