Neural network employing absolute value calculating synapse

Claim 2

A neural network for calculating the absolute difference between an input and a stored weight comprising: means for generating a sequence of voltage pulses including first, second and third voltage pulses; an array of synapse cells arranged in rows and columns, said cells being coupled to said generating means, each cell generating a predetermined sequence of charges in response to said sequence of voltage pulses, said sequence including a first charge representing said input, a second charge representing said weight, and a third charge representing the lesser potential of either said weight or said input, said first, second, and third charges being generated in response to said first, second, and third voltage pulses respectively; a plurality of summing devices coupled to said synapse cells, each device summing said first, second, and third charges generated by a plurality of synapse cells arranged along a single column of said array, the sum of said first charges being calculated in response to said first voltage pulse, the sum of said second charges being calculated in response to said second voltage pulse, and the sum of said third charges being calculated in response to said third voltage pulse; a plurality of accumulators, each of which is coupled to a corresponding one of said summing devices, each accumulator adding said sum of said first charges to said sum of said second charges, and subtracting said sum of said third charges therefrom, to produce an output voltage representing the absolute value difference between said input and said weight for said single column.

Claim 3

The neural network of claim 1 wherein said synapse cells each comprise a pair of field-effect devices coupled in parallel between an internal node and a column line, said column line being coupled to a reference potential and said internal node being coupled to said voltage pulse generating means through a capacitor, one of said field-effect devices having its gate coupled to a row line providing said input and the other said field-effect devices having a floating gate member for storing said weight.

Claim 4

The neural network of claim 2 wherein said first, second, and third charges are generated at said internal node in response to a negative transition of said first, second, and third voltage pulses respectively.

Claim 5

The neural network of claim 3 wherein said summing devices each comprise an operational amplifier configured as an integrator with said column line being coupled to one input of said amplifier.

Claim 6

The neural network of claim 4 wherein said sequence of voltage pulses further comprises a fourth pulse, each said cell generating a fourth charge equal in magnitude to said third charge in response to said fourth pulse.

Claim 7

A neural network comprising: an array of synapse cells arranged in rows and columns, each cell comprising first and second field-effect devices having source nodes that are commonly coupled to an internal node and drain nodes that are coupled to a reference potential, each of said cells further comprising a capacitor coupled to said internal node, and wherein said second field-effect device comprises a floating gate device having a threshold corresponding to a synaptic weight; a plurality of control pulse lines, each control pulse line coupling a pulse stream comprising voltage transitions to said capacitor of each of the cells in the same column of said array; a plurality of column summing lines, each column summing line being coupled to all of the cells in the same column through the drain nodes of said field-effect devices; a plurality of row lines, pairs of said row lines being coupled to all of the cells in the same row wherein one of a pair of row lines selectively couples an input voltage to the gate of said first field-effect device and the other of said pair selectively couples a high positive potential to the control gate of said second field-effect device; said network further comprising a means coupled to said column summing lines for integrating charges developed thereon by said synapse cell, said charges being developed on said column summing lines in response to first, second, and third voltage transitions of said pulse stream wherein said first voltage transition produces a first charge proportional to said input voltage, said second voltage transition produces a second charge proportional to said weight, and said third voltage transition produces a third charge which is proportional to the difference between said reference potential and either said weight or said input voltage.

Claim 8

A neural network comprising: an array of synapse cells arranged in rows and columns, each cell comprising first and second field-effect devices having source nodes that are commonly coupled to an internal node and drain nodes that are coupled to a reference potential, each of said cells further comprising a capacitor coupled to said internal node, and wherein said second field-effect device comprises a floating gate device having a threshold corresponding to a synaptic weight; a plurality of control pulse lines, each control pulse line coupling a pulse stream comprising voltage transitions to said capacitor of each of the cells in the same column of said array; a plurality of column summing lines, each column summing line being coupled to all of the cells in the same column through the drain nodes of said field-effect devices; a plurality of row lines, pairs of said row lines being coupled to all of the cells in the same row wherein one of a pair of row lines selectively couples an input voltage to the gate of said first field-effect device and the other of said pair selectively couples a high positive potential to the control gate of said second field-effect device; said network further comprising a means coupled to said column summing lines for integrating charges developed thereon by said synapse cell, said charges being developed on said column summing lines in response to first, second, and third voltage transitions of said pulse stream wherein said first voltage transition produces a first charge proportional to said input voltage, said second voltage transition produces a second charge proportional to said weight, and said third voltage transition produces a third charge which is proportional to the difference between said reference potential and either said weight or said input voltage.

