This paper proposes a complete self-contained chip set for ANN systems based on a back-propagation model with full-connectivity topology and on-chip learning and refreshing. We believe that our proposal is the first to offer fully hardware-supported on-board learning in continuous analog mode. Another unique novel feature of our 2-chip set design is that it does not require any other supporting chips of any kind and is capable of growing to any arbitrarily large size. Interface and communications with host and/or real-world systems is done via the direct analog and digital I/O ports which allow real-time computations. A representative system of 1 layer with 100 K analog synapses would require 100 SynChips and 20 NeuChips and expected to fill the entire surface of a whole wafer. Such a system which acquires an I/O port of up to 12800 digital bits or 3200 analog channels runs in absolute parallelism and would sustain a performance of 940x10/sup 12/ CPS. Larger systems may be constructed with several wafers.
[1]
Lawrence D. Jackel,et al.
An analog neural network processor with programmable topology
,
1991
.
[2]
H. Shinohara,et al.
A refreshable analog VLSI neural network chip with 400 neurons and 40 K synapses
,
1992
.
[3]
Yannis Tsividis,et al.
A reconfigurable VLSI neural network
,
1992
.
[4]
Carver Mead,et al.
Analog VLSI and neural systems
,
1989
.
[5]
H.P. Graf,et al.
A reconfigurable CMOS neural network
,
1990,
1990 37th IEEE International Conference on Solid-State Circuits.
[6]
Tetsuro Itakura,et al.
Neuro chips with on-chip back-propagation and/or Hebbian learning
,
1992
.
[7]
Ralph Etienne-Cummings,et al.
An analog neural computer with modular architecture for real-time dynamic computations
,
1992
.