Classification Threshold Inversion for Low-Compute AI Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Softmax function computations strain computational resources in embedded devices with limited resources, necessitating a reduction in classification model computations.
Innovation Solution
Implement an inverse transform to derive a non-normalized classification threshold from classification probability values, reducing the need for exponential computations in Softmax functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Softmax function is used in output layer of DNN for classification, then classification accuracy is improved, but computational resource consumption increases
Solution Approach 1:
The patent changes the parameter space by introducing transformed classification thresholds that operate in the pre-Softmax domain. Instead of working with probability values after Softmax transformation, the system transforms thresholds to match the output distribution of the classification model, avoiding the need for expensive exponential computations while maintaining accurate classification decisions
Solution Approach 2:
The patent performs preliminary transformation of classification thresholds before the actual classification process. By pre-computing transformed thresholds that account for the classification model's output characteristics, the system eliminates the need for runtime Softmax computations, reducing computational load during inference while preserving classification accuracy
2Measurement precision
If Softmax function with exponential computations is implemented, then classification probability values are obtained, but device complexity increases
Solution Approach 1:
The patent inverts the traditional approach by not computing probabilities through Softmax and then comparing against thresholds. Instead, it transforms thresholds to match the classification model's native output space, effectively working backwards from the decision boundary to the model output, thereby avoiding exponential computations entirely
3Measurement precision
If classification thresholds are set based on classification probability values, then accurate classification is achieved, but additional computational steps are required
Solution Approach 1:
The patent performs the threshold transformation in advance during model deployment or training, storing the pre-computed transformed thresholds. During actual classification operations, the system directly compares model outputs against these pre-transformed thresholds without requiring runtime probability computations, significantly improving classification speed while maintaining accuracy
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing device (100a) includes a processor. The processor obtains a first classification threshold for classifying data into at least one of a plurality of classes, and outputs a classification result of classifying the data into at least one of the plurality of classes based on an output of a trained classification model and the first classification threshold. The first classification threshold is obtained by a second transform performed on a second classification threshold, the second transform being an inverse transform of a first transform. The first transform corresponds to transforming the output of the trained classification model into a classification probability value of each of a plurality of unit classes constituting the plurality of classes. The second classification threshold is set based on the classification probability values of the plurality of unit classes.