Andrew G. Lee

Khondker Fariha Hossain, Sharif Amit Kamran, Joshua Ong et al.

The rapid accessibility of portable and affordable retinal imaging devices has made early differential diagnosis easier. For example, color funduscopy imaging is readily available in remote villages, which can help to identify diseases like age-related macular degeneration (AMD), glaucoma, or pathological myopia (PM). On the other hand, astronauts at the International Space Station utilize this camera for identifying spaceflight-associated neuro-ocular syndrome (SANS). However, due to the unavailability of experts in these locations, the data has to be transferred to an urban healthcare facility (AMD and glaucoma) or a terrestrial station (e.g, SANS) for more precise disease identification. Moreover, due to low bandwidth limits, the imaging data has to be compressed for transfer between these two places. Different super-resolution algorithms have been proposed throughout the years to address this. Furthermore, with the advent of deep learning, the field has advanced so much that x2 and x4 compressed images can be decompressed to their original form without losing spatial information. In this paper, we introduce a novel model called Swin-FSR that utilizes Swin Transformer with spatial and depth-wise attention for fundus image super-resolution. Our architecture achieves Peak signal-to-noise-ratio (PSNR) of 47.89, 49.00 and 45.32 on three public datasets, namely iChallenge-AMD, iChallenge-PM, and G1020. Additionally, we tested the model's effectiveness on a privately held dataset for SANS provided by NASA and achieved comparable results against previous architectures.

Xiaoyan Bai, Itamar Pres, Yuntian Deng et al.

Language models are increasingly capable, yet still fail at a seemingly simple task of multi-digit multiplication. In this work, we study why, by reverse-engineering a model that successfully learns multiplication via \emph{implicit chain-of-thought}, and report three findings: (1) Evidence of long-range structure: Logit attributions and linear probes indicate that the model encodes the necessary long-range dependencies for multi-digit multiplication. (2) Mechanism: the model encodes long-range dependencies using attention to construct a directed acyclic graph to ``cache'' and ``retrieve'' pairwise partial products. (3) Geometry: the model implements partial products in attention heads by forming Minkowski sums between pairs of digits, and digits are represented using a Fourier basis, both of which are intuitive and efficient representations that the standard fine-tuning model lacks. With these insights, we revisit the learning dynamics of standard fine-tuning and find that the model converges to a local optimum that lacks the required long-range dependencies. We further validate this understanding by introducing an auxiliary loss that predicts the ``running sum'' via a linear regression probe, which provides an inductive bias that enables the model to successfully learn multi-digit multiplication. In summary, by reverse-engineering the mechanisms of an implicit chain-of-thought model we uncover a pitfall for learning long-range dependencies in Transformers and provide an example of how the correct inductive bias can address this issue.