Neural Network Compression via Layered Segmentation and Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for driver assistance and automated driving require special hardware due to limited computing and memory capacity, leading to performance degradation when trained neural networks are deployed on microcontrollers, necessitating specific training methods to achieve efficient compression of neural networks.
Innovation Solution
A method for compressing neural networks by globally optimizing weights and filters, distributing compression across layers to minimize performance loss, while considering interaction between weights and filters, thereby reducing calculation time, energy consumption, and memory requirements without needing special hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If neural networks are trained on supercomputers and then deployed on microcontrollers with limited computing and memory capacity, then the neural network can operate on embedded systems, but its performance capability decreases dramatically
Solution Approach 1:
The patent applies segmentation by dividing the neural network into individual layers and further into filters within each layer. The compression process treats each layer independently, allowing selective removal of filters based on their contribution to performance. This layered segmentation enables the network to be adapted for embedded systems while maintaining acceptable performance by keeping only the most important filters.
Solution Approach 2:
The patent implements local quality by applying different compression rates to different layers of the neural network. Instead of uniform compression, each layer can have its own reduction rate optimized based on its importance and characteristics. This allows critical layers to retain more filters while less critical layers can be more aggressively compressed, maintaining overall performance while reducing the total parameter count for embedded deployment.
2Productivity
If individual layers are compressed independently to reduce computing operations, then computing efficiency increases, but performance loss occurs due to ignoring interaction between layers
Solution Approach 1:
The patent employs feedback mechanisms by using performance metrics from validation data to guide the compression process. After removing filters from layers, the model is evaluated on validation data, and this feedback is used to adjust the compression strategy for subsequent layers. This iterative feedback loop ensures that computing efficiency is improved while performance loss is minimized by learning from actual model behavior at each compression step.
3Use of energy by moving object
If compression is applied to reduce parameters and multiplications, then memory requirement and energy consumption are reduced, but the number of parameters may exceed the predefined maximum complexity
Solution Approach 1:
The patent applies parameter changes by systematically varying the compression rate parameter across different layers. By adjusting this parameter, the method can control the degree of compression applied to each layer, thereby controlling the total number of parameters. This allows optimization of the trade-off between energy consumption (reduced by compression) and parameter count (constrained by maximum complexity requirements).
Data Source
AI summary
A method for compressing a neural network. The method includes: defining a maximum complexity of the neural network; ascertaining a first cost function; ascertaining a second cost function, which characterizes a deviation of a current complexity of the neural network in relation to the defined complexity; training the neural network in such a way that a sum of a first and a second cost function is optimized as a function of parameters of the neural network; and removing those weightings whose assigned scaling factor is smaller than a predefined threshold value.