Neurons in cortical area MT respond to step changes in visual motion with a transient pulse of high firing rate activity before settling down to their sustained level of response (Lisberger & Movshon 1999). These onset transients, unlike the sustained response, show fine-grained temporal structure.

To investigate MT transient dynamics, we recorded extracellularly from area MT of anaesthetised macaque monkeys using methods described in detail in (Kohn & Movshon, 2003). After induction with Ketamine HCl (10 mg kg^-1), anaesthesia was continued during surgery with 1.5-3.5 % isoflurane in a 98 % O₂/2 % CO₂ mixture. Experiments typically lasted 4-5 days, during which anaesthesia and neuromuscular blockade were maintained with sufentanil citrate (4-8 Ág kg^-1 h^-1) and vecuronium bromide (0.1 mg kg^-1 h^-1) in Ringer’s solution containing dextrose (2.5 %). Experiments conformed to local and national guidelines. The adequacy of anaesthesia was monitored continuously (EEG, ECG etc., see Kohn & Movshon (2003) for further details). Animals were humanely killed at the end of experiments.

Using an information theoretic approach (Panzeri & Schultz 2001), we found that the temporal structure in the onset transients can lead to either synergy or redundancy in the information content of spikes nearby in time, with a large variation from cell to cell. However, not all encoded information may be useful, so we compared the amount of information encoded about motion direction with that extracted by two maximum likelihood (ML) decoding strategies. The first strategy utilised complete knowledge of the probability of responses (spike trains) given stimuli, as determined from a training dataset. The second strategy, which we refer to as Poisson decoding, ignored correlations, instead assuming all time bins to be conditionally independent, and the response probability therefore obtained by the product of the spike probability distributions at each point in time. Neither decoder necessarily extracted all of the encoded information, some of which must therefore relate to differences in responses to non-preferred stimuli. The Poisson decoder performed almost as well as the ideal ML decoder, indicating that temporal structure in the onset transients neither helps nor hinders the extraction of motion signals from the spike train.

An adaptive filter implementation of Poisson decoding can be made using such well known mechanisms as synaptic depression and postsynaptic gain control; its near-optimal performance, together with the ubiquity of transient responses, is suggestive of a general mechanism present at each level of visual processing.

This work was supported by the Howard Hughes Medical Institute.