A model of thalamo-cortical interaction for incremental binding in mental contour-tracing
Fig 10
Incremental binding framework and complexity considerations.
Binding computation can abstractly be represented by a graph with nodes coding for information and edges propagating specific semantics. (A) Each neural processor (NProc) captures the processing of a local neural processing site. NProcs encode semantic information states: no evidence (white filling), existing base evidence state (textured filling), or a binding state (black filling). (B) Necessary bi-directional information flow among NProcs to realize incremental binding (left) and available connectivity patterns to implement this information flow using uni-directional couplings (right). The proposed target architecture uses the “Three channels, ternary NProc, Proxy node” pattern (last column). (C) State-dependent binding speed based on a hierarchical incremental binding architecture as proposed by Roelfsema & Houtkamp [15]. The graphs conceptualize the cases of faster (left column) and slower (right column) binding speeds, respectively. Here, the connectivity pattern for ternary NProcs is shown. After an initial phase of evidence accumulation (first row) incremental binding can take place among the hierarchy starting from an NProc for which an attentional seed (red arrow) has been provided. Combining base evidence and neighborhood information, specific NProcs among the hierarchy become (un-)available (indicated by pink star) and speed up or slow down the binding process. (D) Two decomposed sub-signals can be joined by an “and” operation to retrieve the binding state information of an NProc. (E) Main connectivity variants underlying the complexity considerations. If NProcs are extended for feature selectivity per location (denoted by different fillings), full connectivity is required to compute the binding state information (left). Using a proxy node to aggregate the binding state information across features greatly reduces the number of connections (right; see text for details).