Deep Neural Network Activation Thresholds for Feature Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing personal assistant systems using artificial neural networks struggle to detect alternative features in a physical environment that differ from those they were trained on, limiting the extension of rules and functions to related but untrained features.
Innovation Solution
A deep neural network is employed with relaxed activation thresholds at intermediate nodes, allowing it to detect alternative features by satisfying original activation thresholds, enabling the extension of rules and functions to related features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep neural network is trained to detect specific features, then detection accuracy for those features is improved, but the ability to detect alternative or related features deteriorates
Solution Approach 1:
The patent modifies activation parameters (thresholds) of intermediate nodes in the neural network to enable detection of alternative features. By changing these parameters, the network can recognize features similar to but different from those trained on, resolving the contradiction between specialized detection accuracy and general adaptability
2Adaptability or versatility
If activation thresholds are relaxed to detect alternative features, then adaptability is improved, but detection precision for original features deteriorates
Solution Approach 1:
The patent applies different activation thresholds to different nodes within the neural network. Intermediate nodes use relaxed thresholds to detect alternative features, while output nodes maintain strict thresholds to ensure accurate detection of original features. This localized differentiation resolves the contradiction between adaptability and precision
3Adaptability or versatility
If the neural network is extended to detect alternative features, then functionality is improved, but system complexity increases
Solution Approach 1:
The patent makes the existing neural network multi-functional by enabling it to detect both originally trained features and alternative features through parameter modification. This approach avoids the complexity of training separate networks for different features, as one network performs multiple detection functions
Data Source
AI summary
Sensor data is provided to a deep neural network previously trained to detect a feature within the physical environment. Result signals are received from the neural network, and the computing system determines if the feature is present within the physical environment based on the result signals. Responsive to determining that the feature is present, the computing system implements a function of a rule assigned to the feature. Responsive to determining that the feature is not present, the computing system determines whether one or more activation parameters of the neural network have been met indicative of an alternative feature being present within the physical environment. An indication that the activation parameters have been met is output by the computing system, enabling the rule to be extended to the alternative feature.