Firth Bias Reduction in Few-shot Classification

This repository contains all the experiments conducted in the On the Importance of Firth Bias Reduction in Few-Shot Classification paper. For a concise and informal description of our work, check out our paper's website: https://ehsansaleh.github.io/firthfsl

To clone this repo with all three of its sub-modules, run:

git clone --recursive https://github.com/ehsansaleh/firth_bias_reduction.git

The Paper in Pictures

• The MLE Bias in Few-shot Classification

  Here is a visualization to help you get the overall context of typical loss minimization (MLE) bias with only a few samples.

• Firth Bias Reduction in Few Words

  • For 1-Layer Logistic and Cosine Classifiers with the Cross-Entropy Loss

    All you need to do, is replace

    

    with

    

    where U is the uniform distribution over the classes, and lambda is a positive constant. The CE-term with the uniform distribution is basically the (negative) sum of the prediction log-probability values over all data points and classes.

  • General Firth Bias Reduction Form

    Add a log-det of FIM term to your loss minimization problem. That is, replace

    

    with

    

    ,

    This was proven to reduce the bias of your estimated parameters.

• Firth Bias Reduction in a Geometric Experiment

  Here is a simple example show-casing average the MLE's bias from the true parameters in a geometric experiment with a fair coin, and the slow rate at which this bias disappears.

  

• Firth Bias Reduction Improvements in Few-shot Classification Tasks

  Here is the effect of Firth bias reduction campared to typical L2 regularization in 16-way few-shot classification tasks using basic feature backbones and 3-layer logistic classifiers.

  

  Below is the effect of Firth bias reduction on cosine classifiers and S2M2R features.

  

  

The Repository Structure

• code_firth directory contains the Firth regularization code used for the standard ResNet architecture tested on the mini-Imagenet data set.

• code_s2m2rf directory contains the experimnets on mini-Imagenet, CIFAR-FS, and tiered-Imagenet data sets using the WideResNet28 feature stack trained by the S2M2R method.

• code_dcf directory contains our GPU implementation of Distribution Calibration (DC) method and all the experiments performed on it. If you use our GPU implementation of the DC method, please cite our paper "On the Importance of Firth Bias Reduction in Few-Shot Classification, ICLR 2022".

References

• Here is the arxiv link to our paper:
  • The arxiv PDF link: https://arxiv.org/pdf/2110.02529.pdf
  • The arxiv web-page link: https://arxiv.org/abs/2110.02529
• Here is the open-review link to our paper:
  • The open-review PDF link: https://openreview.net/pdf?id=DNRADop4ksB
  • The open-review forum link: https://openreview.net/forum?id=DNRADop4ksB
• Our paper got a spot-light presentation at ICLR 2022.
  • We will update here with links to the presentation video and the web-page on iclr.cc.
• • Here is a web-page for our paper: https://ehsansaleh.github.io/firthfsl
• Here is the bibtex citation entry for our work:

@inproceedings{ghaffari2022fslfirth,
    title={On the Importance of Firth Bias Reduction in Few-Shot Classification},
    author={Saba Ghaffari and Ehsan Saleh and David Forsyth and Yu-Xiong Wang},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=DNRADop4ksB}
}