Programmable neuron for analog non-volatile memory in deep learning artificial neural network

Patent metadata

Field	Value
Publication number	US-11354562-B2
Application number	US-201815936983-A
Country	US
Kind code	B2
Filing date	Mar 27, 2018
Priority date	Jan 3, 2018
Publication date	Jun 7, 2022
Grant date	Jun 7, 2022

Abstract

Numerous embodiments for processing the current output of a vector-by-matrix multiplication (VMM) array in an artificial neural network are disclosed. The embodiments comprise a summer circuit and an activation function circuit. The summer circuit and/or the activation function circuit comprise circuit elements that can be adjusted in response to the total possible current received from the VMM to optimize power consumption.

First claim

What is claimed is:

1. An artificial neural network, comprising:
a vector-by-matrix multiplication array comprising rows and columns of flash memory cells;
a summer circuit for receiving a current from the vector-by-matrix multiplication array and for generating an output voltage in response to the received current, the summer circuit comprising an adjustable circuit element a variable resistor; and
a control system for adjusting the variable resistor in response to a set of bits, wherein the set of bits comprises one or more of configuration bits and trim bits.

2. The artificial neural network of claim 1 , further comprising:
an activation function circuit for receiving the output voltage from the summer circuit as an input and generating an output current in response to the output voltage.

3. The artificial neural network of claim 2 , wherein the activation function circuit executes a hyperbolic tangent function on the input to generate the output current.

4. The artificial neural network of claim 2 , wherein the activation function circuit executes a ReLU function on the input to generate the output current.

5. The artificial neural network of claim 2 , wherein the activation function circuit executes a Sigmoid function on the input to generate the output current.

6. The artificial neural network of claim 1 , wherein the variable resistor comprises discrete resistor elements.

7. The artificial neural network of claim 1 , wherein the variable resistor comprises a MOS transistor.

8. The artificial neural network of claim 1 , wherein the summer circuit comprises an operational amplifier.

9. The artificial neural network of claim 8 , wherein a bias for the operational amplifier is adjusted per vector matrix multiplier array size.

10. The artificial neural network of claim 9 , wherein the bias for the operational amplifier is based on the value of the variable resistor.

11. The artificial neural network of claim 1 , wherein the current from the vector-by-matrix multiplication array is provided on a bit line of the vector-by-matrix multiplication array.

12. The artificial neural network of claim 1 , wherein the current from the vector-by-matrix multiplication array is provided on a source line of the vector-by-matrix multiplication array.

13. The artificial neural network of claim 1 , wherein the flash memory cells are split gate memory cells.

14. An artificial neural network, comprising:
a vector-by-matrix multiplication array comprising rows and columns of flash memory cells;
a summer circuit for receiving a differential input current signal from the vector-by matrix multiplication array and for generating an output voltage in response to the differential input current signal, the summer circuit comprising a variable resistor; and
a control system for adjusting the variable resistor in response to a set of bits, wherein the set of bits comprises one or more of configuration bits and trim bits.

15. The artificial neural network of claim 14 , further comprising:
an activation function circuit for receiving the output voltage from the summer circuit and generating an output current in response to the output voltage.

16. The artificial neural network of claim 15 , wherein the activation function circuit executes a hyperbolic tangent function on the input to generate the output current.

17. The artificial neural network of claim 15 , wherein the activation function circuit executes a ReLU function on the input to generate the output current.

18. The artificial neural network of claim 15 , wherein the activation function circuit executes a Sigmoid function on the input to generate the output current.

19. The artificial neural network of claim 14 , wherein the summer circuit comprises an operational amplifier.

20. The artificial neural network of claim 19 , wherein a bias for the operational amplifier is adjusted per vector matrix multiplier array size.

21. The artificial neural network of claim 20 , wherein the bias for the operational amplifier is a relation to the value of the adjusted circuit element.

22. The artificial neural network of claim 14 , wherein the variable resistor comprises discrete resistor elements.

23. The artificial neural network of claim 14 , wherein the variable resistor comprises a MOS transistor.

24. The artificial neural network of claim 14 , wherein the current from the vector-by-matrix multiplication array is provided on a bit line of the vector-by-matrix multiplication array.

25. The artificial neural network of claim 14 , wherein the current from the vector-by-matrix multiplication array is provided on a source line of the vector-by-matrix multiplication array.

26. The artificial neural network of claim 14 , wherein the flash memory cells are split gate memory cells.

27. A programmable neuron for an artificial neural network, comprising:
a vector-by-matrix multiplication array comprising rows and columns of flash memory cells;
a neuron output circuit for receiving a current from the vector-by-matrix multiplication array and for generating an output in response to the received current, the neuron output circuit comprising an adjustable circuit element comprising a variable resistor; and a control system for adjusting the adjustable circuit element in response to a set of bits, wherein the variable resistor is configured by the set of bits and the set of bits comprises one or more of configuration bits and trimbits.

28. The programmable neuron of claim 27 , wherein the neuron output circuit comprises an activation function circuit for receiving the output voltage from a summer circuit as an input and generating an output current in response to the output voltage.

29. The programmable neuron of claim 28 , wherein the activation function circuit executes a hyperbolic tangent function on the input to generate the output current.

30. The programmable neuron of claim 28 , wherein the activation function circuit executes a ReLU function on the input to generate the output current.

31. The programmable neuron of claim 28 , wherein the activation function circuit executes a Sigmoid function on the input to generate the output current.

32. The programmable neuron of claim 27 , wherein the variable resistor comprises discrete resistor elements.

33. The programmable neuron of claim 27 , wherein the variable resistor comprises a MOS transistor.

34. The programmable neuron of claim 27 , wherein the summer circuit comprises an operational amplifier.

35. The programmable neuron of claim 34 , wherein a bias for the operational amplifier is adjusted per vector matrix multiplier array size.

36. The programmable neuron of claim 35 , wherein the bias for the operational amplifier is based on the value of the adjusted circuit element.

37. The programmable neuron of claim 27 , wherein the current from the vector-by-matrix multiplication array is provided on a bit line of the vector-by-matrix multiplication array.

38. The programmable neuron of claim 27 , wherein the current from the vector-by-matrix multiplication array is provided on a source line of the vector-by-matrix multiplication array.

39. The programmable neuron of claim

Assignees

• Silicon Storage Tech Inc

Inventors

• Tran Hieu Van
• Hong Stanley
• Ly Anh
• Vu Thuan
• Pham Hien
• Nguyen Kha
• Tran Han

Classifications

• G06N3/065 Primary

  Analogue means · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06F17/16

  Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)} · CPC title