Neural Network Training Data Culling via Neuron Activation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for road geometry modeling and object detection in autonomous vehicles are resource-intensive and time-consuming, often relying on unreliable feature detection from image data, which can lead to safety concerns and inefficiencies due to inaccurate object identification.
Innovation Solution
A method for automatically identifying the most informative training data examples for a neural network by processing sensor data, activating neurons, and updating the network with manually labeled images, thereby reducing redundant data and improving object detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional feature detection methods are used for object identification, then object detection can be performed, but the reliability is low and manual measurement is required which increases time and cost
Solution Approach 1:
The neural network performs automatic object detection and identification without requiring manual measurement or intervention. The system serves itself by learning from training data to automatically detect objects, their types, and locations, eliminating the need for human operators to manually measure and identify objects in images.
Solution Approach 2:
The patent replaces manual mechanical measurement methods with an automated neural network-based detection system. The neural network processes images computationally to identify objects, substituting the mechanical process of manual measurement and calculation with an automated information processing system.
2Reliability
If all training data is used for neural network training, then comprehensive coverage is achieved, but redundant data increases processing time and computational resources
Solution Approach 1:
The system extracts and identifies informative training images from the complete training dataset by analyzing neuron activation patterns. Only the most valuable images that provide new information to the neural network are selected for manual labeling and inclusion in the training set, while redundant images are excluded.
Solution Approach 2:
The patent applies different selection criteria to different portions of the training data based on their informativeness. Rather than treating all training images uniformly, the system identifies and processes only those images with high informational value (those that activate sufficient neurons), applying quality-based differentiation to the training dataset.
3Measurement precision
If manual labeling is performed on all training images, then accurate object identification is achieved, but the cost and time consumption increase significantly
Solution Approach 1:
The system extracts only the most informative training images based on neuron activation analysis and directs manual labeling efforts solely to these selected images. This extraction approach ensures that manual labeling resources are concentrated on images that will provide the maximum benefit to training accuracy, rather than labeling all images.
Solution Approach 2:
Instead of performing manual labeling on the complete training dataset, the system performs partial labeling only on the subset of images identified as informative. This partial action approach achieves sufficient training accuracy by focusing labeling efforts on the most critical images, rather than exhaustively labeling every image.
Data Source
AI summary
Methods described herein relate to culling training data for machine learning in a neural network. Methods may include: receiving sensor data from at least one image sensor, where the sensor data is representative of at least one image; processing the at least one image using a neural network; identifying a number of neurons of the neural network that are activated for each of the at least one image; identifying an image as an informative training image in response to the number of neurons activated for the respective image satisfying a predetermined value; and updating the neural network with the informative training image in response to at least one object of the informative training image being manually labeled. The at least one image may be a plurality of images, where the predetermined value may be established based on a distribution of a number of neurons fired for the plurality of images.