MOOD Cost Function for Neural Network Event Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information classification systems are complex, resource-intensive, and not practical for low-power devices, often compromising on flexibility and accuracy in applications like acoustic event detection and classification.
Innovation Solution
A new cost function, referred to as the Many or One Detection (MOOD) cost function, is introduced to train classifiers efficiently, allowing for flexible detection of events by tuning a hyperparameter to achieve desired classification results, such as single or multiple detections within a defined Region of Target (ROT), using a combination of forward and backward passes through the neural network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional classification systems are used, then classification accuracy can be maintained, but device complexity and resource consumption increase significantly
Solution Approach 1:
The patent segments the classification problem into two distinct phases: a training phase that computes optimal class assignments using the Hungarian algorithm, and an inference phase that simply applies these pre-computed assignments. This segmentation allows the complex optimization to be performed offline, resulting in a simple, low-complexity system for actual classification tasks while maintaining high accuracy through optimal matching.
Solution Approach 2:
The patent performs preliminary computation of optimal class assignments during the training phase using the Hungarian algorithm before the actual classification is needed. These pre-computed optimal assignments are then stored and directly applied during inference, eliminating the need for complex real-time optimization and reducing device complexity while preserving classification accuracy.
2Reliability
If conventional classification systems are used, then comprehensive event detection is achieved, but energy consumption and processing requirements increase
Solution Approach 1:
The patent divides the computational workload into offline training (energy-intensive Hungarian algorithm) and online inference (energy-efficient assignment application). This segmentation enables comprehensive event detection with high accuracy while minimizing energy consumption during actual device operation, making it suitable for low-power embedded systems.
Solution Approach 2:
The patent performs energy-intensive computations preliminarily during the training phase, including optimal class assignment computation via the Hungarian algorithm. This preliminary action transfers the energy burden to an offline setting, allowing the device to perform accurate classification with minimal energy consumption during operational phases.
3Measurement precision
If complex classification algorithms are used, then classification precision is improved, but processing speed decreases
Solution Approach 1:
The patent segments the computational process into an offline training phase where the Hungarian algorithm computes optimal assignments with high precision, and an online inference phase where these pre-computed assignments are rapidly applied. This segmentation achieves both high classification precision through optimal matching and high processing speed during inference by avoiding repeated complex computations.
Solution Approach 2:
The patent performs the computationally intensive optimal class assignment computation as a preliminary action during training. This allows the system to achieve high classification precision through optimal matching while maintaining fast processing speed during actual use, as the inference phase only requires applying pre-computed assignments rather than performing complex optimization.
Data Source
AI summary
A classification training system comprises a neural network configured to perform classification of input data, a training dataset including pre-segmented, labeled training samples, and a classification training module configured to train the neural network using the training dataset. The classification training module includes a forward pass processing module, and a backward pass processing module. The backward pass processing module is configured to determine whether a current frame is in a region of target (ROT), determine ROT information such as beginning and length of the ROT and update weights and biases using a cross-entropy cost function and a tunable many-or-one detection (MOOD) cost function, that comprises a tunable hyperparameter for tuning the classifier for a particular task. The backward pass module further computes a soft target value using ROT information and computes a signal output error using the soft target value and network output value.