Neural network compression device

Patent metadata

Field	Value
Publication number	US-12412097-B2
Application number	US-202017298801-A
Country	US
Kind code	B2
Filing date	Jan 8, 2020
Priority date	Jan 18, 2019
Publication date	Sep 9, 2025
Grant date	Sep 9, 2025

Abstract

Processing time of a neural network is shortened, and the number of operations of the neural network is reduced such that a plurality of calculators can be effectively used. A neural network reduction device ( 100 ) that reduces the number of operations of a neural network by an operation device ( 140 ) including a plurality of calculators by reducing the neural network, the neural network reduction device including: a calculator allocation unit ( 102 ) that sets the number of calculators allocated to calculation processing of the neural network; a number-of-operations setting unit ( 103 ) that sets the number of operations of a reduced neural network based on the number of allocated calculators; and a neural network reduction unit ( 104 ) that reduces the neural network such that the number of operations of the neural network by the operation device ( 140 ) is equal to the number of operations set by the number-of-operations setting unit ( 103 ).

First claim

The invention claimed is:

1. A neural network reduction device that reduces a number of operations of a neural network by an operation device including a plurality of calculators by reducing the neural network, the neural network reduction device comprising:
a calculator allocation unit that sets a number of the calculators allocated for calculation processing of the neural network, each of the calculators configured to execute a respective operation portion in each cycle of the calculation processing;
a number-of-operations setting unit that determines a number of operations of the neural network based on the number of the allocated calculators and sets the number of operations for the calculation processing;
a neural network reduction unit that reduces the neural network to generate a reduced neural network, such that the number of operations of the neural network is equal to the number of operations set by the number-of-operations setting unit;
an accuracy verification unit that
calculates accuracy of the reduced neural network, and
compares the accuracy with target accuracy,
wherein in response to the comparison, the calculator allocation unit decreases the number of the allocated calculators based on the accuracy being greater than or equal to the target accuracy or increases the number of the allocated calculators based on the accuracy being less than the target accuracy; and

a number-of-operations correction unit that updates the number of operations of the reduced neural network in response to increases or decreases in the number of the allocated calculators by the calculator allocation unit, wherein the updated number of operations is an integral multiple of the number of the allocated calculators.

2. The neural network reduction device according to claim 1 , wherein the number-of-operations setting unit sets the number of operations of the reduced neural network to be smaller than the number of operations of the neural network before reduction and be an integral multiple of the number of the allocated calculators set by the calculator allocation unit before the reduction.

3. The neural network reduction device according to claim 1 , wherein the number-of-operations setting unit sets the number of operations of the reduced neural network such that a remainder, obtained by dividing the number of operations of the neural network before reduction by the number of the allocated calculators set by the calculator allocation unit, is equal to or more than half of the number of the allocated calculators.

4. The neural network reduction device according to claim 1 , wherein
the neural network includes a plurality of layers,
the calculator allocation unit sets the number of the allocated calculators for each of the layers of the neural network, and
the number-of-operations setting unit sets the number of operations of the reduced neural network for each of the layers of the neural network.

5. The neural network reduction device according to claim 1 , wherein the neural network reduction unit reduces the neural network by pruning processing.

6. The neural network reduction device according to claim 1 ,
wherein the number-of-operations setting unit sets the number of operations of the reduced neural network to be small when the accuracy is equal to or higher than the target accuracy, and the number-of-operations setting unit sets the number of operations of the reduced neural network to be large when the accuracy is lower than the target accuracy.

7. A neural network reduction device that reduces a number of operations of a neural network by an operation device including a plurality of calculators by reducing the neural network, the neural network reduction device comprising:
a number-of-operations setting unit that sets a number of operations for calculation processing of the neural network;
a neural network reduction unit that reduces the neural network to generate a reduced neural network, such that the number of operations of the neural network is equal to the number of operations set by the number-of-operations setting unit;
a calculator allocation unit that sets a number of the calculators allocated for the calculation processing of the neural network, each of the calculators configured to execute a respective operation portion in each cycle of the calculation processing;
a number-of-operations correction unit that receives, from the calculator allocation unit, an indication of decreases in the number of the allocated calculators based on accuracy of the reduced neural network being greater than or equal to a target accuracy or increases the number of the allocated calculators based on the accuracy of the reduced neural network being less than the target accuracy; and
updates the number of operations of the reduced neural network in response to increases or decreases in the number of the allocated calculators set by the calculator allocation unit, wherein the updated number of operations is an integral multiple of the number of the allocated calculators.

Assignees

• Hitachi Astemo Ltd

Inventors

• Ito Hiroaki
• Ono Goichi
• Hirai Riu

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06N3/082 Primary

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/063

  using electronic means · CPC title

• G06N3/04 Primary

  Architecture, e.g. interconnection topology · CPC title