The best k neurons were selected by applying an ANOVA to each neuron in the training set (using the identity labels as levels), and only these neurons were either used or excluded when training and testing the classifier on the single object identity decoding task. For the 16 object stimulus set (upper plots), the results were about as good using only 16 of the most selective neurons compared to using a population of 128 neurons, while for the 7 object stimulus set (lower plots), the results continued to improve slightly as more neurons were added (the difference between the results is probably due to the fact that the there were many more repetitions of each trial type in the 7 object stimulus set, so the classifier could learn the parameters better, and was thus better able to utilize information even in highly noisy neurons). Performance degraded slowly when the top k neurons were excluded.
