Neural networks of a proposed type would be formed with computational units called "neuromorphs" attached in a locally connected array. These neural networks would achieve pattern recognition invariant under rotation and translation by exploiting a combination of network symmetry and biologically inspired image-information-processing concepts. The architecture of these networks can be implemented in software; it is also suitable for implementation in hardware in the form of single-chip integrated circuits that would function in parallel-processing modes and thus be capable of fast recognition of patterns.
A typical previously developed image-recognition system effects a process that includes a feature-extraction subprocess followed by a classification subprocess: First, the system extracts features relevant to the kinds of objects that it seeks to classify. These features tend to fall into three classes: global, local, and relational. These feature vectors are then put into a classifier that has been trained or structured to differentiate the feature vectors into different pattern groups.
A neural network of the proposed type would produce feature vectors that would be global, local, and relational, all at the same time. Consequently, for some applications, single feature vectors would be all that would be needed for classification; moreover, inasmuch as features would be easily differentiable, adequate back-end classifier subsystems could be made less complex than those of previously developed pattern-recognition systems. The proposed neural networks could be made quite compact and suitable for analog array processing systems in which neuromorphs would be incorporated directly into the electronic circuitry of image-detecting arrays of photodetectors. Eventually, efforts to integrate this type of neural pattern-recognition circuitry with image-sensing circuitry should lead to the development of single chips that would perform image-acquisition and pattern-recognition functions analogous to visual preprocessing in eye/brain systems.
The concept of a locally connected network of neuromorphs evolved from previous research on the use of pulse-coupled neural networks for fast, invariant automatic target recognition. More specifically, the proposed network architecture is derived from, but is more efficient than, the architecture of an experimental pattern-recognition neural network based on a biomorphic neuron model that includes both the spiking behavior at the axon hillock and the synapto-dendritic processing of a biological neuron. The proposed neural networks would process information in a manner similar to that of the pulse-coupled neural networks, but could be implemented more easily and could process information at higher rates.
The results of computational simulations of the architecture of the proposed systems indicate an ability to perform translation- and rotation-invariant pattern recognition. Analog integrated circuits for use as building blocks (neuromorphs) of locally connected neural networks of the proposed type were designed and tested in other computational simulations, and the results of these simulations indicate that in comparison with previously developed pattern-recognition systems, the proposed system could achieve faster convergence, could operate with lower power consumption, and could be fabricated with greater integration density.
This work was done by Tyson Thomas of Caltech for NASA's Jet Propulsion Laboratory. For further information, access the Technical Support Package (TSP) free on-line at www.nasatech.com/tsp under the Electronics & Computers category.
This invention is owned by NASA, and a patent application has been filed. Inquiries concerning nonexclusive or exclusive license for its commercial development should be addressed to
the Patent Counsel
NASA Resident Office –JPL; (818) 354-5179.
Refer to NPO-20633.