Approximate calculation device and method suitable for neural network data and weight pre-classification

An approximate calculation and neural network technology, applied in the field of dynamic configuration approximate multiplication calculation, can solve the problems of blindness of adjustment range and results, inability to guarantee real-time calculation, long configuration time, etc., to achieve fast approximate multiplication calculation, reduce overhead, The effect of reducing the delay

Pending Publication Date: 2021-08-13
NANJING PROCHIP ELECTRONIC TECH CO LTD
View PDF0 Cites 1 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0005] There are two existing approximate multiplication dynamic configuration schemes: one method is to control the calculation accuracy by configuring the number of iterations, which requires repeated iterations, takes a lon

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Approximate calculation device and method suitable for neural network data and weight pre-classification
  • Approximate calculation device and method suitable for neural network data and weight pre-classification
  • Approximate calculation device and method suitable for neural network data and weight pre-classification

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0130] In this preferred embodiment, a group of binary numbers is selected as an example to further describe the device and method proposed by the present invention in detail. The selected 16-bit input data is '0001 0101 0110 0101', and the 8-bit weight parameter is '01010100'.

[0131] First, input the 8-bit weight parameter into the weight pre-classification program module. The number of consecutive 0s from the lowest bit of this weight parameter is 2, so it is recorded as '01', which is the second weight classification, and the meaning of classification is 'high precision' . Next, the bit width expansion unit in the weight pre-classification program module splices the 2-bit weight classification characterization parameter '01' into the high bit of the 8-bit original weight parameter, that is, expands its bit width to 10 bits, and the output new weight parameter is '0101010100' .

[0132] Subsequently, the approximate calculation of the hardware part is carried out, and all...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

According to the approximate calculation device and method suitable for neural network data and weight pre-classification, weight parameters of a neural network are pre-processed, fine classification is conducted according to the number of continuous' 0 'from the lowest bit, simple classification is conducted on input data according to the same features, and the two parties are combined to serve as an important basis for configuring an approximate scheme; approximate calculation arrays with different approximation degrees are configured by controlling approximate line positions of the full adder adder tree accumulation circuit and the low-order or approximate adder accumulation circuit. Aiming at the characteristic that weight parameters of a neural network model are often known and fixed, fine pre-classification processing is performed on the weight parameters, fine simple dichotomy processing is performed according to the characteristics of input system data, and approximate multiplication operation arrays with different configurations are dynamically selected by combining the fine pre-classification processing and the dichotomy processing. The selection of the approximation scheme is accurate, the characteristics of the corresponding data and weight are matched in real time, and the blindness, hysteresis and complexity of the dynamic configuration scheme of the existing approximation multiplication are overcome.

Description

technical field [0001] The present invention relates to the technical field of dynamic configuration approximate multiplication calculation, in particular to an approximate calculation device and method suitable for neural network data and weight pre-classification. Background technique [0002] In recent years, people pay more and more attention to energy-saving design. This is because a large number of applications require low-power custom designs, but the amount of data that these hardware needs to process has increased dramatically. In order to meet these two requirements at the same time, approximate computing has become A solution that works. [0003] With the increasing popularity of applications related to human perception, such as neural networks and machine learning, the amount of calculation of on-chip hardware is gradually increasing, but the accuracy does not require complete accuracy and correctness. This is because neural networks have natural Fault tolerance...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
IPC IPC(8): G06F7/533G06F17/17G06N3/063G06K9/62
CPCG06F7/5332G06F17/17G06N3/063G06F18/24
Inventor 王镇
Owner NANJING PROCHIP ELECTRONIC TECH CO LTD
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products