Bias Term Adjustment in Deep Neural Networks for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural networks face challenges in efficiently classifying objects with high computational complexity and sensitivity, particularly when there is prior knowledge that an object has an increased probability of being present in the input, leading to false positives and incorrect classifications.
Innovation Solution
The method involves adjusting the bias of activation functions in deep neural networks without altering the weights, to increase sensitivity to elements with a higher probability of presence, thereby refining the output based on prior knowledge without changing the filter weights or causing hallucinations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional deep neural networks are used for object classification, then the network can process inputs through standard activation functions, but the computational complexity becomes prohibitively high and sensitivity to prior knowledge is insufficient
Solution Approach 1:
The patent changes the bias parameter of activation functions dynamically based on prior knowledge about object probability. By adjusting the bias term rather than retraining all weights, the system incorporates prior knowledge efficiently without prohibitively increasing computational complexity, thereby improving classification reliability while managing device complexity
2Measurement precision
If the network increases sensitivity to elements with higher probability of presence, then classification accuracy improves, but false positives may increase if prior knowledge is incorrectly applied
Solution Approach 1:
The patent applies dynamic adjustment of bias terms based on the specific input and prior knowledge relevance. The bias is not fixed but adapted dynamically - increasing sensitivity only when prior knowledge appropriately applies, thereby improving detection sensitivity while controlling false positive rates through conditional application
3Reliability
If weights of the network are adjusted to incorporate prior knowledge, then sensitivity to known objects increases, but the trained filter weights may be altered causing hallucinations
Solution Approach 1:
The patent extracts and applies only the bias component separately from the trained weights. By taking out the bias term as an independent adjustable parameter and modifying only that while preserving the originally trained weights, the system increases detection sensitivity without altering the integrity of the training and causing hallucinations
Data Source
AI summary
A method of biasing a deep neural network includes determining whether an element has an increased probability of being present in an input to the network. The method also includes adjusting a bias of activation functions of neurons in the network to increase sensitivity to the element. In one configuration, the bias is adjusted without adjusting weights of the network. The method further includes adjusting an output of the network based on the biasing.