Abstract
This document provides critical corrections to the 2021 paper "Learning to Learn Adaptive Classifier–Predictor for Few-Shot Learning." The primary focus is the replacement of placeholder data in the "Fully-supervised (Upper bound)" benchmarks and the clarification of technical phrasing regarding parameter updates and model comparisons. These corrections ensure the accuracy of the archival record for the few-shot learning community.
Introduction
In the original publication, the authors introduced the Adaptive Classifier–Predictor (MPM) to advance few-shot learning (FSL), where models must recognize new categories from only a handful of examples. However, a data entry error resulted in the "Fully-supervised (Upper bound)" results being left as intermediate placeholder records. This correction addresses that error to provide accurate baseline figures for comparison against the proposed MPM method.
Technical Clarifications
Beyond data rectification, several technical descriptions were refined to improve clarity and precision:
- Terminology: The phrase "update the direction of parameters" has been corrected to "the updating direction of parameters."
- Contextual Comparisons: The authors clarified that specific components of their framework are "similar to" existing methods (such as MAML or work by other researchers) rather than being direct implementations or identical copies. This distinction is crucial for understanding the architectural nuances of gradient-based few-shot learning.
Corrected Experimental Results
The corrections apply to two gold-standard benchmarks: miniImageNet (100 classes, 600 samples each) and tieredImageNet (a larger subset with hierarchical class grouping).
miniImageNet Performance
The corrected "Fully-supervised" results provide a more accurate ceiling for model performance. On miniImageNet, the true upper bound is 89.51%. Despite the correction of these baselines, the relative performance of the proposed MPM method remains competitive across various backbones.
As shown in the table above, MPM with a ResNet12 backbone achieved 62.19% in 1-shot and 78.51% in 5-shot scenarios, outperforming contemporary baselines like TADAM and MetaTransfer.
tieredImageNet Performance
For the more challenging tieredImageNet dataset, the corrected upper bound is 91.80%. The MPM method continues to show strong results, particularly in 5-shot scenarios.
On tieredImageNet, MPM using the WRN-28-10 backbone reached 67.58% (1-shot) and a high of 83.93% (5-shot). These results demonstrate that while the gap between few-shot learning and full supervision is closing, a performance gap of approximately 10-15% remains on complex hierarchical datasets.
Conclusion
While the placeholder numbers in the original publication were incorrect, the authors emphasize that the relative performance and the original paper's conclusions remain unchanged. The MPM method remains a robust and competitive approach compared to state-of-the-art models like Meta-LSTM, RelationNet, and SNAIL. These corrections ensure that future researchers have an accurate benchmark for evaluating progress in the field.
Summary updated