ConFeSS: A Framework for Single Source Cross-Domain Few-Shot Learning

Authors: Debasmit Das, Sungrack Yun, Fatih Porikli

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our framework outperforms all meta-learning approaches and produces competitive results against recent cross-domain methods. Additional analyses are also performed to better understand our framework.
Researcher Affiliation Industry Debasmit Das, Sungrack Yun & Fatih Porikli Qualcomm AI Research. {debadas,sungrack,fporikli}@qti.qualcomm.com
Pseudocode Yes Algorithm 1: Con Fe SS framework
Open Source Code No The paper mentions a link to a benchmark repository used for comparison: "The details of the benchmark used for comparison have been introduced in the following repository: https://github.com/IBM/cdfsl-benchmark". However, it does not provide an explicit statement or link for the open-source code of their proposed method (Con Fe SS).
Open Datasets Yes This benchmark uses mini-Image Net (Vinyals et al., 2016), which is a subset of the Image Net (Deng et al., 2009) dataset as the source domain that contains abundantly labeled natural categories. The model learned on the mini-Imagenet dataset is then tested on target datasets containing only a few labeled training data. These target datasets have large domain differences from the source domain, and in order of increasing dissimilarity, they consist of the following: a) Crop Diseases (Mohanty et al., 2016), containing images of different plant disease types, b) Euro SAT (Helber et al., 2019), consisting of different classes of satellite imagery, c) ISIC2018 (Tschandl et al., 2018; Codella et al., 2018), which contains different dermoscopic images of skin lesions, and d) Chest X (Wang et al., 2017), a collection of chest X-Ray images of different lung disease types. Footnotes provide URLs for these datasets.
Dataset Splits Yes For all experiments, the average accuracy over 600 episodes of N-way K-shot setting is reported. Each episode contains randomly sampled K-shot samples per class for adaptation and 15 query samples per class for evaluation, where N is the number of sampled classes.
Hardware Specification Yes All our experiments were conducted in a cluster of nodes where the nodes contain NVIDIA Tesla V100 GPUs with a mix of 16GB and 32GB memory.
Software Dependencies No The paper mentions using "Adam as the optimizer" and "Res Net-10 backbone" but does not specify version numbers for these or other software libraries/frameworks (e.g., Python, PyTorch/TensorFlow versions).
Experiment Setup Yes We use Adam as the optimizer with a learning rate of 0.001. ... we use the pre-training batch size Nb = 50 and the augmentation size Nt = 3... The masking module M( ) consists of a small two-layer feed-forward network with a hidden layer dimension of 256. We set temperature τ = 1. Also, we set λpos = 10 3, λneg = 10 2, λdiv = 10 2, and λreg = 10 2. The numbers of training epochs for Step 1, Step 2, and Step 3 in Algorithm 1 are set as 400, 15, and 50, respectively.