Channel-specific adaptive temporal integration

Douglas McLelland1, Pamela M. Baker and Wyeth Bair

Dept. Physiology, Anatomy and Genetics, Univ of Oxford, Oxford UK
1 douglas(dot)mclelland(at)dpag(dot)ox(dot)ac(dot)uk


Sensory systems show adaptive computation, to better match sensitivity to stimulus statistics. For example, in adaptive temporal integration (ATI) described by Bair and Movshon (2004), when stimuli are fast-moving, direction selective (DS) cells in visual cortex show a narrow window of temporal integration, but this broadens substantially for slow-moving stimuli. Here, we test the plausibility of one candidate model to account for this behaviour, in which the DS cell is driven by two independent channels with different stimulus preferences and integration characteristics. We find that this model does indeed show stimulus-dependent ATI. Further, we find that it is capable of simultaneously supporting two different profiles of temporal integration for independent signals.

Visual Stimulus

Stimulus 1. The stimulus consists of a dynamically moving sinusoidal grating, optimised for the size, spatial frequency and orientation preference of the DS cell. On each 10 ms "frame", the grating moves randomly (actually, according to an m-sequence) by a fixed spatial phase in either the preferred or anti-preferred direction. By setting the value of the spatial phase jump on each frame, we set the equivalent temporal frequency (ETF, the temporal frequency of the grating if it moved in the same direction for several consecutive frames). The stimulus is uncorrelated in the motion domain, allowing for the use of reverse correlation techniques, such as spike-trigerred averaging, to estimate kernels of integration.

To view this stimulus, click the button below to launch the iModel stimulus viewer. Then click Play or scroll through frames manually to see the sinusoidal stimulus move back and forth.

Stimulus to be added.

Stimulus 2. A second stimulus comprised two superimposed, orthogonally oriented gratings, with dynamic motion characteristics as above, and independent random motion. The orientation of the stimulus was rotated by 45 degrees, such that both component gratings have preferred and non-preferred directions of motion. The ETF of each grating can be set independently. To view this stimulus....

Model Results

Result 1. We tested a network population model, DS_Post_Fac_T2 with the dynamic motion grating, over a range of ETF values from 1.56 to 25 Hz, and generated spike-triggered averages (STAs) for spike trains from DS cells. The STAs can be viewed by clicking the button below, and selecting "Dir Mod STA" result,

Please get a Java compatible browswer to see this. Takes a few seconds to load.

It is apparent that for low ETF values, STAs are broad, whereas for high ETF values, STAs are narrow. Thus, a 2-channel model is capable of supporting two different temporal profiles of STA. It remains to be tested whether a model with multiple channels can support a continuous range of STA widths.

Result 2. We tested the same network population model as above, DS_Post_Fac_T2 with the dual-grating stimulus, with a slow ETF (6 Hz) for one grating and a fast ETF (25 Hz) for the other. We then calculated STAs for a single set of output spikes, but separately against the motion of each grating. The STAs can be viewed using the same button as above, and selecting the "Dir Mod Mask STA" plots.

STAs thus calculated for a single set of spikes were broad for the motion of the slow grating and narrow for the motion of the fast grating. Switching the identity of the fast and slow gratings likewise switched the broad and narrow STAs.


The following models are relevant to the points here: