A Probabilistic Re-Intepretation of Confidence Scores in Multi-Exit Models.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) often require significant computational resources for inference.
• Existing multi-exit models assume fixed confidence thresholds for early prediction.
• Training multiple branches of a model can be complex and require extensive hyperparameter tuning.

Purpose of the Study:

• To propose a novel, hyperparameter-free method for jointly training multi-exit deep neural networks.
• To enable efficient inference by dynamically determining the optimal exit point for each input sample.
• To improve the computational efficiency of deep learning models without sacrificing accuracy.

Main Methods:

• Developed a training approach for multi-exit DNNs using trained confidence scores for each branch.
• Implemented a joint training strategy that regularizes confidence score calculation during model training.
• Evaluated the proposed method on image classification benchmarks using various neural architectures and early-exit criteria.

Main Results:

• Demonstrated a method to train multi-exit models effectively without manual hyperparameter selection.
• Showcased the potential for reduced inference time by leveraging early exits based on learned confidence.
• Achieved competitive performance on image classification tasks, validating the approach across different models.

Conclusions:

• The proposed training strategy offers an efficient and effective way to develop multi-exit deep neural networks.
• This method facilitates adaptive inference, reducing computational cost for samples that do not require full network processing.
• The learned confidence scores provide a robust mechanism for early stopping, enhancing model efficiency.