Neural Network Compression Through Temporal Activation Pruning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing size and computational demands of convolutional neural networks (CNNs) used in driver assistance systems and automated driving pose a challenge due to hardware constraints in vehicles, necessitating efficient pruning methods that minimize inference and training times while maintaining network stability and functionality.
Innovation Solution
A method for pruning CNNs by identifying unstable elements based on temporal activation differences, using time-varying input data to determine and remove elements with significant activation fluctuations, ensuring homogeneous pruning and retraining for improved stability and reduced network size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the size of the neural network is reduced by removing filters, then the inference time and computational effort are reduced, but the stability and functionality of the network may deteriorate
Solution Approach 1:
The patent applies preliminary action by computing temporal activation differences across multiple input frames before performing pruning. This allows the system to identify unstable filters based on their behavior over time before removal, ensuring that only filters with consistently low activation differences are pruned. The preliminary computation of stability metrics across temporal sequences enables informed pruning decisions that maintain network reliability while achieving compression.
2Quantity of substance
If more filters are removed to significantly reduce network size, then computational resources required are reduced, but the quality of output may deteriorate
Solution Approach 1:
The patent implements feedback by using temporal activation differences as a metric to guide the pruning process. The system continuously monitors filter activations across multiple input frames and uses this feedback information to identify which filters can be safely removed. This feedback mechanism ensures that filters contributing significantly to output quality (those with high activation differences) are preserved, while only filters with consistently low impact across temporal sequences are pruned, thereby maintaining output quality while reducing filter count.
3Device complexity
If pruning is performed without considering temporal stability, then the network size is reduced quickly, but the network may become unstable under varying input conditions
Solution Approach 1:
The patent applies dynamics by considering the temporal behavior of filter activations across multiple input frames. Instead of static pruning based on single-frame metrics, the system dynamically evaluates how filter activations change over time, capturing the dynamic response of the network to varying inputs. This dynamic assessment of temporal activation differences enables the identification of filters that are truly stable across different conditions, ensuring network stability is maintained while achieving compression.
Data Source
Figure 1~2
Figure 3
AI summary
The invention relates to a method for compressing a neural network (5), wherein the neural network (5) is supplied with temporally changing input data (7), wherein temporal activation differentials (20) of elements of the neural network (5) are determined for temporally adjacent input data, and wherein the determined temporal activation differentials (20) are provided in order to prune the neural network (5) according to the determined temporal activation differentials (20). The invention also relates to a device (1), a motor vehicle, a computer program and a data carrier signal.