Attention is thought to improve processing of selected stimuli. It has been demonstrated that firing rates can be modulated by attention and that their enhancement can improve the signal-to-noise ratio for discriminating stimulus features. However, from these often small effects observed at the single neuron level, it is not clear how to explain the massive enhancements in perceptual performance known from psychophysical studies.
Combining multielectrode recordings of field potentials with current methods from computational neuroscience, we investigate effects of selective attention on object representation using recordings of epidural field potentials (EFPs). We trained two monkeys to attend to one of two sequences of objects, which were simultaneously presented in both hemifields. The task required the monkeys to identify the re-occurrence of the initial object in the attended hemifield. LFPs were recorded with an array of 37 epidural electrodes, covering parts of area V4 and V1, while the monkeys were performing the task.
Analysis of object encoding was done with standard support vector machines using EFP wavelet power coefficients, allowing to estimate the probability of correct classification of the objects presented in dependence on the attentional state. Classifying on all electrodes, we found a performance of up to 94% correct (1200 ms window, chance level 17%), i.e. the EFP-data enabled nearly perfect object identification. Almost all stimulus-specific information was concentrated in the frequency range from 50 to 100 Hz. Classification on data from 5 electrodes covering V4 resulted in 52% performance for non-attended stimuli, and 63% for attended stimuli. We also classified the direction of attention from the EFP-signal, which reached a performance of up to 91% for all electrodes, and of up to 83% for the 5 electrodes from V4 (chance level 50%).
Our results clearly show that attention substantially improves the encoding of visual stimuli in gamma-band neuronal activity. Using artificial data sets derived from the recorded EFP activities, we investigated which specific features in the data were responsible for the attentional gain in performance. In particular, we first scaled independently either the mean values or the variances of the data in the non-attended condition to the corresponding numbers observed in the data for the attended condition. Then we quantified the resulting gain in performance by classifying with support vector machines trained on these new, artificial data sets. The analysis revealed that the small improvements found in the signal-to-noise ratios under attention contribute only to a small extent to the large improvements in performance. The major part of the attentional enhancement of the neural representation can rather be explained by stimulus- and electrode-specific shifts in the mean wavelet amplitudes, which render the neural activity pattern more distinct from each other. This result uncovers a novel mechanism for attentional enhancement, which could only be discovered by using multichannel recordings combined with a thorough data analysis guided by specific hypotheses about principles of neural coding.
Support contributed by: DIP METACOMP, BMBF Bernstein Group Bremen,
and EU Grant BIND MECT-CT-20095-024831 and BACS FP6-IST-027140.