초록 ▼

A network architecture for the programmable emulation of large artificial neural networks ANN having digital operation employs a plurality L of neuron units of identical structure, each equipped with m neurons, the inputs (E) thereof being connected to network inputs (EN) multiplied or branching via

A network architecture for the programmable emulation of large artificial neural networks ANN having digital operation employs a plurality L of neuron units of identical structure, each equipped with m neurons, the inputs (E) thereof being connected to network inputs (EN) multiplied or branching via individual input registers (REGE). The outputs (A) of the neuron units are connectable to network outputs (AN) at different points in time via individual multiplexers (MUX) and individual output registers (REGA) and the neuron units have individual auxiliary inputs via which signals can be supplied to them that represent weighting values (W) for weighting the appertaining neural connections and represent thresholds (0) for weighting input signals.

대표청구항 ▼

A network architecture for programmable emulation of artificial neural networks (ANN) having digital operation, characterized by a plurality L of neuron units of an identical structure, each of said units being equipped with m neurons, said neuron units having neuron inputs (E) which are connected t

A network architecture for programmable emulation of artificial neural networks (ANN) having digital operation, characterized by a plurality L of neuron units of an identical structure, each of said units being equipped with m neurons, said neuron units having neuron inputs (E) which are connected to output lines (EN) branching via individual output registers (REGE), said registers having outputs (A) which are selectively switchable by network outputs (AN) via individual multiplexers (MUX) and individual output registers (REGA), and a plurality of individual auxiliary inputs via which signals can be supplied that represent weighting values (W) for weighting said neuron inputs; and further characterized in that the m mutually identical neurons each comprise four inputs to each of which a respective digital multiplier (MUL) is allocated, said multipliers having plural inputs, one multiplier input of said plural inputs being connected to the said neuron inputs and another multiplier input thereof being connected to one of the individual auxiliary inputs; each said multiplier having a result output, the result outputs of the multipliers (MUL) are each respectively connected to an individual input of a digital adder (ADD) allocated to the appertaining neuron; the result output of the adder (ADD) is connected via a register (REG) to a first input of a digital accumulator having a result output, the result output of the accumulator (AKK) is connected to a signal input of a function element (FKT) individually allocated to the appertaining neuron, being connected thereto via a multiplier (MUX) individually allocated to the appertaining neuron and via a further register (REG), the signal output of said function element (FKT) being connected via a further register (REG) to the appertaining network output (AN); a second input of the accumulator (AKK) is connected via a further register (REG) to a further output of the multiplexer (MUX); and a selection line for the selection of a discriminator function (WDF) is connected to a selection input of each of the function elements (FKT), being connected thereto via a register (REG); and further characterized in that the nonmultiplied or non-branching inputs are provided for the neurons, and including four neurons, whereby a first of these inputs is connected to a first adder input of a further adder (ADD) that is allocated to a first neuron, a second of these inputs is connected to a first adder input of a second adder (ADD) that is allocated to a second neuron, a third of these inputs that is allocated to a first adder input of a second adder (ADD) of a third neuron and a fourth of these inputs is allocated to a first adder input of a second adder (ADD) that is allocated to a fourth neuron, and whereby every neuron has an individual auxiliary input allocated to it to which a signal can be supplied that represents a weighting value and that is connected via a register to a second adder input of the second adder (ADD); and a further multiplexer (MUX) is provided for each of the individual accumulators (AKK) to receive the output therefrom, the one output of said further multiplexer being connected to receive the output of the second adder (ADD) and the output of said further multiplexer being connected to the one output of the first multiplexer (MUX) for the formation of an accumulator path.