Visual attention

Below are some scripts and simulations related to the papers on visual attention and my PhD thesis. The model we developped is based on the continuum neural field theory. It is made of several 2D maps of, say \(N \times N\) units, each of them having a membrane potential \(r(x,t)\) evolving according to the differential equation:

\[\begin{eqnarray} \frac{\partial r}{dt}(x,t) = - r(x,t) + \int_y w(x,y) f(r(y,t)) + I(x,t) + h \end{eqnarray}\]

where \(w(x,y)\) is the weight between the units at position x and y, \(I(x,t)\) is the input for the position x at time t, h is a baseline and f a transfer function usually taken as a heaviside, sigmoid or some monotonically increasing positive function. By appropriately choosing the intra-map connectivity (with the function w) and the inter-map connectivity (defining the input I(x,t)), one can obtain interesting behavior. One we were interested in is visual attention, i.e. the ability to focus ressources on a subset of the visual input. We first consider covert visual attention, i.e. the ability to focus on specific parts (spatial-based or feature-based) of the visual input without eye-movements (the scene is static). The experiments we carried out are classical visual search tasks where a subject must report where a target is. In this kind of experiment, it has been shown experimentally that the reaction time increases linearly with the number of the stimuli that are similar to the target of the search. In the videos, different settings are considered by varying the number of distractors. In our experiments, the distractors always share at least one feature with the target.

Number of distractors Simulation
3
8
14

Some additional videos made when running the simulation on the Intercell cluster

Trial number Simulation
1
2
3

Finally, we can add an additional component to the model. The previous model has an eye-centered working memory storing the position of the stimuli that have already been inspected. If one wishes to allow the agent to perform eye movements, it is sufficient to update this representation with the intended eye movement. Making use of an anticipation map we proposed in another publication, we can now perform overt attention, inspecting a visual scene with eye movements without reinspecting a stimulus that has been already inspected. This is what we demonstrate on the video below where the agent is trying to scan, with eye movements, the blue targets. Sometimes, some distractors catch covertly the attention because they are larger on the retina and produce a strong excitation. However, with this covert shift, the model discovers that attention is indeed on a distractor and decides to covertly switch the locus of attention.