Shared Neural Network Feature Extraction for Multi-Task Driving AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous and assisted driving systems face high computational loads due to the need to process diverse scenarios and tasks simultaneously, limiting the functions and features that can be provided.
Innovation Solution
The use of elastic weight consolidation (EWC) on a base convolutional neural network (CNN) allows for shared feature extraction across multiple tasks, reducing computational load and minimizing 'forgetting' of previously learned tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple separate neural networks are used for different tasks, then each task can be performed with high accuracy, but the computational load increases significantly
Solution Approach 1:
The patent merges multiple separate neural networks into a single shared neural network that performs multiple tasks simultaneously. The network architecture includes a shared base network followed by task-specific head networks, allowing multiple functions (object detection, lane finding, traffic light recognition) to be executed within one unified system, thereby reducing computational overhead while maintaining task accuracy
Solution Approach 2:
The base neural network is designed as a universal feature extractor that serves multiple tasks. The shared layers extract general features from sensor data that are then utilized by different task-specific output layers, enabling one network to perform multiple functions without requiring separate dedicated networks for each task
2Productivity
If a shared neural network is used for multiple tasks, then computational load is reduced, but the network may forget previously learned tasks
Solution Approach 1:
The neural network is segmented into distinct functional components: a shared base network for feature extraction and task-specific head networks for individual tasks. This segmentation allows the base network to learn shared representations while task-specific heads maintain specialized knowledge, preventing catastrophic forgetting through architectural separation
Solution Approach 2:
The patent employs pre-training and fine-tuning strategies where the base network is first trained on general features, then task-specific heads are added and trained sequentially. This preliminary action of establishing a robust base network before adding specialized tasks helps preserve previously learned capabilities while adapting to new tasks
3Reliability
If the neural network is fine-tuned for each task sequentially, then each task achieves optimal performance, but the weights are overwritten and previous tasks are forgotten
Solution Approach 1:
Different parts of the neural network have different qualities and functions: the base network maintains stable, shared feature extraction capabilities while task-specific head networks undergo localized fine-tuning. This local quality differentiation allows task-specific optimization without disrupting the globally shared knowledge stored in the base network
Solution Approach 2:
The patent extracts task-specific parameters and weights from the overall network into separate task-specific head networks. This extraction isolates task-specific knowledge from the shared base network, allowing independent optimization of each task without overwriting shared representations, thereby preventing catastrophic forgetting
Data Source
AI summary
A system includes a processor for performing one or more autonomous driving or assisted driving tasks based on a neural network. The neural network includes a base portion for performing feature extraction simultaneously for a plurality of tasks on a single set of input data. The neural network includes a plurality of subtask portions for performing the plurality of tasks based on feature extraction output from the base portion. Each of the plurality of subtask portions comprise nodes or layers of a neutral network trained on different sets of training data, and the base portion comprises nodes or layers of a neural network trained using each of the different sets of training data constrained by elastic weight consolidation to limit the base portion from forgetting a previously learned task.