Rollback Ensemble With Multiple Local Minima in Fine-Tuning Deep Learning Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Image retrieval is challenging, requiring generalized features for untrained classes with limited data.
• Pretrained deep networks often face stagnated loss reduction during fine-tuning, hindering generalization.

Purpose of the Study:

• To propose a novel fine-tuning method for pretrained deep networks to enhance generalized feature learning in image retrieval.
• To address the stagnation phenomenon observed during deep network fine-tuning for retrieval tasks.

Main Methods:

• A new fine-tuning strategy involving rolling back some network weights to their pretrained values.
• Implementation of a multihead ensemble structure to leverage multiple local minima.

Main Results:

• The proposed rollback scheme guides the learning path towards gentle basins, yielding more generalized features.
• The multihead ensemble creates synergy, further improving generalization performance.
• Achieved state-of-the-art results on the Inshop and SOP datasets.

Conclusions:

• The novel fine-tuning method significantly enhances generalization capabilities in image retrieval.
• The rollback scheme and multihead ensemble are effective strategies for overcoming fine-tuning limitations and improving performance.