Neural network using stochastic processing

• US 4,972,363 A
• Filed: 02/01/1989
• Issued: 11/20/1990
• Est. Priority Date: 02/01/1989
• Status: Expired due to Fees

• Current Assignee: The Boeing Co.
• Original Assignee: The Boeing Co.
• Inventors: Holt, Fredrick B., Nguyen, Dziem D.
• Primary Examiner(s): Smith, Jerry
• Assistant Examiner(s): TRAMMELL, JAMES P
• Application Number:
  US07/305,962
• Time in Patent Office: 657 Days
• Field of Search: 364/148, 364/807, 364/513, 364/713, 364/602, 365/215, 365/125, 307/201
• US Class Current: 708/801
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks

Abstract

An apparatus and method for implementing a neural network having N nodes coupled to one another by interconnections having interconnect weights T_ij that quantify the influence of node j on node i. The apparatus comprises a node circuit for each node and a data processor. The data processor receives one or more library members, and transmits the interconnect weights to the node circuits. The data processor also stores a current state vector, and receives input data representing a library member to be retrieved. The data processor then performs an iteration in which the current state vector is sent to the node circuits, and an updated state vector is received from the node circuits, the iteration being commenced by setting the current state vector equal to the input data. Each node circuit comprises one or more stochastic processors for multiplying the state vector elements by the corresponding interconnect weights, to determine the updated state vector. Each stochastic processor preferably includes means for generating a pseudorandom sequence of numbers, and using such sequence to encode the interconnect weights and state vector elements into stochastic input signals that are then multiplied by a stochastic multiplier comprising delay means and an AND gate.

Claims

1. An apparatus for implementing a neural network having N nodes coupled to one another by interconnections having interconnect weights T_ij that quantify the influence of node j on node i, the apparatus comprising:

a plurality of node circuits, one associated with each node;

data processing means including means for receiving library data representing one or more library members, means for determining interconnect weights T_ij from the library data and for transmitting interconnect weight data representing the interconnect weights to each node circuit, means for storing current state vector data representing a current state vector comprising N state vector elements, means for receiving input data representing a library member to be retrieved, means for performing one or more iterations until the current state vector reaches an equilibrium state, each iteration comprising transmitting the current state vector data to the node circuits, receiving updated state vector data representing an updated state vector from the node circuits, and setting the current state vector data equal to the updated state vector data, and means for initializing the iteration by setting the current state vector data equal to the input data; and

each node circuit comprising means for receiving the interconnect weight and current state vector data, and one or more stochastic processors operating in parallel, each stochastic processor including means for multiplying data representing one interconnect weight by data representing a corresponding state vector element to produce a product signal, means for summing the product signals to thereby determine the updated state vector data, and means for transmitting the updated state vector data to the data processing means.

2. The apparatus of claim 1, wherein each stochastic processor includes means for generating a pseudorandom sequence of numbers, means for using said pseudorandom sequence to encode data representing an interconnect weight and data representing a state vector element into first and second stochastic input signals respectively, and a stochastic multiplier for receiving the stochastic input signals and producing a stochastic output signal representing the product of the interconnect weight and the state vector element.

3. The apparatus of claim 2, wherein each stochastic multiplier comprises delay means for delaying the first stochastic input signal to produce a delayed first stochastic input signal, and an AND gate for performing a logical AND operation on the delayed first stochastic input signal and the second stochastic input signal to produce the stochastic output signal.

4. The apparatus of claim 1, wherein the data processing means includes means for determining bias data for each node from the interconnect weight data and the input data, and means for transmitting the bias data to the respective node circuits, and wherein each node circuit comprises means for summing the product signals and the bias data to thereby determine the updated state vector data.

5. The apparatus of claim 1, wherein each node circuit comprises N stochastic processors, each stochastic processor including means for multiplying data representing a predetermined state vector element by data representing a corresponding interconnect weight, each node circuit comprises means for causing the stochastic processors to operate such that N product signals are produced in parallel.

6. The apparatus of claim 5, wherein each stochastic processor includes means for generating a pseudorandom sequence of numbers, means for using said pseudorandom sequence to encode data representing an interconnect weight and data representing a state vector element into first and second stochastic input signals respectively, and a stochastic multiplier for receiving the stochastic input signals and producing a stochastic output signal representing the product of the interconnect weight and the state vector element.

7. The apparatus of claim 6, wherein each stochastic multiplier comprises delay means for delaying the first stochastic input signal to produce a delayed first stochastic input signal, and an AND gate for performing a logical AND operation on the delayed first stochastic input signal and the second stochastic input signal to produce the stochastic output signal.

8. The apparatus of claim 5, wherein the stochastic processors produce the respective product signals in stochastic form, and wherein the the means for summing the product signals includes an array of analog adders connected to the stochastic processors for summing the stochastic product signals to produce a stochastic summation signal, and means for integrating the stochastic summation signal to produce the updated state vector data.

9. The apparatus of claim 5, wherein the stochastic processors produce the respective product signals in stochastic form, and wherein the means for summing the product signals comprises counting means for counting each stochastic product signal to produce a corresponding nonstochastic digital count signal, and an array of digital adders for adding the digital count signals to produce the updated state vector data.

10. The apparatus of claim 5, wherein the stochastic processors produce the product signals in optical form, and wherein the means for summing the product signals comprises a CCD element for receiving and summing the optical product signals to produce an electronic summation signal, and means for integrating the electronic summation signal to produce the updated state vector data.

11. The apparatus of claim 5, wherein the stochastic processors produce the respective product signals in optical form, and wherein the means for summing the product signals comprises a photodetector for receiving each product signal and producing a corresponding electrical signal, and digital adder means for adding the electrical signals to produce the update state vector data.

12. The apparatus of claim 1, wherein either the interconnect weight or the state vector elements comprise bipolar operands and the other of the interconnecting weights or state vector elements not comprising bipolar operands comprises unipolar operands, wherein each stochastic processor includes conversion means, first and second stochastic multiplies connected to the conversion means, and a subtractor coupled to the stochastic multipliers, the conversion means for converting the bipolar operand into a positive signal and a negative signal, the positive signal being equal to the bipolar operand if the bipolar operand is greater than zero and otherwise being equal to zero, the negative signal being equal to the absolute value of the bipolar operand if the bipolar operand is less than zero and equal to zero otherwise, the first stochastic multiplier including means for multiplying the positive signal by the unipolar operand to produce a first intermediate signal, the second stochastic multiplier including means for multiplying the negative operand by the unipolar operand to produce a second intermediate signal, and the subcontractor including means for subcontracting the second intermediate signal from the first intermediate signal to produce the product signal.

13. A method for implementing a neutral network having N nodes coupled to one another by interconnections having interconnect weights T_ij that quantify the influence of node j on node i, the method comprising:

providing data processing means and a plurality of node circuits, one node circuit for each node;

at the data processing means, receiving library data representing one or more library members, determining interconnect weights T_ij from the library data and transmitting interconnect weight data representing the interconnect weights to each node circuit;

storing current state vector data representing a current state vector comprising N state vector elements, receiving input data representing a library member to be retrieved, performing one or more iterations until the current state vector reaches an equilibrium state, each iteration comprising transmitting the current state vector data to node circuits, receiving updated state vector data representing an updated state vector from the node circuits, setting the current state vector data equal to the updated state vector data, and initializing the iteration by setting the current state vector data equal to the input data; and

at each node circuit, receiving the interconnect weight and current state vector data, stochastically multiplying in parallel the interconnect weight data and the state vector data to produce product signals, summing the product signals to thereby determine the updated state vector data, and transmitting the updated state vector data to the data processing means.