Classification Threshold Transform for Low-Compute Model Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The computational strain caused by exponential computations in Softmax functions when implementing classification models on embedded devices with limited resources, leading to increased computational load and reduced recognition accuracy.
Innovation Solution
Implementing an inverse-transformable function to convert classification model outputs into non-normalized classification thresholds, reducing the need for exponential computations by using a second transform to derive a first classification threshold for classifying data into classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Softmax function is used for classification, then classification accuracy is improved, but computational load increases
Solution Approach 1:
The patent changes the parameter space by transforming classification thresholds through inverse Softmax function. Instead of computing Softmax probabilities and comparing against 0.5 threshold, the patent computes transformed thresholds using inverse Softmax on logit values, maintaining classification accuracy while reducing computational complexity by avoiding exponential operations.
Solution Approach 2:
The patent applies the inverse of the Softmax function (logistic function) to transform the classification thresholds. By inverting the transformation direction, the patent works backwards from probability space to logit space, allowing thresholds to be computed without exponential operations while preserving the decision boundary.
2Loss of information
If Softmax function is used for classification, then classification probability values are obtained, but resource consumption increases
Solution Approach 1:
The patent performs preliminary transformation of classification thresholds using inverse Softmax function during the model training or deployment phase. By pre-computing and storing the transformed thresholds in logit space, the patent eliminates the need for real-time exponential computations during inference, reducing resource consumption while preserving classification probability information.
3Measurement precision
If exponential computations are performed, then classification probability values are accurate, but processing time increases
Solution Approach 1:
The patent changes the computational parameter from exponential domain to linear domain by applying inverse Softmax transformation. This parameter change allows thresholds to be computed using simple linear operations instead of expensive exponential computations, significantly reducing processing time while maintaining the accuracy of classification decisions.
Data Source
AI summary
An information processing device 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.