In this work, we propose a novel framework to enhance the efficiency and accuracy of neural operators through self-composition, offering both theoretical guarantees and practical benefits. Inspired by iterative methods in solving numerical partial differential equations (PDEs), we design a specific neural operator by repeatedly applying a single neural operator block, we progressively deepen the model without explicitly adding new blocks, improving the model's capacity. To train these models efficiently, we introduce an adaptive train-and-unroll approach, where the depth of the neural operator is gradually increased during training. This approach reveals an accuracy scaling law with model depth and offers significant computational savings through our adaptive training strategy. Our architecture achieves state-of-the-art (SOTA) performance on standard benchmarks. We further demonstrate its efficacy on a challenging high-frequency ultrasound computed tomography (USCT) problem, where a multigrid-inspired backbone enables superior performance in resolving complex wave phenomena. The proposed framework provides a computationally tractable, accurate, and scalable solution for large-scale data-driven scientific machine learning applications.

Identifying sitters in historical paintings is a key task for art historians, offering insight into their lives and how they chose to be seen. However, the process is often subjective and limited by the lack of data and stylistic variations. Automated facial recognition is capable of handling challenging conditions and can assist, but while traditional facial recognition models perform well on photographs, they struggle with paintings due to domain shift and high intra-class variation. Artistic factors such as style, skill, intent, and influence from other works further complicate recognition. In this work, we investigate the potential of foundation models to improve facial recognition in artworks. By fine-tuning foundation models and integrating their embeddings with those from conventional facial recognition networks, we demonstrate notable improvements over current state-of-the-art methods. Our results show that foundation models can bridge the gap where traditional methods are ineffective. Paper page at https://www.idiap.ch/paper/artface/

We investigated the adaptation and performance of Masked Autoencoders (MAEs) with Vision Transformer (ViT) architectures for self-supervised representation learning on one-dimensional (1D) ultrasound signals. Although MAEs have demonstrated significant success in computer vision and other domains, their use for 1D signal analysis, especially for raw ultrasound data, remains largely unexplored. Ultrasound signals are vital in industrial applications such as non-destructive testing (NDT) and structural health monitoring (SHM), where labeled data are often scarce and signal processing is highly task-specific. We propose an approach that leverages MAE to pre-train on unlabeled synthetic ultrasound signals, enabling the model to learn robust representations that enhance performance in downstream tasks, such as time-of-flight (ToF) classification. This study systematically investigated the impact of model size, patch size, and masking ratio on pre-training efficiency and downstream accuracy. Our results show that pre-trained models significantly outperform models trained from scratch and strong convolutional neural network (CNN) baselines optimized for the downstream task. Additionally, pre-training on synthetic data demonstrates superior transferability to real-world measured signals compared with training solely on limited real datasets. This study underscores the potential of MAEs for advancing ultrasound signal analysis through scalable, self-supervised learning.

Effective String Theory (EST) offers a robust non-perturbative framework for describing confinement in Yang-Mills theory by treating the confining flux tube between a static quark-antiquark pair as a thin, vibrating string. While EST calculations are typically carried out using zeta-function regularization, certain problems-such as determining the flux tube width-are too complex to solve analytically. However, recent studies have demonstrated that EST can be explored numerically by employing deep learning techniques based on generative algorithms. In this work, we provide a brief introduction to EST and this novel numerical approach. Finally, we present results for the width of the Nambu-Gotö EST.

Session-based recommendation (SR) models aim to recommend items to anonymous users based on their behavior during the current session. While various SR models in the literature utilize item sequences to predict the next item, they often fail to leverage semantic information from item titles or descriptions impeding session intent identification and interpretability. Recent research has explored Large Language Models (LLMs) as promising approaches to enhance session-based recommendations, with both prompt-based and fine-tuning based methods being widely investigated. However, prompt-based methods struggle to identify optimal prompts that elicit correct reasoning and lack task-specific feedback at test time, resulting in sub-optimal recommendations. Fine-tuning methods incorporate domain-specific knowledge but incur significant computational costs for implementation and maintenance. In this paper, we present multiple approaches to utilize LLMs for session-based recommendation: (i) in-context LLMs as recommendation agents, (ii) LLM-generated representations for semantic initialization of deep learning SR models, and (iii) integration of LLMs with data-driven SR models. Through comprehensive experiments on two real-world publicly available datasets, we demonstrate that LLM-based methods excel at coarse-level retrieval (high recall values), while traditional data-driven techniques perform well at fine-grained ranking (high Mean Reciprocal Rank values). Furthermore, the integration of LLMs with data-driven SR models significantly out performs both standalone LLM approaches and data-driven deep learning models, as well as baseline SR models, in terms of both Recall and MRR metrics.

