1. Information Processing by Neuronal Populations Chapter 6: Single-neuron and ensemble contributions to decoding simultaneously recoded spike trains Information Processing by Neuronal Populations Chapter 6: Single-neuron and ensemble contributions to decoding simultaneously recoded spike trains Mark laubach, Nandakumar S. Narayanan, Eyal Y. Kimch Summarized by Seung-Joon Yi

2. Summary Narayanan et al. (2005)  Used statistical classifier to quantify the relationship between neural firing rates and a categorical measure of behavior  Estimates of decoding were made for all possible combinations of neurons in a given ensemble  Used LPF and decimation to reduce the complexity of data  Used a wavelet-based feature extraction algorithm  Trained and tested classifiers  Used MI information to quantify the relationship Main contribution  Showed that neural interactions depend on the size of the neural population © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 2

3. Background Early view  Population vectors, constructed from weighted averages of the responses of single neurons, can accurately predict behavior variables. Information theoretic framework  Synergistic coding scheme: Ensembles encode more than the sum of the component neurons.  Redundant coding scheme: Ensembles can be less noisy and less prone to errors © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 3

4. Procedure overview © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 4

5. Decimation and feature extraction © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 5 Decimating spike train to make SDF Feature extraction using wavelet based method

6. Classification © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 6  Two principal components: PC1 and PC2  Two types of behaviors: Short RT and Long RT  Linear DA (straight line) vs. Regularized DA (curved line)

7. Which classifier is better? © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 7 Good classifier vs. Lazy classifier  Similar accuracy : 90% vs. 90.9%  Different mutual information  (b): 0.4395+0.6639-0.9085 = 0.1949 bits  (c): 0.4395+0.1311-0.5555 = 0.015 bits

8. Base results from random data © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 8 Classifiers can perform with a high level of bias  Thresold of MI: 0.06bits

9. Synergy and redundancy © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 9 Individual predictive information: Pi=Ic-Ii

10. Three major results © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 10 Individual neuron information is not predictive of the neuron’s contribution to an ensemble of neurons 96% of neurons contributed redundant information The type of interactions depended on the # of neurons in the ensemble

11. A pair of redundant neurons © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 11 Misclassification on the same trials

12. Role of correlated noise in redundant interactions © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 12 Shuffling the trial orders for the neurons increased accuracy. Correlations in the trial-by-trial variability

13. Simulated networks with different interrelationships © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 13 Direct neural interactions are not necessary  High signal strength: redundancy  Low signal strength: independence  Medium, similar signal strength: redundancy  Medium, different signal strength: synergy

14. Ensemble size vs. relationships The larger the ensemble size, the greater the level of redundancy in general  Synergy: A few cells in a more complex behavioral settings  Redundancy: many cells in a more simple behavioral setting © 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 14