Temporal learning neural network

• US 5,704,016 A
• Filed: 03/17/1995
• Issued: 12/30/1997
• Est. Priority Date: 03/23/1994
• Status: Expired due to Fees

• Current Assignee: Kabushiki Kaisha Square Enix
• Original Assignee: Kabushiki Kaisha Square Enix
• Inventors: Matsumoto, Gen, Shigematsu, Yukifumi
• Primary Examiner(s): Hafiz, Tariq R.
• Assistant Examiner(s): Rhodes, Jason W.
• Application Number:
  US08/405,963
• Time in Patent Office: 1,019 Days
• Field of Search: 395/27, 395/24, 395/21, 395/11, 395/22, 395/23
• US Class Current: 706/41
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A temporal learning neural network includes a plurality of temporal learning neural processing elements and an input/output control section. Each element includes a calculation device and a learning device. The calculation device includes an input memory section and a response calculation circuit. The learning device includes a learning processing circuit and a history evaluation circuit. The calculation circuit calculates a sum of a total summation value of a product of input values and connection efficacies, and an internal potential, compares the sum with a predetermined threshold value, outputs a 1 or 0 signal depending on the comparison and substitutes internal potential of a next time for the sum. The processing circuit receives an input history evaluation value when the calculation circuit has produced an output 1 signal which strengthens, weakens or leaves unchanged the connection efficacies depending on the comparison. The evaluation circuit obtains an input history value, compares the obtained input history value with the learning threshold value, generates an evaluation signal and distributes the evaluation signal to the input memory section. The input/output control section is provided with input terminals and output terminals, sends signals input from the calculation circuit and evaluation circuit to the input memory section, receives signals output from the calculation circuit and evaluation circuit, and effects communication with each of the processing elements. This process is an input temporal associative learning process.

Claims

1. A temporal learning neural network for implementing input temporal associative learning, the neural network comprising:

a substrate,

a plurality of temporal learning neural processing elements integrated on said substrate, and

an input/output control section including input terminals and output terminals;

each of said processing elements comprising;

learning means, and

calculation means comprising;

an input memory section with inputting means for inputting a plurality of signal pulses, and means for storing the signal pulses input from said inputting means, and

a response calculation circuit which calculates a sum of (a) a total summation value of a product of (i) input values to said input memory section and (ii) corresponding input terminal connection efficacies, and (b) an internal potential remaining in the processing element at a current time, which is obtained by decaying the internal potential of the processing element at a preceding time, the sum forming an accumulated value, said response calculation circuit including comparing means for comparing the accumulated value with a predetermined threshold value and, (i) when the accumulated value exceeds the predetermined threshold value, outputting an output 1 signal, and storing, as the internal potential for the processing element at a subsequent time, a value obtained by deducting a constant from the accumulated value, and (ii) when the accumulated value does not exceed the predetermined threshold value, outputting an output 0 signal and storing the accumulated value as the internal potential of the processing element of the subsequent time, said learning means comprising;

a history evaluation circuit for determining an input history evaluation value, and

a learning processing circuit for generating new connection efficacies in a processing element wherein said calculation circuit has output an output 1 signal, by receiving an input history evaluation value from said history evaluation circuit, the input history evaluation value (i) strengthening the connection efficacies when the evaluation value is positive, (ii) weakening the connection efficacies when the evaluation value is negative and (iii) leaving the connection efficacies unchanged when the evaluation value is zero, wherein said history evaluation circuit includes;

means for using an input history value of the processing element of the preceding time for the input history value of the processing element of a current time, which is obtained by accumulating the input signals with decay,

means for comparing the input history value with an enhancement learning threshold value and a decreased learning threshold value, generating the input history evaluation value whereby a changing direction for the connection efficacies is (i) positive when the obtained input history value is equal to or higher than the enhancement learning threshold value, (ii) negative when the obtained input history value is lower than the enhancement learning threshold value and is equal to or higher than the decreased threshold value, or (iii) zero when the obtained input history value is lower than the decreased threshold value, and

means for distributing the input history evaluation value to the input memory section of each of the processing elements, and

wherein said input/output control section includes

means for receiving at said input terminals (i) the output 1 and output 0 signals output from said response calculation circuit and (ii) the input history evaluation value from said history evaluation circuit, and

means for sending, from said output terminals, the signals input from said response calculation circuit and from said history evaluation circuit, to said input memory section of each of said plurality of processing elements.

2. A neural network according to claim 1, further comprising:

a plurality of substrates on each of which said plurality of processing elements have been integrated, and wherein temporal associative learning is applied to said processing elements between said plurality of substrates to form an associative memory system for storing temporal relationships; and

means for training selected output processing elements using at least one of teaching and induction signals.

3. A neural network according to claim 1, wherein said plurality of processing elements are divided into a plurality of groups having input terminals and output terminals connected respectively to said output terminals and said input terminals of said input/output control section.

4. A neural network according to claim 1, wherein said input/output control section further comprises means for connecting to an external circuit.