A massively parallel architecture for a self-organizing neural pattern recognition machine
Computer Vision, Graphics, and Image Processing
A competitive distribution theory of neocortical dynamics
Neural Computation
Data Mining: Practical Machine Learning Tools and Techniques, Second Edition (Morgan Kaufmann Series in Data Management Systems)
Input Feedback Networks: Classification and Inference Based on Network Structure
Proceedings of the 2008 conference on Artificial General Intelligence 2008: Proceedings of the First AGI Conference
Counting objects with biologically inspired regulatory-feedback networks
IJCNN'09 Proceedings of the 2009 international joint conference on Neural Networks
Unsupervised Segmentation With Dynamical Units
IEEE Transactions on Neural Networks
| Hi-index | 0.00 |
Simultaneous patterns within images may have conflicting interpretations depending on context (other representations concurrently inferred). This causes significant problems known as 'the binding problem' and 'the superposition catastrophe' for recognition algorithms that incorporate parameter optimization (including neural networks). Previously oscillatory dynamics have been proposed to better separate patterns and address such problems. Another dynamic method, independent of oscillation, is proposed that infers which representations fit together. It works by cycling activation between inputs and outputs. Inputs activate contending representations which in turn inhibit their representative inputs. Inputs utilized by multiple representations are more ambiguous and are inherently inhibited more. The inhibited inputs then affect representation activity, which again affects inputs. The cycling is repeated until a steady state is reached. This method allows simultaneous evaluation of representations and can determine what set of representations best fit the whole image. The implementation of feedback dynamics for separating patterns is described in detail and key examples are demonstrated by simulations.