Abstract: A method, apparatus, and computer program product for identifying objects. Pixels for an image of an object are received. A plurality of pixels from the pixels are selected a plurality of times to form groups of sample pixels. Locations in the image for the plurality of pixels in a group of sample pixels are randomly selected for each group in the groups of sample pixels. A vector having a plurality of elements is formed using the groups of sample pixels. Each element in the plurality of elements in the vector corresponds to a group from the groups of sample pixels. The object is identified using the vector.

Abstract: A self-organizing adaptive replicated (SOAR) for creating a replicate of human expert behavior. The SOAR can be embedded invisibly within multiple types of systems to observe, adapt and grow to emulate a user's interactive behavior and performance level. The system yields near equivalent responses to near equivalent stimuli in real time. The SOAR is based on a three layer perceptron type architecture which guarantees arbitrary M to N mapping of continuous valued spaces. The architecture uses a competitive, additive, and layer independent learning rule which insures excellent rapid learning. A self-organizing, adaptive algorithm permits the SOAR to adapt to the true classification space. The SOAR has applications in areas such a speech recognition, target detection, pattern recognition of multi-feature data, electro-mechanical subsystem control and resource allocation and optimization.

Abstract: A system (20) for producing an assignment of a plurality of entities (12) to a plurality of objects (16). This assignment is optimized subject to fixed constraints. The system (20) represents the assignment as a connectionist processing architecture having multiple processing elements (22), (24). These processing elements (22), (24) include a first class of processing elements (22) each representing groups of the entities (12) and a second class of processing elements (24) representing the objects (16). Also a plurality of interconnections (26) between the first and second classes of processing elements (22), (24) having variable weighted connection strengths which are a function of the constraints and also a function a random noise factor. The system (20) also includes a means for selectively assigning the entities (28) one-by-one to the object based on the strength of the interconnections (26).

Abstract: A fine grained multi-planar clutter rejection processor (10) for correlating multiple sets of data. The processor (10) maps each set of data onto a plurality of arrays (28-34). The data includes target data which is correlated between sets and clutter which is uncorrelated between sets of data. The system also includes a means for shifting (40) the positions of the second and subsequent arrays in a pattern which is larger for each successive array. In addition, a correlation identification unit (78) identifies the coordinate locations in the first array (28) which contain data points and which also contain data points in subsequent arrays in their shifted positions. In this way, data points identified in this manner are correlated and the remaining data points can be discarded as clutter. The processor (10) system is able to handle a very large number of data points per scan (over 100,000) over a high number of scans (such as eight).

Abstract: A correlation processor (22) for solving large correlation problems involving multi-dimensional data (12, 16, 20). The system correlates vectors in a first set with vectors in a second set. The system includes a plurality of input devices (24) for accepting the entities (16) in the second set of data (14). Also, a set of processing elements (26) are included for producing a correlation output signal that is a first predetermined function of its inputs. A set of variable strength interconnections (28) couple each of the input devices (24) to each of the processing elements (26) through a connection having a strength that is a second predetermined function of the first set of multi-dimensional data (12), wherein the weighted connection (28) to each processing element (26) is proportional to one of the entities in the first set of multi-dimensional data.

Abstract: The neural engine (20) is a hardware implementation of a neural network for use in real-time systems. The neural engine (20) includes a control circuit (26) and one or more multiply/accumulate circuits (28). Each multiply/accumulate circuit (28) includes a parallel/serial arrangement of multiple multiplier/accumulators (84) interconnected with weight storage elements (80) to yield multiple neural weightings and sums in a single clock cycle. A neural processing language is used to program the neural engine (20) through a conventional host personal computer (22). The parallel processing permits very high processing speeds to permit real-time pattern classification capability.

Abstract: A system is provided for associating a set of data values where each data value belongs to one of a plurality of classes and also to one of a plurality of types. The system includes an array of memory cells arranged in rows and columns, each memory cell storing one data value. An energy computation cell for each column is coupled for computing an energy value proportional to the likelihood that the data values in each column comprise one of the desired associations. A system control module to the energy computation cells for transferring the data values of a given class and type from one memory cell to another memory cell of the same class and type, within a given row of memory cells when the replacement lowers to the energy value for the column. The energy values computed by all the energy computation cells are summed to determine a total energy value. Each time data values are transferred, the total energy value of all of the energy cells is computed.

Abstract: A method of training a multilayer perceptron type neural network to provide a processor for fusion of target angle data detected by a plurality of sensors. The neural network includes a layer of input neurons at least equal in number to the number of sensors plus the maximum number of targets, at least one layer of inner neurons, and a plurality of output neurons forming an output layer. Each neuron is connected to every neuron in adjacent layers by adjustable weighted synaptic connections.

Abstract: A modular, cellular automaton system (26) for identifying patterns in multi-dimensional data. The system includes a regular two-dimensional array (40) of interconnected processing cells (44) and data cells (42), wherein each data cell (42) is located at a particular x-y location on a cartesian coordinate system. The system (26) includes a means for accepting the data and for transmitting the data to data cells (42) having x-y coordinate locations which correspond to the coordinates of the data points. The system (26) sums data values surrounding each of the data values and compares the sums with a threshold. Sums are set to zero if the threshold is not exceeded. In an iterative process, the summing and thresholding is repeated until the sum is set to zero or the sum no longer changes on two consecutive iterations. In this way, data cells (42) which remain at non-zero values together form the pattern to be identified in the data.

