Neural network system and circuit for use therein

• US 4,994,982 A
• Filed: 12/22/1988
• Issued: 02/19/1991
• Est. Priority Date: 12/23/1987
• Status: Expired due to Fees

• Current Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Original Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Inventors: Gobert, Jean, Duranton, Marc, Sirat, Jacques-Ariel
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/289,648
• Time in Patent Office: 789 Days
• Field of Search: 364/300, 364/513, 364/200, 364/900, 307/201
• US Class Current: 706/41
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A neural network system comprises a memory for storing in binary code the synaptic coefficients indicative of the interconnections among the neurons. Means are provided for simultaneously supplying all the synaptic coefficients associated with a given neuron. Digital multipliers are provided for determining the product of the supplied synaptic coefficients and the relevant neuron states of the neurons connected to said given neuron. The multipliers deliver their results into an adder tree for determining the sum of the products. As a result of the parallel architecture of the system high operating speeds are attainable. The modular architecture enables extension of the system.

Claims

1. A neuronal circuit for a neural network system which circuit enables digital determination of new states of M neurons by determining products of states of N input neurons and M input vectors of N synaptic coefficients which are associated with each interconnection between neurons and which are stored in a programmable memory,

characterized in that the neuronal circuit comprises an integrated circuit for performing for a given neuron chosen from the M neurons, parallel processing of all synaptic coefficients of the given neuron and which comprises:

(a) the programmable digital memory which is integrated in the circuit for storing the N×

M synaptic coefficients of all M neurons encoded on p bits, and for simultaneously supplying all synaptic coefficients of the given neuron,

(b) N digital multipliers for determining the products of the N synaptic coefficients of the given neuron and the states of the N input neurons,

(c) an adder tree for determining the sum S of all said products for the given neuron, and

(d) a digital state memory for receiving the states of the N input neurons which may be encoded from 1 to n bits in a reconfigurable mode, the circuit being controlled by control means,

where N, M, p, and n represent positive integers and S represents a number.

2. A circuit as claimed in claim 1, characterized in that the N multipliers are formed by logic AND-gates.

3. A circuit as claimed in claim 2, characterized in that the state of a neuron is encoded on n=1 bit, the state memory comprising N storage elements, the determination of the sum S taking place during a single control cycle after the reading of the synaptic coefficients.

4. A circuit as claimed in claim 2, characterized in that the state of each neuron is encoded on n bits, the state memory comprising N×

n storage elements, each neuron state (or each relevant synaptic coefficient) being read by means of an associated selector from a group A of selectors which successively selects the n bits of the neuron state (or said relevant synaptic coefficient) in order to supply a signal EM which is applied to a multiplier comprising logic AND-gates, the group A of selectors being controlled by the control means via a signal SELA, the adder tree comprising, connected to its output, an adder and a shift accumulator so that the sum S is determined in n (or p) shift control cycles after the start of the reading of the synaptic coefficients in the programmable memory.

5. A circuit as claimed in claim 4, characterized in that in operating with signed synaptic coefficients, the adder tree comprises adders for operating on the sign bits of the N products.

6. A circuit as claimed in claim 5, characterized in that the neuron states are encoded in 2'"'"'s-complement, the adder tree comprising, connected to its output, an adder/subtractor and a shift accumulator, the "substract" operation being performed at the instant at which the sign bit of the neuron state is selected and if this sign bit represents a negative value.

7. A circuit as claimed in claim 5 characterized in that

the neuron states are encoded as logic levels K=+1/-1, the circuit further comprising N switches, controlled by a selection signal (SEL), for enabling reading of the synaptic coefficients from the programmable memory as they are stored therein when K=+1 or in inverted logic levels when K=-1.

8. A circuit as claimed in claim 7, characterized in that each neuron state is encoded in 2'"'"'s-complement, the state memory comprising N.n storage elements, a selector of a group A of N selectors successively selecting the n bits (B_n . . . B₁), starting with the least significant bit B₁, in order to produce upon each selection a signal EM which is applied to the multiplier ET relating to the neuron state, the group A of selectors being controlled by the control means, using a signal SELA, each synaptic coefficient being read by means of one of the N switches controlled by a respective signal SEL, the values or the logic levels of the signals EM and SEL being determined so that the signal SEL copies the sign bit when it is selected and is zero when the other bits are selected, and so that the signal EM copies all bits when they are selected, the adder tree comprising, connected to its output, an adder and a shift accumulator so that the sum S is determined in n shift control cycles after the reading of the synaptic coefficients in the programmable memory.

