M. Szabo, R. Almeida, G. Deco, M. Stetter — A neuronal model for the shaping of feature selectivity in IT by visual categorization (2005)

M. Szabo, R. Almeida, G. Deco, M. Stetter (2005) A neuronal model for the shaping of feature selectivity in IT by visual categorization, Neurocomputing 65–66 (2005) 195–201

The part of the brain that handles complex forms ( inferior temporal cortex : IT ) will change how it responds to stimuli as you learn to categorize objects based on their features. IT learns to ignore features useless for categorization while becoming more sensitive ( selective ) to features that are useful.

The authors propose that this effect might be explained by interaction between IT and another part of the brain, speculatively pref-frontal cortex (PFC), a region implicated in attention and judgement. A competing hypothesis would be that area IT changes how it interprets visual input, rather than relying on such top-down judgements. It is possible that both mechanisms are at play, but on different timescales. However, this paper explains experimental data using only the attention model.

Their model uses Poisson-process neurons with excitatory and inhibitory interactions. To simplify things, they consider a task with only two features, one of which is relevant for categorization, and the other is not. At first glance, the network they build looks a fair bit more complicated than it needs to be. However, all the components are important for achieving the observed results. The paper also uses a much simplified version of the spiking model, but does not present the equations or an explanation. Instead, the reader is directed to http://www.ncbi.nlm.nih.gov/pubmed/11524578.

The model is presented with two features A and B. A network is set up to generate the inverted signals !A and !B using biased competition and biased interactions with excitatory and inhibitory populations. For each layer, an additional set of excitatory and inhibitory cells is attached that have nothing to do with the features or categorization. These populations exist to maintain network activity at a stable non-zero level no matter what the stimulus. All four categories : A, B, !A, !B, project to a second layer with two categories C1 and C2. The model does not know ahead of time which combination of features will be useful for classifying stimuli into either C1 or C2. The model is trained to categorize stimuli where only the A, !A contrast is useful, and B, !B is irrelevant. The model learns to associate A with C1 and !A with C2. Feedback interactions from the categorization units to the feature units enhance the contrast of A and !A to incoming stimulation by interactive activation.

The time-courses of responses in the trained model line up with recorded time-courses. Initial stimulus presentation leads to a weak but measurable selectivity, which is followed after a short lag with a large amplification of selectivity. This late amplification is interpreted as feedback information from the categorisation units, which tentatively reside in prefrontal cortex.

notes :
  • Studies investigating plasticity in IT during task training need to control for the possibility of task-dependent modulation
  • Studies also need to accept the possibility that IT plasticity might be due entirely to IT-PFC interactions
  • The time-courses of activation can help illuminate whether the altered responses involve IT-PFC interactions

No comments:

Post a Comment