Deep Neural Network Confidence via Neuron Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning techniques used in vehicles for object classification and parameter estimation often lack access to ground truth data, leading to incorrect classifications and estimations in real-time due to the absence of absolute values.
Innovation Solution
A system that employs a deep neural network to generate a distribution of predictions based on sensor data, where neurons are disabled after each prediction to generate subsequent predictions with fewer neurons, allowing for the calculation of distribution variation and confidence parameter determination, enabling accurate classification and measurement of objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a deep neural network uses all neurons for prediction, then prediction speed is improved, but measurement precision deteriorates due to inability to assess confidence
Solution Approach 1:
The neural network is segmented into multiple groups of neurons that can be independently activated or deactivated. This allows the system to use different neuron subsets for different predictions, enabling both fast predictions (using all neurons) and confidence assessment (using selective neuron deactivation to generate multiple predictions).
Solution Approach 2:
The system performs preliminary actions by pre-arranging neuron groups and their activation patterns. Confidence assessment is prepared in advance through the structured approach of disabling specific neuron groups and generating multiple predictions, allowing the system to have confidence metrics ready before making final classification decisions.
2Measurement precision
If multiple predictions are generated with different neuron configurations, then measurement precision is improved through confidence assessment, but productivity deteriorates due to multiple processing steps
Solution Approach 1:
The neural network is divided into segments (neuron groups) that can be selectively activated. This segmentation allows the system to generate multiple predictions efficiently by activating different neuron segments rather than reprocessing the entire network, thus maintaining productivity while achieving confidence assessment.
Solution Approach 2:
The system uses partial action by activating only necessary neuron groups for each prediction scenario. Instead of fully processing all neurons for every prediction, the system selectively activates relevant neuron segments, reducing computational overhead while still generating multiple predictions for confidence assessment.
3Adaptability or versatility
If the neural network is trained with diverse training data, then adaptability is improved, but measurement precision deteriorates due to increased distribution variation
Solution Approach 1:
The system implements feedback by calculating distribution variation from multiple predictions and using this information to assess confidence. When training data diversity causes high distribution variation, the feedback mechanism detects this through confidence scoring, allowing the system to identify when predictions may be unreliable despite diverse training data, thus maintaining adaptability while monitoring precision.
Solution Approach 2:
The system dynamically adjusts its confidence assessment based on the actual distribution variation observed in predictions. Rather than treating all predictions equally, the system adapts its confidence metrics dynamically according to the measured variation, allowing it to handle diverse training data effectively while maintaining precision through context-aware confidence evaluation.
Data Source
AI summary
A distribution of a plurality of predictions generated by a deep neural network using sensor data is calculated, and the deep neural network includes a plurality of neurons. At least one of a measurement or a classification corresponding to an object is determined based on the distribution. The deep neural network generates each prediction of the plurality of predictions with a different number of neurons.