Convolutional expert neural system (ConExNS)
AssigneeInternational Business Machines Corporation
InventorsManuel Aparicio, IV, Samuel E. Otto
Life of this patent
Estimating the term…
Abstract
Based on the recognition that transfer functions of Boolean completeness can be expressed as a domain within a periodic function, an architecture of a artificial neuron which is fully generalized in application and capable of rapid learning with minimal memory requirements while maintaining content addressability of memory, is provided. Full functionality of this architecture is demonstrated for an input vector containing two values. Extension to three variables shows the potential for generality of this architecture to N-valued input vectors.
Claims29
A method of adaptively altering the activation function of an artificial neuron including the steps of
- storing in a memory a number of weighting values at least equal to the number of values which are included in an input vector and a wave value corresponding to at least one of the slope and the period of a wave function representing at least one activation function of said artificial neuron,
- computing at least one periodic wave function from said period and said weighting values, including computing phase and amplitude values of said periodic wave function from said weighting values,
- deriving at least one value from said periodic wave function corresponding to a value included in an input vector,
- adding all said values derived from said at least one periodic function,
- writing at least one input vector to memory, and
- counting and storing the number of input vectors written to said memory.
A method of adaptively altering a state of an artificial neuron including the steps of
- writing at least one input vector to memory of said artificial neuron,
- determining a threshold value based upon a number of vectors written to said memory, and
- determining whether a current input vector has been written to memory on the basis of said threshold value.
An artificial neuron including
- means for storing in a memory a number of weighting values at least equal to the number of values which are included in an input vector and a wave value corresponding to at least one of the slope and the period of a wave function representing at least one transfer function,
- means for computing at least one periodic wave function from said period and said weighting values, including computing phase and amplitude values of said periodic wave function from said weighting values,
- means for deriving at least one value from said periodic wave function corresponding to a value included in an input vector,
- means for adding all said values derived from said at least one periodic function,
- means for writing at least one input vector to memory, and
- means for counting and storing the number of input vectors written to said memory.
A neural network for developing an output signal responsive to a sequential set of concurrent input variables comprising an input vector, each sequential set of input variables separated by Hamming distances, said neural network comprising:
- means for assigning a selected weight to each of said input variables;
- phase generation means for generating a phase value equal to the sum of all positive weights assigned to said set of input variables;
- and
- output signal generation means for generating a periodic transfer function domain as said output signal having a phase equal to said phase value, and a period equal to the sum of magnitudes of ones of said weights assigned to each of said input variables which are erroneous for a current input vector
- whereby the phase and period of said transfer function expresses said Hamming distance between a given input vector and a previously stored input vector.
A neural network for developing an output signal responsive to a sequential set of concurrent input variables comprising an input vector, each sequential set of input variables separated by Hamming distances, said neural network comprising:
- means for assigning a selected weight to each of said input variables;
- amplitude generation means for generating an amplitude value equal to the sum of the absolute value of all weights assigned to said set of input variables;
- and
- output signal generation means for generating a periodic transfer function domain as said output signal having an amplitude equal to said amplitude value and a period equal to the sum of magnitudes of ones of said weights assigned to each of said input variables which are erroneous for a current input vector
- whereby the amplitude and period of said transfer function expresses said Hamming distance between a given input vector and a previously stored input vector.
A neural network for developing an output signal responsive to a sequential set of concurrent input variables comprising an input vector, each sequential set of input variables separated by Hamming distances, said neural network comprising:
- means for assigning a selected weight to each of said input variables;
- amplitude generation means for generating an amplitude value equal to the sum of the absolute value of all weights assigned to said set of input variables;
- phase generation means for generating a phase value equal to the sum of all positive weights assigned to said set of input variables;
- and
- output signal generation means for generating a periodic transfer function domain as said output signal having an amplitude equal to said amplitude value, a phase equal to said phase value, and a period equal to the sum of magnitudes of ones of said weights assigned to each of said input variables which are erroneous for a current input vector
- whereby the amplitude, phase and period of said transfer function expresses said Hamming distance between a given input vector and a previously stored input vector.
Dependencies are inferred from the claim text.