Multi-Layer Neural Network Model Training with Expanded Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep multi-layer neural network models, such as VGGNet, ResNet, and Xception, face challenges in resource-limited devices due to high calculation amounts and slow processing speeds, making them unsuitable for smartphones and hardware with limited memory, despite efforts like Network Pruning and Parameter Quantization, which either simplify the model but reduce performance or are difficult to apply effectively.
Innovation Solution
The method involves expanding the number of input feature maps or nodes in specific layers of the neural network during training and application, allowing for more weight connections and computations while maintaining storage efficiency by not requiring additional storage for expanded feature maps, and using gradient values from expanded maps to train the model effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high depth network models (VGGNet, ResNet, Xception) are used to enhance precision, then service precision is improved, but calculation amount increases and processing speed decreases
Solution Approach 1:
The patent segments the neural network into a backbone network and a detection head, where the backbone extracts features and the head performs detection. This segmentation allows the main computational burden to be separated from the detection function, enabling faster processing while maintaining precision through the specialized detection head architecture.
Solution Approach 2:
The patent introduces spatial dimension expansion by generating multiple scaled versions of the input image (1x, 2x, 3x, 4x) and processing them through different resolution paths. This dimensional transformation allows the network to capture both fine-grained and coarse-grained features simultaneously, improving detection precision without linearly increasing computational cost.
2Measurement precision
If network models are designed deeper and wider to enhance precision, then detection accuracy is improved, but model size increases making it unsuitable for resource-limited devices
Solution Approach 1:
The network is divided into a shared backbone that extracts common features across all scales and a separate detection head for each scale. This segmentation reduces redundancy and allows the model to achieve high accuracy with a more compact structure compared to a monolithic deep network.
Solution Approach 2:
The backbone network serves multiple functions by extracting features that are reused across different detection heads and scale levels. This multi-functionality reduces the total number of parameters needed while maintaining high detection accuracy across various object sizes.
3Productivity
If Network Pruning is applied to simplify the model, then calculation amount is reduced, but it is difficult to determine which filters are removable limiting practical application
Solution Approach 1:
The patent implements dynamic feature map generation where the number and scale of feature maps are adjusted based on the input image characteristics and detection requirements. This dynamic adaptation allows the model to automatically optimize its computational burden without requiring manual pruning or complex configuration.
Solution Approach 2:
The network performs preliminary feature extraction at multiple scales before the final detection stage, preparing feature representations in advance. This preliminary processing distributes the computational workload across different resolution levels, making the model more efficient and easier to apply in practice.
4Quantity of substance
If Parameter Quantization is applied to reduce storage space, then storage space is decreased, but reduction of indication precision causes network performance to worsen
Solution Approach 1:
The patent addresses the storage-performance tradeoff by operating in multiple resolution dimensions simultaneously. By processing feature maps at different scales (1x, 2x, 3x, 4x), the network captures comprehensive information that compensates for any precision loss from quantization, maintaining high performance while using efficient storage.
Data Source
AI summary
The present disclosure provides a training and application method of a multi-layer neural network model, apparatus and a storage medium. In a forward propagation of the multi-layer neural network model, the number of input feature maps is expanded and a data computation is performed by using the expanded input feature maps.