Claim 9

The neural network of claim 6 further comprising a means for accumulating said charges integrated by said integrating means.

Claim 10

The neural network of claim 7 wherein said accumulating means adds the sum of the first charges to the sum of the second charges, and subtracts twice the sum of the third charges therefrom, to produce an output voltage representing the absolute difference between said input voltage and said synaptic weight.

Claim 11

The neural network of claim 8 wherein said integrating means comprises a switched capacitor circuit.

Claim 12

The neural network of claim 9 further comprising a clocking means for providing clock signals to said integrator and accumulating means, predetermined ones of said clock signals being active during the time between a negative transition and the next positive transition of said pulse stream.

Claim 13

A neural network for calculating the absolute difference between an input and a stored weight comprising: means for generating a sequence of voltage pulses including first, second and third voltage pulses; an array of synapse cells arranged in rows and columns, said cells being coupled to said generating means, each cell generating a predetermined sequence of charges in response to said sequence of voltage pulses, said sequence including a first charge representing said input, a second charge representing said weight, and a third charge representing the larger potential of either said weight or said input, said first, second, and third charges being generated in response to said first, second, and third voltage pulses, respectively; a plurality of summing devices coupled to said synapse cells, each device summing said first, second, and third charges generated by a plurality of synapse cells arranged along a single column of said array, the sum of said first charges being calculated in response to said first voltage pulse, the sum of said second charges being calculated in response to said second voltage pulse, and the sum of said third charges being calculated in response to said third voltage pulse; a plurality of accumulators, each of which is coupled to a corresponding one of said summing devices, each accumulator adding said sum of said first charges to said sum of said second charges, and subtracting said sum of said third charges therefrom, to produce an output voltage representing the absolute value difference between said input and said weight for said single column.

Claim 14

A neural network for calculating the absolute difference between an input and a stored weight comprising: means for generating a sequence of voltage pulses including first, second and third voltage pulses; an array of synapse cells arranged in rows and columns, said cells being coupled to said generating means, each cell generating a predetermined sequence of charges in response to said sequence of voltage pulses, said sequence including a first charge representing said input, a second charge representing said weight, and a third charge representing the larger potential of either said weight or said input, said first, second, and third charges being generated in response to said first, second, and third voltage pulses, respectively; a plurality of summing devices coupled to said synapse cells, each device summing said first, second, and third charges generated by a plurality of synapse cells arranged along a single column of said array, the sum of said first charges being calculated in response to said first voltage pulse, the sum of said second charges being calculated in response to said second voltage pulse, and the sum of said third charges being calculated in response to said third voltage pulse; a plurality of accumulators, each of which is coupled to a corresponding one of said summing devices, each accumulator adding said sum of said first charges to said sum of said second charges, and subtracting said sum of said third charges therefrom, to produce an output voltage representing the absolute value difference between said input and said weight for said single column.

Claim 15

The neural network of claim 11 wherein said synapse cells each comprise a pair of field-effect devices coupled in parallel between an internal node and a column line, said column line being coupled to a reference potential and said internal node being coupled to said voltage pulse generating means through a capacitor, one of said field-effect devices having its gate coupled to a row line providing said input and the other said field-effect devices having a floating gate member for storing said weight.

Claim 16

The neural network of claim 12 wherein said first, second, and third charges are generated at said internal node in response to a negative transition of said first, second, and third voltage pulses, respectively.

Claim 17

The neural network of claim 13 wherein said summing devices each comprise an operational amplifier configured as an integrator with said column line being coupled to one input of said amplifier.

Claim 18

The neural network of claim 14 wherein said sequence of said voltage pulses further including a fourth pulse which generates a fourth charge equal in magnitude to said third charge.