CNN Stride and Dilation Control for Compute-Accuracy Tradeoffs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional frame skipping techniques for neural networks fail to achieve a flexible implementation to control the trade-off between computational amounts and accuracy with a single model, particularly in convolutional neural network models.
Innovation Solution
An information processing apparatus and method that transforms stride and dilation parameters using a transformation parameter to adjust the computational load and accuracy of convolutional neural networks, allowing a single model to operate on devices with varying computational capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If frame skipping is used to reduce computational amounts, then processing speed improves, but accuracy deteriorates
Solution Approach 1:
The patent applies dynamics by making the neural network model adaptable to different computational resources through parameter transformation. The system dynamically adjusts the model's stride and dilation parameters based on available computational resources, allowing the same model to operate efficiently on both high-performance and low-performance devices without sacrificing accuracy on capable devices or overwhelming limited resources.
Solution Approach 2:
The patent transforms parameters of the neural network model (specifically stride and dilation parameters) to adjust computational load. By changing these parameters through a transformation function that considers available computational resources, the system can reduce computational amounts on resource-constrained devices while maintaining the ability to achieve high accuracy on devices with sufficient computational power.
2Adaptability or versatility
If multiple neural network models are prepared for different computational resources, then adaptability improves, but device complexity increases
Solution Approach 1:
The patent achieves universality by creating a single neural network model that can function across multiple computational resource levels. Through parameter transformation, one model serves multiple purposes - it can be adapted to run on high-performance servers, mid-range devices, and low-power mobile devices without requiring separate model versions, thus eliminating the complexity of managing multiple models while maintaining broad adaptability.
Solution Approach 2:
The patent uses parameter transformation to enable a single model to adapt to different computational resources. By transforming the model's stride and dilation parameters based on available resources, the system achieves versatility across different device types without the complexity of maintaining separate models for each device category.
Data Source
AI summary
According to an embodiment, an information processing apparatus includes one or more hardware processors configured to function as a memory control unit, a transformation unit, a first convolutional neural network (CNN), and a second CNN unit. The memory control unit reads a first stride parameter used for controlling an output resolution and a first dilation parameter used for controlling an input resolution from a memory device. The transformation unit transforms the first stride parameter to a second stride parameter and transforms the first dilation parameter to a second dilation parameter by using a transformation parameter. The first CNN unit executes first CNN processing of a feature vector by using at least the second stride parameter. The second CNN unit executes second CNN processing with an output vector of the first CNN unit as an input by using at least the second dilation parameter.