Neural network accommodating parallel synaptic weight adjustments in a single cycle

• US 5,247,206 A
• Filed: 05/19/1992
• Issued: 09/21/1993
• Est. Priority Date: 03/12/1992
• Status: Expired due to Term

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Castro, Hernan A.
• Primary Examiner(s): Hudspeth, David R.
• Application Number:
  US07/885,528
• Time in Patent Office: 490 Days
• Field of Search: 307/201, 395/23, 395/24
• US Class Current: 706/35
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A neural network providing correlation learning in a synapse cell coupled to a circuit for parallel implementation of weight adjustment provides the learning portion of the synaptic operation and includes a floating gate device having a corresponding floating gate member that stores the connection weight of the cell. Parallel weight adjustments are performed in a single operational cycle utilizing floating gate technology and control signals that facilitate programming/erasing operations.

Claims

1. A neural network including an array of cells for computing the product of an input signal and a connection weight to generate a neural output signal, each of said cells comprising an associated circuit for implementing weight updates in a single learning cycle wherein said circuit comprises a charge storage means for storing an electrical charge representing said connection weight, and a logic means for coupling said neural output signal to said charge storage means, said network further comprising a means for generating a voltage pulse stream coupled to said charge storage means, said voltage pulse stream having a magnitude capable of producing an electric field sufficient to effectuate charge transfer to/from said charge storage means.

2. The network of claim 1 wherein said neural output is provided along pairs of column lines, each of said pairs being coupled to a column of said cells within said array.

3. The network of claim 2 wherein said logic means comprises an intermediate node coupled to said charge storage means, a first field-effect device coupled between said intermediate node and a first column line, and a second field-effect device coupled between said intermediate node and a second column line, said first and second column lines comprising one of said pair of column lines within said network;

said logic means further comprising a means for providing a control signal and its inverse to the gates of said first and second field-effect devices, respectively, such that when said control signal is in a first logic state said charge storage means is coupled to said first column line, and when said control signal is in a second logic state said charge storage means is coupled to said second column line.

4. The network of claim 3 wherein said charge storage means comprises a floating gate device having a floating gate member for storage of said charge.

5. The network of claim 4 wherein said floating gate device includes a control gate coupled to receive said voltage pulse stream and a source/drain region coupled to said intermediate node.

6. The network of claim 5 wherein the frequency of said voltage pulse stream is proportional to the magnitude of said input signal.

7. The network of claim 5 wherein the frequency of said voltage pulse stream is proportional to the magnitude of said neural output.

8. The network of either claim 6 or 7 wherein said control signal represents the polarity of said input signal.

9. The network of claim 3 wherein said charged storage means comprises:

a charge storage node;

third and fourth field-effect devices coupled in parallel between said intermediate node and said charge storage node, said third field-effect device having its gate coupled to said intermediate node and said fourth field-effect device having its gate coupled to said charge storage node; and

a capacitor means for coupling said voltage pulse stream to said charge storage node.

10. The network of claim 9 wherein the frequency of said voltage pulse stream is proportional to the magnitude of said input signal.

11. The network of claim 9 wherein the frequency of said voltage pulse stream is proportional to the magnitude of said neural output.

12. The network of either claim 10 or 11 wherein said control signal represents the polarity of said input signal.