Large-batch training has become a cornerstone in accelerating the training of deep neural networks, yet it poses challenges in optimization and generalization. Existing optimizers like AdamW present performance degradation during language models' large-batch training, due to the information bottleneck in attention layers caused by the sharp increase of max attention logit. While the LAMB optimizer partially addresses this issue, some attention layers still face this issue. The reason is that $l_2$-norm-based trust ratios in LAMB are less effective in directly influencing the max value of query/key weights. Furthermore, the weight-wise trust ratio in LAMB is error-prone as it overlooks relationships of weight values within rows or columns. Building on these observations, we propose a novel optimizer, MERIT, which leverages the max-norm to calculate the trust ratio to constrain the max attention logit more effectively. Moreover, we further construct element-wise trust ratios to provide more robust update scaling by focusing on local weight structures. Extensive experiments of large-batch training across various sizes of GPT-2 models demonstrate the superior performance of MERIT. Notably, during the training of GPT-2 Medium, MERIT enables a 6k batch size without any performance degradation compared to the standard batch size (480) with 48B training tokens. This work highlights the importance of considering the max attention logit and finer-granularity trust ratio in large-batch training. It successfully improves the training stability and paves the way for larger batch usage, enabling faster development and iteration of large language models. Code is available at https://github.com/NUS-HPC-AI-Lab/MERIT.

The Adam optimizer remains one of the most widely used optimizers in deep learning, and effectively tuning its hyperparameters is key to optimizing performance. However, tuning can be tedious and costly. Freeze-thaw Bayesian Optimization (BO) is a recent promising approach for low-budget hyperparameter tuning, but is limited by generic surrogates without prior knowledge of how hyperparameters affect learning. We propose Adam-PFN, a new surrogate model for Freeze-thaw BO of Adam's hyperparameters, pre-trained on learning curves from TaskSet, together with a new learning curve augmentation method, CDF-augment, which artificially increases the number of available training examples. Our approach improves both learning curve extrapolation and accelerates hyperparameter optimization on TaskSet evaluation tasks, with strong performance on out-of-distribution (OOD) tasks.

Commonsense visual-question answering often hinges on knowledge that is missing from the image or the question. Small vision-language models (sVLMs) such as ViLT, VisualBERT and FLAVA therefore lag behind their larger generative counterparts. To study the effect of careful commonsense knowledge integration on sVLMs, we present an end-to-end framework (NLKI) that (i) retrieves natural language facts, (ii) prompts an LLM to craft natural language explanations, and (iii) feeds both signals to sVLMs respectively across two commonsense VQA datasets (CRIC, AOKVQA) and a visual-entailment dataset (e-SNLI-VE). Facts retrieved using a fine-tuned ColBERTv2 and an object information-enriched prompt yield explanations that largely cut down hallucinations, while lifting the end-to-end answer accuracy by up to 7% (across 3 datasets), making FLAVA and other models in NLKI match or exceed medium-sized VLMs such as Qwen-2 VL-2B and SmolVLM-2.5B. As these benchmarks contain 10-25% label noise, additional finetuning using noise-robust losses (such as symmetric cross entropy and generalised cross entropy) adds another 2.5% in CRIC, and 5.5% in AOKVQA. Our findings expose when LLM-based commonsense knowledge beats retrieval from commonsense knowledge bases, how noise-aware training stabilises small models in the context of external knowledge augmentation, and why parameter-efficient commonsense reasoning is now within reach for 250M models.

A new generation of AI models generates step-by-step reasoning text before producing an answer. This text appears to offer a human-readable window into their computation process, and is increasingly relied upon for transparency and interpretability. However, it is unclear whether human understanding of this text matches the model's actual computational process. In this paper, we investigate a necessary condition for correspondence: the ability of humans to identify which steps in a reasoning text causally influence later steps. We evaluated humans on this ability by composing questions based on counterfactual measurements and found a significant discrepancy: participant accuracy was only 29%, barely above chance (25%), and remained low (42%) even when evaluating the majority vote on questions with high agreement. Our results reveal a fundamental gap between how humans interpret reasoning texts and how models use it, challenging its utility as a simple interpretability tool. We argue that reasoning texts should be treated as an artifact to be investigated, not taken at face value, and that understanding the non-human ways these models use language is a critical research direction.

We revisit the classical Sobel operator to ask a simple question: Are first-order edge maps sufficient to drive an all-dense multilayer perceptron (MLP) for handwritten character recognition (HCR), as an alternative to convolutional neural networks (CNNs)? Using only horizontal and vertical Sobel derivatives as input, we train an MLP on MNIST and EMNIST Letters. Despite its extreme simplicity, the resulting network reaches 98% accuracy on MNIST digits and 92% on EMNIST letters -- approaching CNNs while offering a smaller memory footprint and transparent features. Our findings highlight that much of the class-discriminative information in handwritten character images is already captured by first-order gradients, making edge-aware MLPs a compelling option for HCR.