|Consistency of model with neurophysiological recordings|
The multiplicative modulation of tuning curves
Treue and Trujillo (1999) recorded the response from macaque monkey MT neurons when two RDPs are displayed inside the receptive field (RF) of a neuron. One of the RDPs (denoted Pattern A in the figure, reproduced from Treue and Trujillo, 1999) always moves in the anti-preferred direction of the neuron. The second (Pattern B) moves in one of 12 directions. They obtained neural recordings under three conditions: (i) attention to Pattern A, (ii) attention to Pattern B, and (iii) attention to a task-irrelevant fixation point, corresponding to the baseline sensory response. The response of the neuron was enhanced in the second condition, and suppressed in the first, and the modulation was multiplicative.
To investigate if these results can be accounted for by the saliency hypothesis for tracking, a pool of MT neurons tuned to stimulus moving with the same speed but in 12 different directions, (0o,30o, ..., 330o), was simulated with the saliency network of. Twelve RDPs, moving in each of the 12 preferred directions, were generated with the Psychtoolbox. The output (saliency) of the model neuron tuned to 60o was computed at the location of the attended stimuli
Monotonically decreasing modulation
Treue and Trujillo (2004) showed that the extent of modulation
follows a monotonically decreasing function of the angular difference
between the attended motion direction and the neuron's preferred
direction. For this, they used the set-up of the figure below
(reproduced from Trujillo and Treue, 2004). Two RDPs were used, one displayed inside the receptive field of the neuron
whose response was recorded, the other outside. The two RDPs moved in the same
direction. Two settings were used, depending on the location the macaque was
attending to. In the first setting, termed
To reproduce these results using the saliency hypothesis, as before, 12 RDPs, moving in each of the 12 preferred directions, were generated with the Psychtoolbox. In the attend fixation condition, no moving stimulus was used to train the top-down weights and the response of the 12 neurons was computed without the feature selection mechanism. In the case of the attend same condition, the RDP outside the RF is attended to. This was simulated by computing the responses of the 12 neurons to that RDP, and including the top-down feature weights in the saliency computations for the RDP inside the RF. The output (saliency) of the model neuron tuned to 60o was computed at the location of the attended stimuli.The model faithfully replicates the modulation trend observed in the neuronal recordings, showing enhancement for neurons whose preferred direction is close to that being attended, and suppression for anti-preferred directions. Finally, the monotonic fall of the modulation ratio is also accurately replicated by the model.