Neural Network Compression Using Impact Maps for Task-Specific Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Trained neural networks used in applications like autonomous driving are computationally intensive due to being trained for multiple tasks, necessitating a method to generate a compressed network capable of performing specific tasks with low computational cost, especially in real-time and safety-critical systems.
Innovation Solution
A method involving a first model portion generating an impact map to assess the importance of neural network components, followed by a second model portion creating a compressing map to delete unnecessary components, reducing weights and biases to near zero, and using a regularization term to increase compression, resulting in a task-specific efficient network architecture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a neural network is trained for multiple tasks to improve versatility, then adaptability is improved, but computational cost increases
Solution Approach 1:
The patent segments the neural network by identifying and removing unnecessary components for specific tasks. The method calculates importance values for individual network components and selectively removes those with low importance for the target task, dividing the network into essential and non-essential parts.
Solution Approach 2:
The patent extracts only the necessary components for the specific task from the multi-task neural network. By calculating importance values and comparing them against thresholds, the method extracts and retains only those components that contribute significantly to the target task performance.
2Adaptability or versatility
If a neural network is trained for multiple tasks to improve adaptability, then adaptability is improved, but device complexity increases
Solution Approach 1:
The patent segments the neural network architecture by identifying task-specific components. The method divides the network into task-relevant and task-irrelevant parts through importance calculation, allowing selective retention of only necessary components for the target task.
Solution Approach 2:
The patent extracts unnecessary network components for the specific task by calculating importance values and removing those below a threshold. This extraction process simplifies the network architecture while preserving task-specific performance.
3Speed
If network components are removed to reduce computational cost, then processing speed is improved, but measurement precision deteriorates
Solution Approach 1:
The patent applies local quality by assigning different importance values to different network components based on their specific contribution to the target task. The method selectively preserves components with high local importance while removing those with low importance, maintaining precision where needed while improving speed.
Solution Approach 2:
The patent changes the parameter of component importance through calculation and thresholding. By dynamically determining which components meet the importance threshold for the target task, the method adjusts the network configuration to optimize both speed and accuracy.
Data Source
AI summary
A device and a method for generating a compressed network from a trained neural network are provided. The method includes: a model generating a compressing map from first training data, the compressing map representing the impact of model components of the model to first output data in response to the first training data; generating a compressed network by compressing the trained neural network in accordance with the compressing map; the trained neural network generating trained network output data in response to second training data; the compressed network generating compressed network output data in response to the second training data; training the model by comparing the trained network output data with the compressed network output data.