Exploring connections of spectral analysis and transfer learning in medical imaging

In this paper, we use spectral analysis to investigate transfer learning and study model sensitivity to frequency shortcuts in medical imaging. By analyzing the power spectrum density of both pre-trained and fine-tuned model gradients, as well as artificially generated frequency shortcuts, we observe notable differences in learning priorities between models pre-trained on natural vs medical images, which generally persist during fine-tuning. We find that when a model's learning priority aligns with the power spectrum density of an artifact, it results in overfitting to that artifact. Based on these observations, we show that source data editing can alter the model's resistance to shortcut learning.

Source Matters: Source Dataset Impact on Model Robustness in Medical Imaging

Transfer learning has become an essential part of medical imaging classification algorithms, often leveraging ImageNet weights. However, the domain shift from natural to medical images has prompted alternatives such as RadImageNet, often demonstrating comparable classification performance. However, it remains unclear whether the performance gains from transfer learning stem from improved generalization or shortcut learning. To address this, we investigate potential confounders -- whether synthetic or sampled from the data -- across two publicly available chest X-ray and CT datasets. We show that ImageNet and RadImageNet achieve comparable classification performance, yet ImageNet is much more prone to overfitting to confounders. We recommend that researchers using ImageNet-pretrained models reexamine their model robustness by conducting similar experiments. Our code and experiments are available at https://github.com/DovileDo/source-matters.

Multi-frame-based Cross-domain Image Denoising for Low-dose Computed Tomography

Computed tomography (CT) has been used worldwide for decades as one of the most important non-invasive tests in assisting diagnosis. However, the ionizing nature of X-ray exposure raises concerns about potential health risks such as cancer. The desire for lower radiation dose has driven researchers to improve the reconstruction quality, especially by removing noise and artifacts. Although previous studies on low-dose computed tomography (LDCT) denoising have demonstrated the potential of learning-based methods, most of them were developed on the simulated data collected using Radon transform. However, the real-world scenario significantly differs from the simulation domain, and the joint optimization of denoising with the modern CT image reconstruction pipeline is still missing. In this paper, for the commercially available third-generation multi-slice spiral CT scanners, we propose a two-stage method that better exploits the complete reconstruction pipeline for LDCT denoising across different domains. Our method makes good use of the high redundancy of both the multi-slice projections and the volumetric reconstructions while avoiding the collapse of information in conventional cascaded frameworks. The dedicated design also provides a clearer interpretation of the workflow. Through extensive evaluations, we demonstrate its superior performance against state-of-the-art methods.

STEP: Learning N:M Structured Sparsity Masks from Scratch with Precondition

Recent innovations on hardware (e.g. Nvidia A100) have motivated learning N:M structured sparsity masks from scratch for fast model inference. However, state-of-the-art learning recipes in this regime (e.g. SR-STE) are proposed for non-adaptive optimizers like momentum SGD, while incurring non-trivial accuracy drop for Adam-trained models like attention-based LLMs. In this paper, we first demonstrate such gap origins from poorly estimated second moment (i.e. variance) in Adam states given by the masked weights. We conjecture that learning N:M masks with Adam should take the critical regime of variance estimation into account. In light of this, we propose STEP, an Adam-aware recipe that learns N:M masks with two phases: first, STEP calculates a reliable variance estimate (precondition phase) and subsequently, the variance remains fixed and is used as a precondition to learn N:M masks (mask-learning phase). STEP automatically identifies the switching point of two phases by dynamically sampling variance changes over the training trajectory and testing the sample concentration. Empirically, we evaluate STEP and other baselines such as ASP and SR-STE on multiple tasks including CIFAR classification, machine translation and LLM fine-tuning (BERT-Base, GPT-2). We show STEP mitigates the accuracy drop of baseline recipes and is robust to aggressive structured sparsity ratios.

Scale up with Order: Finding Good Data Permutations for Distributed Training

Gradient Balancing (GraB) is a recently proposed technique that finds provably better data permutations when training models with multiple epochs over a finite dataset. It converges at a faster rate than the widely adopted Random Reshuffling, by minimizing the discrepancy of the gradients on adjacently selected examples. However, GraB only operates under critical assumptions such as small batch sizes and centralized data, leaving open the question of how to order examples at large scale -- i.e. distributed learning with decentralized data. To alleviate the limitation, in this paper we propose D-GraB that involves two novel designs: (1) $\textsf{PairBalance}$ that eliminates the requirement to use stale gradient mean in GraB which critically relies on small learning rates; (2) an ordering protocol that runs $\textsf{PairBalance}$ in a distributed environment with negligible overhead, which benefits from both data ordering and parallelism. We prove D-GraB enjoys linear speed up at rate $\tilde{O}((mnT)^{-2/3})$ on smooth non-convex objectives and $\tilde{O}((mnT)^{-2})$ under PL condition, where $n$ denotes the number of parallel workers, $m$ denotes the number of examples per worker and $T$ denotes the number of epochs. Empirically, we show on various applications including GLUE, CIFAR10 and WikiText-2 that D-GraB outperforms naive parallel GraB and Distributed Random Reshuffling in terms of both training and validation performance.

