Show Tag: architecture

Select Other Tags

Ravulakollu et al. loosely use the super colliculus as a metaphor for their robotic visual-auditory localization.

Lee and Mumford interpret the visual pathway in terms of Bayesian belief propagation: each stage in the processing uses output from the one further up as contextual information and output from the one further down as evidence to update its belief and corresponding output.

Each layer thus calculates probabilities of features of the visual display given noisy and ambiguous input.

Krauzlis et al. argue that attention may not so much be a explicit mechanism but a phenomenon emerging from the need of distributed information processing systems (biological and artificial) for centralized coordination:

According to that view, some centralized control estimates the state of (some part of) the world and modulates both action and perception according to the state which is estimated to be the most plausible at any given point.

Krauzlis et al. localize this central control in the basal ganglia.

De Kamps and van der Velde argue for combinatorial productivity and systematicity as fundamental concepts for cognitive representations. They introduce a neural blackboard architecture which implements these principles for visual processing and in particular for object-based attention.

Deco and Rolls introduce a system that uses a trace learning rule to learn recognition of more and more complex visual features in successive layers of a neural architecture. In each layer, the specificity of the features increases together with the receptive fields of neurons until the receptive fields span most of the visual range and the features actually code for objects. This model thus is a model of the development of object-based attention.

Friston's predictive coding model predicts a hierarchical cortical system.

Functional segregation and integration are complementary principles of organization of the brain.

Backward connections in the visual cortex show less topographical organization (`show abundant axonal bifurcation'), are more abundant than forward connections.

The visual cortex is hierarchically organized.

It seems a bit unclear to me what determines the hierarchy of the visual cortex if backward connections are predominant.

Feedforward connections in the visual cortex seem to be driving while feedback connections seem to be modulatory.

If what Friston means by `models that do not show conditional independence' includes SOM, then that would explain why I can't find an error signal. Maybe the prior constraint invoked by SOMs is similarity between stimuli?

Possibly, this is a point for future work: model cortico-collicular connections as prediction. But, in Friston's framework, there would have to be ascending connections, too.

The cognitivist interpretation of the terms 'bottom-up' and 'top-down' is that of hypothesis-driven or expectation-driven processing., respectively.

The anatomical interpretation of the terms 'bottom-up' and 'top-down' is that of feedforward vs. feedback connections in a processing hierarchy, respectively.

The terms 'bottom-up' and 'top-down' can mean different, related things depending on context. Engel et al. list four:

  • anatomical
  • cognitivist
  • gestaltist
  • (neural) dynamicist

Grossberg's ART and Friston's theory of cortical responses appeal to the anatomical interpretation of 'top-down' and 'bottom-up' processing and stress feedback as well as feedforward connections.

The temporal binding model implies that related activity of neurons across populations leads to binding of different aspects of stimuli.

The temporal binding theory does not rely on a hierarchical architecture.

Hierarchical architectures for information processing have the benefit of being able to re-use processing primitives for different purposes.

Separating visual processing into channels by the kind of feature it is based on is beneficial for efficient coding: feature combinations can be coded combinatorially.

There are very successful solutions to isolated problems in computer vision (CV). These solutions are flat, however in the sense that they are implemented in a single process from feature extraction to information interpretation. A CV system based on such solutions can suffer from redundant computation and coding. Modeling a CV

All visual areas from V1 to V2 and MT are retinotopic.

The ventral pathway of visual processing is weakly retinotopically organized.

The complexity of features (or combinations of features) neurons in the ventral pathway react to increases to object level. Most neurons react to feature combinations which are below object level, however.

The dorsal pathway of visual processing consists of areas MST (motion area), and visual areas in the posterior parietal cortex (PPC).

The complexity of motion patterns neurons in the dorsal pathway are responsive to increases along the pathway. This is similar to neurons in the ventral pathway which are responsive to progressively more complex feature combinations.

Receptive fields in the dorsal pathway of visual processing are less retinotopic and more head-centered.

There seem to be few, if any, neurons in the pulvinar which receive input from and project to neurons in the same region of MT.

Pulvinar neurons seem to receive input and project to different layers in visual cortex:

They receive input from layer 5 and project to layers one and three.

Connectivity between pulvinar and MT is similar to connectivity between pulvinar and visual cortex.

Stone speaks of the 'conservative nature of evolution' which recycles solutions and applies them wherever they fit. According to this, it is likely that any mechanisms found in visual processing operate in many if not all places of the brain dealing with different but structurally similar functions.