Although deep learning models have had great success in natural languageprocessing and computer vision, we do not observe comparable improvements inthe case of tabular data, which is still the most common data type used inbiological, industrial and financial applications. In particular, it ischallenging to transfer large-scale pre-trained models to downstream tasksdefined on small tabular datasets. To address this, we propose VisTabNet -- across-modal transfer learning method, which allows for adapting VisionTransformer (ViT) with pre-trained weights to process tabular data. Byprojecting tabular inputs to patch embeddings acceptable by ViT, we candirectly apply a pre-trained Transformer Encoder to tabular inputs. Thisapproach eliminates the conceptual cost of designing a suitable architecturefor processing tabular data, while reducing the computational cost of trainingthe model from scratch. Experimental results on multiple small tabular datasets(less than 1k samples) demonstrate VisTabNet's superiority, outperforming bothtraditional ensemble methods and recent deep learning models. The proposedmethod goes beyond conventional transfer learning practice and shows thatpre-trained image models can be transferred to solve tabular problems,extending the boundaries of transfer learning. We share our exampleimplementation as a GitHub repository available athttps://github.com/wwydmanski/VisTabNet.
We present a detailed analysis of the high-mass binary system V1216 Sco, aneclipsing Algol-type binary hosting a $\beta$ Cephei pulsator, with an orbitalperiod of 3.92 days. This system was analyzed using TESS photometry andhigh-resolution spectroscopy from SALT HRS to investigate its orbitalparameters, stellar properties, and evolutionary history. The TESS light curve,comprising over 12000 data points, revealed five independent pulsationfrequencies within the beta Cephei range of 5 to 7 d$^{-1}$. Spectroscopic analysis provided radial velocities and disentangledatmospheric parameters, enabling precise orbital and evolutionary modeling. Thesystem features a primary star of 11.72 M$_{\odot}$ and a secondary of 4.34M$_{\odot}$. The secondary star has a radius near its Roche lobe, indicatingrecent or ongoing mass transfer. Evolutionary modeling with MESA-binarysuggests that V1216 Sco underwent a case A mass transfer scenario, where masstransfer began while the donor was still on the main sequence. The system'sevolutionary models indicate an age of 15 to 30 million years, highlighting thesignificant impact of binary interactions on stellar evolution. This study underscores the value of combining observational and theoreticalapproaches to understanding complex systems like V1216 Sco, emphasizing therole of mass transfer in shaping binary star evolution.
Superradiance, the phenomenon enabling energy extraction through radiationamplification, is not universal to all black holes. We show that semi-classicalbackreaction can induce superradiance, even when absent at the classical level.Specifically, we compute the quasinormal modes of a massless scalar fieldprobing a family of rotating `quantum' black holes in three-dimensional anti-deSitter space, accounting for all orders of backreaction due to quantumconformal matter. A subset of these modes is identified as superradiant,leading to the formation of quantum black hole `bombs'. All such quantum blackholes have curvature singularities shrouded by horizons. Thus, whilebackreaction enforces cosmic censorship, it also renders the black holesdynamically unstable. Further, we find all thermally unstable black holes aredynamically unstable, though the converse does not hold generally. Our findingsthus suggest a semiclassical version of the Gubser-Mitra conjecture on blackhole stability. This motivates us to propose a stability criterion for quantumblack holes.
Despite the growing interest in human-AI decision making, experimentalstudies with domain experts remain rare, largely due to the complexity ofworking with domain experts and the challenges in setting up realisticexperiments. In this work, we conduct an in-depth collaboration withradiologists in prostate cancer diagnosis based on MRI images. Building onexisting tools for teaching prostate cancer diagnosis, we develop an interfaceand conduct two experiments to study how AI assistance and performance feedbackshape the decision making of domain experts. In Study 1, clinicians were askedto provide an initial diagnosis (human), then view the AI's prediction, andsubsequently finalize their decision (human-AI team). In Study 2 (after amemory wash-out period), the same participants first received aggregatedperformance statistics from Study 1, specifically their own performance, theAI's performance, and their human-AI team performance, and then directly viewedthe AI's prediction before making their diagnosis (i.e., no independent initialdiagnosis). These two workflows represent realistic ways that clinical AI toolsmight be used in practice, where the second study simulates a scenario wheredoctors can adjust their reliance and trust on AI based on prior performancefeedback. Our findings show that, while human-AI teams consistently outperformhumans alone, they still underperform the AI due to under-reliance, similar toprior studies with crowdworkers. Providing clinicians with performance feedbackdid not significantly improve the performance of human-AI teams, althoughshowing AI decisions in advance nudges people to follow AI more. Meanwhile, weobserve that the ensemble of human-AI teams can outperform AI alone, suggestingpromising directions for human-AI collaboration.
Non-identically distributed data is a major challenge in Federated Learning(FL). Personalized FL tackles this by balancing local model adaptation withglobal model consistency. One variant, partial FL, leverages the observationthat early layers learn more transferable features by federating only earlylayers. However, current partial FL approaches use predetermined,architecture-specific rules to select layers, limiting their applicability. Weintroduce Principled Layer-wise-FL (PLayer-FL), which uses a novel federationsensitivity metric to identify layers that benefit from federation. Thismetric, inspired by model pruning, quantifies each layer's contribution tocross-client generalization after the first training epoch, identifying atransition point in the network where the benefits of federation diminish. Wefirst demonstrate that our federation sensitivity metric shows strongcorrelation with established generalization measures across diversearchitectures. Next, we show that PLayer-FL outperforms existing FL algorithmson a range of tasks, also achieving more uniform performance improvementsacross clients.
Recent Bayesian analyses of heavy ion collision data have established anon-trivial temperature dependence of the shear and bulk viscosity per entropy.Motivated by this, we consider higher derivative corrections to realistic,bottom-up holographic models of quark-gluon plasma based on five-dimensionalEinstein-dilaton theories and determine the dilaton potentials in the higherderivative terms by matching the Bayesian analyses. A byproduct of our analysisis the bulk viscosity that follows from the holographic V-QCD theory. Higherderivative corrections when treated perturbatively lead to tension withexisting data. We investigate possible resolutions.