9. A circuit as claimed in the claim 7, characterized in that each neuron state is encoded on n bits in 2'"'"'s-complement, the most-significant bit being the sign bit, the state memory comprising N.n storage elements, each switch associated with each neuron state being controlled by a signal SEL supplied by a selector from a group B of selectors which successively select the first until the n^th bit of the corresponding neuron state, starting with the least-significant bit, simultaneously each selector from a group A of selectors successively selecting, shifted through one bit with respect to the group B, a logic value 0 until the (n-1)^th bits of the neuron states, each output EM₁ of the selector of the group A and each output SEL of the selector of the group B being applied to an exclusive-OR-gate which supplies the signal EM applied to the logic AND-gate multipliers, the group A and the group B of selectors being controlled by the control circuit, via the signals SELA, SELB, respectively, the adder tree comprising, connected to its output an adder and a shift accumulator so that the sum S is determined in n shift control cycles after the reading of the synaptic coefficients in the programmable memory.

10. A circuit as claimed in claim 2, characterized in that the neuron states are encoded as logic levels K=+1/-1, the circuit further comprising N switches, controlled by a selection signal (SEL), for enabling reading of the synaptic coefficients from the programmable memory as they are stored therein when K=+1 or in inverted logic levels when K=-1.

11. A circuit as claimed in any one of the claim 1, characterized in that it comprises two state memories so that one memory stores the input neuron states and the other memory stores the new neuron states, the latter states being the relevant states when the renewal of all input neuron states has been effected.

12. A circuit as claimed in any one of the claim 1, characterized in that the number of neurons N at the input and the number of neurons M at the output are equal, thus forming a fully connected neural network.

13. A neuronal device, comprising a neuronal circuit as claimed in claim 1 and comprising a digital converter element which subjects the sum S to a programmable function and which supplies results which enable updating of the state of the given neuron.

14. A device as claimed in claim 13, characterized in that the converter element is either a processor or a digital memory.

15. A neuronal arrangement comprising n devices as claimed in claim 13, each of which device is provided with a digital state memory which stores 1 bit of each state, the states being encoded on n bits, the devices being connected in parallel in order to increase a processing speed.

16. A neural network system comprising a device as claimed in claim 13, controlled by a central processor.

17. A circuit as claimed in any one of the claim 1, characterized in that it also comprises a digital converter element which subjects the sum S to a programmable function and which supplies results enabling the updating of the state of the given neuron.

18. A circuit as claimed in claim 17, characterized in that the converter element is a digital memory.

19. A neuronal arrangement comprising a device as claimed in claim 13, characterized in that it comprises k=k₁ ×

k₂ devices capable of storing k×

N×

M synaptic coefficients organised in k₁ ×

M neurons of k₂ ×

N synaptic coefficients.

20. An arrangement as claimed in claim 19, characterized in that it comprises at least two devices which are connected in parallel for operating on synaptic coefficients encoded on an increased number of bits.

21. A neuronal arrangement comprising a circuit as claimed in claim 17, characterized in that it comprises k=k₁ ×

k₂ circuits for storing k×

N×

M synaptic coefficients organised in k₁ ×

M neurons of k₂ ×

N synaptic coefficients.

22. An arrangement as claimed in claim 21, characterized in that it comprises at least two circuits which are connected in parallel in order to operate on synaptic coefficients encoded on an increased number of bits.

23. A neuronal arrangement comprising n circuits as claimed in claim 17, each of which circuit is provided with a digital state memory which stores 1 bit of each state, the states being encoded on n bits, the circuits being connected in parallel in order to increase a processing speed.

24. A third-order neural network system, comprising N circuits or devices or N second-order arrangements as claimed in claim 1, 17, 13, 21, or 23, first outputs of the N circuits or devices or arrangements (70₁ . . . 70_N) being multiplied by the N states of the neurons in N multipliers (71₁ . . . 71_N) whose second outputs are added in an adder tree (73), a converter element subjecting a third output of the converter element to a programmable function and supplying results enabling the updating of the state of each given neuron.

25. A neural network system of the order q which is higher than 3, wherein the second-order circuits or arrangements as claimed in claim 24 are replaced by systems of the order q-1.

26. A neural network system comprising a q order system as claimed in claim 25, controlled by a central processor.

27. A neural network system comprising a third order system as claimed in claim 24, controlled by a central processor.

28. A neural network system comprising a circuit as claimed in claim 17, controlled by a central processor.

29. A neural network system comprising an arrangement as claimed in claim 21, 22, 23, 19, 20, or 15, controlled by a central processor.