3) Sequence-seeking and counter streams

advertisement

CS560 Assignment 3

Due February 4,2004

Kartik Vishwanath

Sequence seeking and Counter Streams – Shimon Ullman

Part I

Abstract

The paper presents a model for information processing in the brain (Neocortex to be more specific). The basic process is to seek a sequence of transformations linking source and target representations. The source representation is obtained from sensory perception.

The destination is a stored model. This process is termed as sequence seeking . The search for the patterns is bi-directional “bottom up” as well as “top down”. The searches are performed in two separate paths that seek to meet. A track being traversed leaves a trace in the complementary stream. During the entire search process a maximal number of patterns that can be co-active are explored. Once a sequence has been used successfully its use is reinforced by increasing its relative strength. Other techniques are applied to increase the speed in finding match for instance “Express lines” are used to traverse more than a single level in the ascending route and activate patterns that seem more probable or to constrain patterns that seem improbable. The entire operation of sequence seeking is implemented in a structure called counter streams which defines two complementary pathways that seek to meet.

Conclusion

The basic operation of the neocortex is to seek a set of transformations from the source to target representation. The individual transformations themselves maybe different for different parts of the cortex but the overall procedure of source to destination sequence is common. Many sequences are explored simultaneously both “top-down” as well as

“bottom-up” and successful sequences are learnt for future mappings.

Part II

1) Explain why bidirectional search is more efficient in time or space than unidirectional search. Illustrate your argument using reasonable numbers for the branching factor and for the number of levels lying between the sensory end.

Bidirectional search is more efficient in terms of time and space compared to unidirectional search. Consider a case of unidirectional breadth first search (which seems more applicable in this scenario). We start from a root node and try out various nodes at the next level and so on. Thus depending on the branching factor and the depth of the

solution the complexity increases exponentially in other words O(b^d) where b=branching factor and d=depth of the solution. Now if we keep every node in memory the space complexity will also be O(b^d). With the same approach in bi-directional search and keeping the branching factor same the time and space complexity will only be

O(b^(d / 2)) because our depth of the solution becomes limited to finding a common node between the “top-down” and “bottom-up” searches. Thus overall the bidirectional search is a more optimal solution.

2) Explain why it would it be dangerous for a descending pathway to actuate an ascending pathway.

The ascending pathway can actuate a pathway in the higher level because it represents a stored model or a hypothesis of what the sensory perception could mean. If the descending pathway attempts the same it will be dangerous because that would lead to confusion in the sensory perception of the world. A real world example can be thought of in terms of the phenomenon of hallucinations. In this phenomenon people visually perceive objects or events that do not exist. Relating this with this paper one can conclude that in such cases a pathway from a stored model actuated a pathway of a lower level that corresponded to sensory data being interpreted wrongly. Moreover the stored models are learnt over a period of time and that might not be a correct representation of the currently perceived model (In other words the currently perceived model may be entirely new) thus if descending paths are allowed to actuate ascending paths they might lead to use of incorrect or old models that have not been updated.

Download