Difference between revisions of "InternalRepresentationResearch"
From aHuman Wiki
(Automated page entry using MWPush.pl) |
(No difference)
|
Revision as of 01:27, 16 June 2015
Internal Representation of Perceived Sensor Data
@@Home -> NeoCortexResearch -> InternalRepresentationResearch
References
- High level requirements - see http://code.google.com/p/ahuman/wiki/aMatterRequirements
Final Design
- spatial poolers - several, overlapping
- temporal poolers - several
Constructing Regions
Source parameters:
- input rectangular region of size X,Y
- outputs is also some rectangle, where each output item is probability of corresponding sequence
- low-probability items will not be activated if use accumulated activation
- spacial factor defines how big are recognised spacial patterns
- temporal factor defines how complex can be recognised sequence - temporal pattern
- depth factor defines how complex is spacial-temporal pattern
Sequences and Columns:
- item corresponds to cortical column
- column stores spacial-temporal sequence
- until allocated (filled), column is empty or half-empty
- column defines maximum sequence size
- half-empty sequence can be recognised, not trying to enlarge - means supporting variable-length sequences
- so column can be: empty, partial/full unrecognized/recognised
- column is recognised if frequence is high enough
- sequence can be forgotten by column
Regions:
- biologically plausible is having several regions equivalent in inputs/output sizes
- can enlarge from bottom (intput) to top (output)
- can have asymmetric sizes compression
- if area overlap is zero, then input item is presented only in one output item; it means spacial inputs patterns are mapped to several outputs items, no one item represents whole picture; input pattern is represented by output pattern
- if area overlap is not zero - then it can require too many regions to enable represent any inputs pattern by one outputs item
- if find most active column on top level it is classical Jeff's HTM