SAFLEX: Self-Adaptive Augmentation via Feature Label Extrapolation
Authors: Mucong Ding, Bang An, Yuancheng Xu, Anirudh Satheesh, Furong Huang
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted extensive empirical evaluations to highlight SAFLEX s superior performance and adaptability. On eight medical images (Yang et al., 2023), SAFLEX elevates popular augmentation techniques like Rand Augment (Cubuk et al., 2020) and Mixup (Zhang et al., 2018), boosting performance by up to 3.6%. On seven tabular datasets, SAFLEX shows compatibility with categorical data and effectively enhances Cut Mix (Yun et al., 2019). Furthermore, SAFLEX improves image augmentations from diffusion models, yielding an average improvement of 1.9% in fine-grained classification and out-of-distribution generalization against three diffusion-augmentation methods, harnessing on their pre-trained expertise. We also validate SAFLEX s integration with contrastive learning through a CLIP fine-tuning experiment. |
| Researcher Affiliation | Academia | Mucong Ding, Bang An, Yuancheng Xu, Anirudh Satheesh, Furong Huang Department of Computer Science University of Maryland mcding@cs.umd.edu |
| Pseudocode | Yes | The pseudo-code of SAFLEX for cross-entropy loss is shown as Algorithm 1. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing the source code for the SAFLEX methodology or provide a link to a code repository. |
| Open Datasets | Yes | We assess multi-class classification across eight medical image datasets from Med MNIST (Yang et al., 2023)... Our experiments span seven tabular datasets... Our evaluation encompasses two tasks: (1) Fine-grained classification on a CUB dataset subset (Wah et al., 2011)... (2) OOD generalization on an i Wild Cam subset from the Wilds dataset (Koh et al., 2021)... We use a Res Net-50 model (He et al., 2016) pretrained on Image Net (Deng et al., 2009). |
| Dataset Splits | No | The paper mentions using a "validation set" (Dval) and that fine-tuning performance is done using it, but it does not specify the exact percentages or counts for the train/validation/test splits within the main text required for reproduction. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions models like "Res Net-18", "Adam optimizer", "Stable Diffusion v1.5", and "CLIP", but it does not specify versions of ancillary software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | In line with (Yang et al., 2021), we train a Res Net-18 model (He et al., 2016) using the Adam optimizer (Kingma & Ba, 2014) for 100 epochs. ... We consider backbone models such as the sample Multilayer Perceptron (MLP) with two hidden layers and 256 neurons each and tranformer-based models like SAINT (Somepalli et al., 2022) (without contrastive pretraining). These models undergo training with dropout (Srivastava et al., 2014) and, in certain cases, batch normalization, for 200 epochs. ... We use a Res Net-50 model (He et al., 2016) pretrained on Image Net (Deng et al., 2009). ... we employ Rand Augment, following hyperparameter setups as described in (Koh et al., 2021). |