RZSR: Reference-based Zero-Shot Super-Resolution with Depth Guided Self-Exemplars

Recent methods for single image super-resolution (SISR) have demonstrated outstanding performance in generating high-resolution (HR) images from low-resolution (LR) images. However, most of these methods show their superiority using synthetically generated LR images, and their generalizability to real-world images is often not satisfactory. In this paper, we pay attention to two well-known strategies developed for robust super-resolution (SR), i.e., reference-based SR (RefSR) and zero-shot SR (ZSSR), and propose an integrated solution, called reference-based zero-shot SR (RZSR). Following the principle of ZSSR, we train an image-specific SR network at test time using training samples extracted only from the input image itself. To advance ZSSR, we obtain reference image patches with rich textures and high-frequency details which are also extracted only from the input image using cross-scale matching. To this end, we construct an internal reference dataset and retrieve reference image patches from the dataset using depth information. Using LR patches and their corresponding HR reference patches, we train a RefSR network that is embodied with a non-local attention module. Experimental results demonstrate the superiority of the proposed RZSR compared to the previous ZSSR methods and robustness to unseen images compared to other fully supervised SISR methods.

GraB: Finding Provably Better Data Permutations than Random Reshuffling

Random reshuffling, which randomly permutes the dataset each epoch, is widely adopted in model training because it yields faster convergence than with-replacement sampling. Recent studies indicate greedily chosen data orderings can further speed up convergence empirically, at the cost of using more computation and memory. However, greedy ordering lacks theoretical justification and has limited utility due to its non-trivial memory and computation overhead. In this paper, we first formulate an example-ordering framework named herding and answer affirmatively that SGD with herding converges at the rate $O(T^{-2/3})$ on smooth, non-convex objectives, faster than the $O(n^{1/3}T^{-2/3})$ obtained by random reshuffling, where $n$ denotes the number of data points and $T$ denotes the total number of iterations. To reduce the memory overhead, we leverage discrepancy minimization theory to propose an online Gradient Balancing algorithm (GraB) that enjoys the same rate as herding, while reducing the memory usage from $O(nd)$ to just $O(d)$ and computation from $O(n^2)$ to $O(n)$, where $d$ denotes the model dimension. We show empirically on applications including MNIST, CIFAR10, WikiText and GLUE that GraB can outperform random reshuffling in terms of both training and validation performance, and even outperform state-of-the-art greedy ordering while reducing memory usage over $100\times$.

Maximizing Communication Efficiency for Large-scale Training via 0/1 Adam

1-bit communication is an effective method to scale up model training, and has been studied extensively on SGD. Its benefits, however, remain an open question on Adam-based model training (e.g. BERT and GPT). In this paper, we propose 0/1 Adam, which improves upon the state-of-the-art 1-bit Adam via two novel designs: (1) adaptive variance state freezing, which eliminates the requirement of running expensive full-precision communication at early stage of training; (2) 1-bit sync, which allows skipping communication rounds with bit-free synchronization over Adam's optimizer states, momentum and variance. In theory, we provide convergence analysis for 0/1 Adam on smooth non-convex objectives, and show the complexity bound is better than original Adam under certain conditions. On various benchmarks such as BERT-Base/Large pretraining and ImageNet, we demonstrate on up to 128 GPUs that 0/1 Adam is able to reduce up to 90% of data volume, 54% of communication rounds, and achieve up to 2X higher throughput compared to the state-of-the-art 1-bit Adam while enjoying the same statistical convergence speed and end-to-end model accuracy on GLUE dataset and ImageNet validation set.

Progressive Joint Low-light Enhancement and Noise Removal for Raw Images

Low-light imaging on mobile devices is typically challenging due to insufficient incident light coming through the relatively small aperture, resulting in a low signal-to-noise ratio. Most of the previous works on low-light image processing focus either only on a single task such as illumination adjustment, color enhancement, or noise removal; or on a joint illumination adjustment and denoising task that heavily relies on short-long exposure image pairs collected from specific camera models, and thus these approaches are less practical and generalizable in real-world settings where camera-specific joint enhancement and restoration is required. To tackle this problem, in this paper, we propose a low-light image processing framework that performs joint illumination adjustment, color enhancement, and denoising. Considering the difficulty in model-specific data collection and the ultra-high definition of the captured images, we design two branches: a coefficient estimation branch as well as a joint enhancement and denoising branch. The coefficient estimation branch works in a low-resolution space and predicts the coefficients for enhancement via bilateral learning, whereas the joint enhancement and denoising branch works in a full-resolution space and progressively performs joint enhancement and denoising. In contrast to existing methods, our framework does not need to recollect massive data when being adapted to another camera model, which significantly reduces the efforts required to fine-tune our approach for practical usage. Through extensive experiments, we demonstrate its great potential in real-world low-light imaging applications when compared with current state-of-the-art methods.

Variance Reduction in Training Forecasting Models with Subgroup Sampling

In real-world applications of large-scale time series, one often encounters the situation where the temporal patterns of time series, while drifting over time, differ from one another in the same dataset. In this paper, we provably show under such heterogeneity, training a forecasting model with commonly used stochastic optimizers (e.g. SGD) potentially suffers large gradient variance, and thus requires long time training. To alleviate this issue, we propose a sampling strategy named Subgroup Sampling, which mitigates the large variance via sampling over pre-grouped time series. We further introduce SCott, a variance reduced SGD-style optimizer that co-designs subgroup sampling with the control variate method. In theory, we provide the convergence guarantee of SCott on smooth non-convex objectives. Empirically, we evaluate SCott and other baseline optimizers on both synthetic and real-world time series forecasting problems, and show SCott converges faster with respect to both iterations and wall clock time. Additionally, we show two SCott variants that can speed up Adam and Adagrad without compromising generalization of forecasting models.