Abstract: An Adaptive Network For In-Band Signal Separation (26) and method for providing in-band separation of a composite signal (32) into its constituent signals (28), (30). The input to the network (26) is a series of sampled portions of the composite signal (32). The network (26) is trained with at least one of said composite signals (28) (30) using a neural network training paradigm by presenting one or more of the constituent signals (28) (30) to said network (28). The network (26) may be used to separate multiple speech signals from a composite signal from a single sensor such as a microphone.

Abstract: A cellular network assignment processor (10) for solving optimization problems utilizing a neural network architecture having a matrix of simple processing cells (12) that are highly interconnected in a regular structure. The cells (12) accept as input, costs in an assignment problem. The position of each cell (12) corresponds to the position of the cost in the associated constraint space of the assignment problem. Each cell (12) is capable of receiving, storing and transmitting cost values and is also capable of determining if it is the maximum or the minimum of cells (12) to which it's connected. Operating on one row of cells (12) at a time the processor (10) determines if a conflict exists between selected connected cells (12) until a cell (12) with no conflict is found in each row. The end result is a chosen cell (12), in each row, the chosen cells (12) together representing a valid solution to the assignment problem.

Abstract: A neural network signal processor (NSP) (20) that can accept, as input, unprocessed signals (32), such as those directly from a sensor. Consecutive portions of the input waveform are directed simultaneously to input processing units, or "neurons" (22). Each portion of the input waveform (32) advances through the input neurons (22) until each neuron receives the entire waveform (32). During a training procedure, the NSP 20 receives a training waveform (30) and connective weights, or "synapses" (28) between the neurons are adjusted until a desired output is produced. The NSP (20) is trained to produce a single response while each portion of the input waveform is received by the input neurons (22). Once trained, when an unknown waveform (32) is received by the NSP (20), it will respond with the desired output when the unknown waveform (32) contains some form of the training waveform (30).

Abstract: An information processor (10) for solving assignment problems uses a matrix of individual processing cells. The location of each cell within the processor (10) corresponds to the position of the costs in the associated constraint space of the assignment problem, and each cell contains a cost register (12) that is stored with an associated cost value. A noise generator (14) associated with each cell is used to trigger a variable threshold (16) in each cell so that cost values may be transmitted by each cell only when the signal from the noise generator (14) exceeds the threshold. When triggered, the cell disables all other cells along each dimensional axes stemming from the cell from contributing to the current tentative solution. This disabling is removed at the beginning of the next cycle. Successive solutions to the assignment problem are evaluated by adding the cost values in an accumulator (30) for cells that represent each particular solution.

Abstract: A Fine-Grained Microstructure Processor (FMP) (10) for solving assignment and correlation problems that utilizes a pair of arrays (22, 24) of cells (26) to represent the position of two set of two-dimensional data points (14). The arrays (22, 24) are divided by a predefined shaped (16) having a plurality of regions (18). The FMP (10) counts the number of data points (14) in each region (18) and then finds the difference between the resulting sums from corresponding regions (18) in the two arrays (22, 24). The differences for each regions (18) are added together by an accumulator (34) to determine a correlation factor. The data in the second array (24) is then shifted until a different data point (14) occupies a particular position with respect to the predefined shape (16) and a new correlation factor is found.

Abstract: A wavefront vector correlation processor provides rapid, weighted correlation of a given vector with other vectors in a given vector space. In one embodiment the processor includes an interconnected two-dimensional array of simple processing elements that is an electronic analogue of the X,Y plane. Each processing element is connected to neighboring processing elements and is capable of transmitting a pulse to its neighboring processing elements in response to an input pulse. An initial pulse initiated at a processing element having a first classification causes electronic ripples to spread out in chain/reaction among neighboring processing elements. When the electronic ripple reaches processing elements having a second classification, a controller circuit senses these events and measures the time between the initial pulse and the detected pulse. This measured time may then be used to provide correlation between the vector of the first classification and vectors having a second classification.

Abstract: An ultra-high speed two-dimensional coordinate transform processor (10) contains layers having conducting lines representing coordinate lines in coordinate systems. Layers in a first set have conducting lines representing a first coordinate system. Layers in a second set have conducting lines representing a second coordinate system and are positioned adjacent the layers in the first set. A signal is applied at a point where lines in the first set intersect. This point represents a point in the first coordinate system. An induced signal is sensed in the layers in the second set at a point where coordinate lines intersect. Since the detected signal will be strongest at the point nearest the applied signal, the intersecting point having the strongest signal will represent the point in the second coordinate system that is the transform of the first point. The processor (10) is faster than conventional coordinate transform methods and is inexpensive to produce.

Abstract: An analog associative processor (12) discriminates between intersects representing targets and intersects representing ghosts from angle-only information from multiple sensors (10). The analog associative processor (12) is constructed using a multi-layer substrate that is an analog of the real-world sensor emplacements, angular sensor traces, and coordinate system. An algorithm is implemented by the analog associative processor (12) which detects ghosts by counting pulses received by each intersect from the other intersects. When the number of pulses reaching a given intersect reaches a predetermined threshold, the intersect is identified as a ghost. The analog associative processor (12) then continues until the number of remaining intersects equals the number of total targets. The remaining intersects are then identified as true targets.