Abstract: The apparatus for the recognition of speech includes an acoustic preprocessor, a visual preprocessor, and a speech classifier that operates on the acoustic and visual preprocessed data. The acoustic preprocessor comprises a log mel spectrum analyzer that produces an equal mel bandwidth log power spectrum. The visual processor detects the motion of a set of fiducial markers on the speaker's face and extracts a set of normalized distance vectors describing lip and mouth movement. The speech classifier uses a multilevel time-delay neural network operating on the preprocessed acoustic and visual data to form an output probability distribution that indicates the probability of each candidate utterance having been spoken, based on the acoustic and visual data. The training system includes the speech recognition apparatus and a control processor with an associated memory.

Abstract: A signal processing apparatus has a circuit network which is formed by connecting a plurality of neuron units into a network, each of the neuron being provided with a self-learning means having a weight function varying means and a weight function generating means for generating a variable weight function of the weight function varying means, on the basis of a positive or negative error signal obtained as a result of the comparison between an output signal and a teaching signal. In order to obtain a positive error signal .delta..sub.j(+) and a negative error signal .delta..sub.j(-), there is provided a differential coefficient calculating means for calculating two kinds of differential coefficients for a neuron response function, the calculation being done on the basis of the output signal from the neuron unit.

Abstract: The apparatus for the recognition of speech comprises an acoustic preprocessor, a visual preprocessor, and a speech classifier that operates on the acoustic and visual preprocessed data. The acoustic preprocessor comprises a log mel spectrum analyzer that produces an equal mel bandwidth log power spectrum. The visual processor detects the motion of a set of fiducial markers on the speaker's face and extracts a set of normalized distance vectors describing lip and mouth movement. The speech classifier uses a multilevel time-delay neural network operating on the preprocessed acoustic and visual data to form an output probability distribution that indicates the probability of each candidate utterance having been spoken, based on the acoustic and visual data.

Abstract: An improved method and apparatus for correcting for splay is provided. A document distorted by the curvature of a page of text away from a platen is converted to a digital image. The digital image is the manipulated to remove the distortion by fitting the lines of text in an unsplayed portion to a skew line, which represents the deviation of lines of text in the digital image from horizontal. Then the splay is determined for each line of text. Once the skew and the splay are determined, an inverse transformation is done to straighten the lines of text. A horizontal stretching is also applied to the text to correct for the projection angle of the original document.

Abstract: The invention provides an improved method and apparatus for compression of palettized images. Input symbols in an M-ary alphabet are binarized based on a context model of the input data, where the binarization is selected to provide good compression by a binary encoder. The particular binarization is determined from a reindexing table which maps each input symbol to a number of binary values. The mapping is determined from the images to be compressed, and is typically transmitted with the compressed images as overhead. The mapping is a local minimum of the bitwise entropy of the binarization. With or without reindexing the input, the symbols can be converted compressed in parallel, with the bits of the input symbols buffered and reordered as necessary to ensure that bits needed for context of a bit being decoded are available before the decompressor decodes the bit being decoded.

Abstract: A three layer artificial neural network having an N terminal input, a two cell hidden and a single cell output layer generates an output parity signal indicating whether an even or an odd number of binary bits are asserted at the N terminal input. The two hidden layer neural cells have activation functions that cause deviations about a linear response characteristic that allow the classification of a signal representative of the number of asserted input bits into odd and even groups. This network represents a significant reduction in the number of hidden units previously required because of the particular form of activation transfer characteristic used in the two hidden layer neural cells.

Abstract: A method and apparatus for placing digital data on plain paper. One embodiment of the present invention allows for the digital data to undergo encryption before being placed on the plain paper. In one embodiment, a photocopier is used for transferring digital encrypted data to and from a plain piece of paper. The photocopier allows digital data to be stored onto plain paper after encryption, such that the digital data is secure. The photocopier also includes a device to recognize the encrypted digitized pixels on the page such that they may be decrypted and the original image reproduced.

Abstract: An artificial network for encoding the binary on-state of one-out-of-N inputs, say j, when only one state is on at a time wherein the jth on-state is represented by a suitable output level of an N-input MP type neuron operating in the non-saturated region of the neuron output nonlinearity. A single line transmits the encoded amplitude level signal to a decoder having N single input neural networks.

Abstract: The present invention relates to the interrelationships between nature (as mediated by evolution and genetic algorithms) and nurture (as mediated by gradient-descent supervised learning) in a population of neural networks for pattern recognition. The Baldwin effect is demonstrated that learning can change the rate of evolution of the population's genome - a "pseudo-Lamarkian" process, in which information learned is ultimately encoded in the genome by a purely Darwinian process. Selectivity is shown for this effect: too much learning or too little learning in each generation leads to slow evolution of the genome, whereas an intermediate amount leads to most rapid evolution. For a given number of learning trials throughout a population, the most rapid evolution occurs if different individuals each receive a different number of learning trials, rather than the same number.

Abstract: A circuit employing logical gates for calculating activation function derivatives on stochastically encoded signals. In one embodiment a two input nonlinear circuit calculates neuron activation functions suitable for gradient-descent learning.