MOS multi-layer neural network including a plurality of hidden layers interposed between synapse groups for performing pattern recognition

A MOS multi-layer neutral network receiving an m-bit input and generating an n-bit output, said neural network comprising:

• an input layer having m neurons each selectively arranged to receive a corresponding bit value of said m-bit input, each neuron including first and second serially coupled CMOS inverters, an output from each first CMOS inverter being connected to an input of a corresponding second CMOS inverter to generate a non-inverted output therefrom, said output from each first CMOS inverter also providing an inverted output associated with each neuron;
• an input synapse group having a plurality of transistor synapse means coupled to respective ones of said inverted and non-inverted outputs from said input layer to form a transistor group matrix, said transistor synapse means generating a plurality of output row bias signals, each transistor synapse means being assigned a predetermined weight value (current driving strength) in relation to its position along the transistor group matrix, each row in the transistor group matrix being responsive to a respective bias signal pre-assigned to provide a select unit weight value bias signal onto said each row and also being responsive to a sum bias signal from each of said transistor synapse means respectively coupled thereto and driven by corresponding ones of said inverted and non-inverted outputs,
• wherein each row is associated with one of said output row bias signals generated in response to the selected unity weight value bias signal and the respective sum bias signal coupled thereto;
• an output layer having n neurons each selectively arranged to output a corresponding bit value of said n-bit output;
• at least one hidden layer, each coupled in succession between said input layer and said output layer and including n neurons, wherein a first hidden layer of said at least one hidden layer couples first preselected n output row bias signals from said input synapse group to respective ones of its corresponding n neurons,
• each said at least one hidden layer being logically coupled to transfer the correspondingly input preselected n output row bias signals to both said output layer and to every subsequent hidden layer coupled thereto in succession;
• and
• at least one transfer synapse group, each associated with a corresponding one of said at least one hidden layers and providing respective n first output row bias signals, each transfer synapse group having a plurality of transfer synapse means coupled to corresponding ones of inverted and noninverted outputs from said respective n neurons and further coupled to second preselected n output row bias signals from said input synapse group to form a transfer group matrix for generating said respective first n output row bias signals, each transfer synapse means being assigned a predetermined weight value (current driving strength) as a function of its position along the respective transfer group matrix,
• wherein each row in a respective transfer group matrix is coupled to each of the n neurons of each successive said at least one transfer synapse group or, if said respective transfer synapse group is the last in succession, then to corresponding inputs of said n neurons of said output layer, and
• wherein said input synapse group couples a supply voltage to each neuron of said output layer and said at least one hidden layer such that when a corresponding bit value into an input layer neuron is "1", then when a connection weight value is selected to be positive, a first supply voltage is connected with a connection strength of a corresponding connection weight value through a PMOS transistor whose gate is connected to each corresponding inverted output, and when a connection weight value is selected to be negative, a second supply voltage is connected instead through an NMOS transistor whose gate is connected to each corresponding non-inverted output,
• when a corresponding bit value into said input layer is "0", then when a connection weight value is selected to be positive, a first supply voltage is connected with a connection strength of a corresponding connection weight value through a PMOS transistor whose gate is connected to each corresponding non-inverted output, and when a connection weight value is selected to be negative, a second supply voltage is connected instead through an NMOS transistor whose gate is connected to each corresponding inverted output, and
• when said input bit value is "1" or "0" and a connection weight value is "0, " nothing is connected, and
• wherein each transfer synapse group is composed of PMOS and NMOS transistors for connecting inverted and noninverted neuron outputs associated therewith.