An electrically trainable artificial neural network (ETANN) with10240 `floating gate' synapses

摘要

The use of floating-gate nonvolatile memory technology for analogstorage of connection strengths, or weights, has previously beenproposed and demonstrated. The authors report the analog storage andmultiply characteristics of a new floating-gate synapse and furtherdiscuss the architecture of a neural network which uses this synapsecell. In the architecture described 8192 synapses are used tointerconnect 64 neurons fully and to connect the 64 neurons to each of64 inputs. Each synapse in the network multiplies a signed analogvoltage by a stored weight and generates a differential currentproportional to the product. Differential currents are summed on a pairof bit lines and transferred through a sigmoid function, appearing atthe neuron output as an analog voltage. Input and output levels arecompatible for ease in cascade-connecting these devices into multilayernetworks. The width and height of weight-change pulses are calculated.The synapse cell size is 2009 μm