Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Discretized differentiable neural network search method based on entropy loss function

A search method and neural network technology, applied in neural learning methods, biological neural network models, neural architectures, etc., can solve the problems of small number of pruned edges, inaccuracy, and inability to guarantee discarded weights, etc., to reduce the loss of discretization accuracy , significant effect

Pending Publication Date: 2020-10-13
UNIVERSITY OF CHINESE ACADEMY OF SCIENCES
View PDF0 Cites 4 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0005] Moreover, and more importantly, during discretization, DARTS combines candidate operations and edges with a weighted sum (the weights are learnable), and keeps a fixed number of candidates with strong weights while discarding others, however, There is no guarantee that the dropped weights are relatively small
This discretization process introduces significant inaccuracies in the structure of each unit, and the accumulation of such inaccuracies eventually leads to a well-optimized supernetwork not necessarily producing high-quality subnetworks, especially when the discarded candidates are still With moderate weights, and / or relatively small number of pruned edges compared to supernets
Discretization leads to a sharp drop in the accuracy of the supernetwork and also hurts the performance of the search structure in the retraining phase

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
  • Discretized differentiable neural network search method based on entropy loss function
  • Discretized differentiable neural network search method based on entropy loss function
  • Discretized differentiable neural network search method based on entropy loss function

Examples

Experimental program
Comparison scheme
Effect test

Embodiment approach

[0029] According to a preferred embodiment of the present invention, the constructed search space is a unit-based hypernetwork search space, denoted as O, where each element is a fixed operation, denoted as o(*).

[0030] In a further preferred embodiment, the supernetwork is composed of 8 cell structures stacked, including 6 normal cells and 2 reduction cells;

[0031] The initial channel number of each unit is 16, including 6 nodes, and the alternative operations for node connection include 7.

[0032] Preferably, the operation is 3x3 and 5x5 atrous separable convolution, 3x3 and 5x5 separable convolution, 3x3 average pooling, 3x3 maximum pooling and cross-layer connection.

[0033] Among them, within each unit, the purpose of the search is to determine an operation for each pair of nodes.

[0034] In the present invention, as figure 1 As shown, record (i, j) as a pair of nodes, where 0≤i≤j≤N-1, N is the number of input edges reserved for each node;

[0035] According to ...

Embodiment 1

[0134] 1. Database:

[0135] The commonly used CIFAR10 and ImageNet datasets are used to evaluate the network architecture search method described in the present invention. Among them, CIFAR10 consists of 60,000 images with a spatial resolution of 32×32. The images are evenly distributed across 10 categories, with 50,000 training images and 100,000 test images; ImageNet contains 1,000 categories, including 1.3 million high-resolution training images and 50,000 validation images. These images are uniformly distributed across the class.

[0136] According to the commonly used settings, the mobile setting is adopted. In the test phase, the input image size is fixed at 224×224, and the structure is searched on CIFAR10 and then migrated to the ImageNet dataset.

[0137] 2. Compare the classification errors of the network structures searched by various search methods in the present invention and the prior art on the CIFAR10 data set, and the results are as shown in Table 1:

[01...

experiment example 1

[0155] Under different target network configurations, the DARTS of prior art and the search method described in the present invention (DA 2 S) search results are compared, and the results are shown in Table 4.

[0156] Table 4

[0157]

[0158] It can be seen from Table 4 that under different configurations, DARTS has a great loss of precision in the discretization process, while the search method of the present invention has a great improvement in the loss of precision, decreasing from [77.75-78.00] to [0.21-21.29].

[0159] further, Figure 10 Shows the change curve of the softmax value of the operation weight of the method DARTS in the standard unit on CIFAR10 during the search process; Figure 11 Shows the change curve of the softmax value of the operation weight in the descending unit of the method DARTS on CIFAR10 during the search process; Figure 12 It shows the network structure searched when the method DARTS is configured on CIFAR10 to select 3 out of 14 items...

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

The invention discloses a discretized differentiable neural network search method based on an entropy loss function. According to the method, a new loss term is designed into a constraint loss term suitable for different target network structure configurations based on an entropy function according to the characteristics of system entropy minimization driving system element (weight) sparsity and discretization so as to reduce discretization errors. According to the discretized differentiable neural network search method based on the entropy loss function, a discretized friendly target networkstructure is obtained through one-time search, and discretization precision loss existing in an existing search algorithm is greatly reduced; parameters of an entropy function-based structural constraint loss function may be modified to be adapted to search for arbitrarily configured network structures.

Description

technical field [0001] The invention belongs to the field of neural network structure search for automatic machine learning, and specifically relates to a discretized and differentiable neural network search method based on entropy loss function, which is used to eliminate discretization in the one-time differentiable neural network search method based on weight sharing. error. Background technique [0002] Network Architecture Search (NAS) aims to automatically search neural networks in a very large space not well covered by human expertise. To alleviate the computational burden of individually evaluating sampled network architectures based on reinforcement learning and evolutionary algorithms, the researchers propose a one-shot search method that first optimizes a supernetwork containing all possible architectures and then samples subnetworks from it for evaluation. This method speeds up NAS by 3 to 4 orders of magnitude through the shared weight mechanism. [0003] A ty...

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): G06N3/04G06N3/08G06F16/901G06F16/903
CPCG06N3/08G06F16/9024G06F16/903G06N3/045
Inventor 刘畅田运杰焦建彬叶齐祥
Owner UNIVERSITY OF CHINESE ACADEMY OF SCIENCES
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products