Kevin Smith

Yewon Lee, Andrew Z. Li, Philip Huang et al. · mit

Planning for sequential robotics tasks often requires integrated symbolic and geometric reasoning. TAMP algorithms typically solve these problems by performing a tree search over high-level task sequences while checking for kinematic and dynamic feasibility. This can be inefficient because, typically, candidate task plans resulting from the tree search ignore geometric information. This often leads to motion planning failures that require expensive backtracking steps to find alternative task plans. We propose a novel approach to TAMP called Stein Task and Motion Planning (STAMP) that relaxes the hybrid optimization problem into a continuous domain. This allows us to leverage gradients from differentiable physics simulation to fully optimize discrete and continuous plan parameters for TAMP. In particular, we solve the optimization problem using a gradient-based variational inference algorithm called Stein Variational Gradient Descent. This allows us to find a distribution of solutions within a single optimization run. Furthermore, we use an off-the-shelf differentiable physics simulator that is parallelized on the GPU to run parallelized inference over diverse plan parameters. We demonstrate our method on a variety of problems and show that it can find multiple diverse plans in a single optimization run while also being significantly faster than existing approaches.

Mengzhou Hu, Sahar Alkhairy, Ingoo Lee et al.

Gene set analysis is a mainstay of functional genomics, but it relies on curated databases of gene functions that are incomplete. Here we evaluate five Large Language Models (LLMs) for their ability to discover the common biological functions represented by a gene set, substantiated by supporting rationale, citations and a confidence assessment. Benchmarking against canonical gene sets from the Gene Ontology, GPT-4 confidently recovered the curated name or a more general concept (73% of cases), while benchmarking against random gene sets correctly yielded zero confidence. Gemini-Pro and Mixtral-Instruct showed ability in naming but were falsely confident for random sets, whereas Llama2-70b had poor performance overall. In gene sets derived from 'omics data, GPT-4 identified novel functions not reported by classical functional enrichment (32% of cases), which independent review indicated were largely verifiable and not hallucinations. The ability to rapidly synthesize common gene functions positions LLMs as valuable 'omics assistants.

Mahesh Madhav, Allen Lee, Andres Mejia et al.

The march toward developing relevant and robust CPU benchmarks continues with the introduction of SPEC CPU 2026, the next generation suite for measuring processor performance. This paper details the methodology behind its creation, showcasing a process centered on community collaboration and principled development. The suite is built upon a foundation of modern, open-source applications, selected and hardened through a process that emphasizes workload diversity, portability, and software longevity. A key contribution is Rolling-Round-Robin Rate, a novel and standardized approach to running heterogeneous, multiprogrammed workloads that addresses a long-standing gap in benchmarking practice. Additionally, the suite features an expanded set of multithreaded benchmarks and introduces workloads with distinct microarchitectural profiles, reflecting the demands of contemporary software. By detailing our principled approach to benchmark selection, adaptation, and validation, we demonstrate how the SPEC CPU 2026 suite sets the standard for performance evaluation in the next era of computer architecture research and development.

Yue Liu, Christos Matsoukas, Fredrik Strand et al.

Vision transformers have demonstrated the potential to outperform CNNs in a variety of vision tasks. But the computational and memory requirements of these models prohibit their use in many applications, especially those that depend on high-resolution images, such as medical image classification. Efforts to train ViTs more efficiently are overly complicated, necessitating architectural changes or intricate training schemes. In this work, we show that standard ViT models can be efficiently trained at high resolution by randomly dropping input image patches. This simple approach, PatchDropout, reduces FLOPs and memory by at least 50% in standard natural image datasets such as ImageNet, and those savings only increase with image size. On CSAW, a high-resolution medical dataset, we observe a 5 times savings in computation and memory using PatchDropout, along with a boost in performance. For practitioners with a fixed computational or memory budget, PatchDropout makes it possible to choose image resolution, hyperparameters, or model size to get the most performance out of their model.

Christos Matsoukas, Johan Fredin Haslum, Moein Sorkhei et al.

Transfer learning is a standard technique to transfer knowledge from one domain to another. For applications in medical imaging, transfer from ImageNet has become the de-facto approach, despite differences in the tasks and image characteristics between the domains. However, it is unclear what factors determine whether - and to what extent - transfer learning to the medical domain is useful. The long-standing assumption that features from the source domain get reused has recently been called into question. Through a series of experiments on several medical image benchmark datasets, we explore the relationship between transfer learning, data size, the capacity and inductive bias of the model, as well as the distance between the source and target domain. Our findings suggest that transfer learning is beneficial in most cases, and we characterize the important role feature reuse plays in its success.

Christos Matsoukas, Johan Fredin Haslum, Moein Sorkhei et al.

Convolutional Neural Networks (CNNs) have reigned for a decade as the de facto approach to automated medical image diagnosis, pushing the state-of-the-art in classification, detection and segmentation tasks. Over the last years, vision transformers (ViTs) have appeared as a competitive alternative to CNNs, yielding impressive levels of performance in the natural image domain, while possessing several interesting properties that could prove beneficial for medical imaging tasks. In this work, we explore the benefits and drawbacks of transformer-based models for medical image classification. We conduct a series of experiments on several standard 2D medical image benchmark datasets and tasks. Our findings show that, while CNNs perform better if trained from scratch, off-the-shelf vision transformers can perform on par with CNNs when pretrained on ImageNet, both in a supervised and self-supervised setting, rendering them as a viable alternative to CNNs.

Johan Fredin Haslum, Christos Matsoukas, Karl-Johan Leuchowius et al.

High content imaging assays can capture rich phenotypic response data for large sets of compound treatments, aiding in the characterization and discovery of novel drugs. However, extracting representative features from high content images that can capture subtle nuances in phenotypes remains challenging. The lack of high-quality labels makes it difficult to achieve satisfactory results with supervised deep learning. Self-Supervised learning methods have shown great success on natural images, and offer an attractive alternative also to microscopy images. However, we find that self-supervised learning techniques underperform on high content imaging assays. One challenge is the undesirable domain shifts present in the data known as batch effects, which are caused by biological noise or uncontrolled experimental conditions. To this end, we introduce Cross-Domain Consistency Learning (CDCL), a self-supervised approach that is able to learn in the presence of batch effects. CDCL enforces the learning of biological similarities while disregarding undesirable batch-specific signals, leading to more useful and versatile representations. These features are organised according to their morphological changes and are more useful for downstream tasks -- such as distinguishing treatments and mechanism of action.

Lennart Alexander Van der Goten, Kevin Smith

MRI scans provide valuable medical information, however they also contain sensitive and personally identifiable information that needs to be protected. Whereas MRI metadata is easily sanitized, MRI image data is a privacy risk because it contains information to render highly-realistic 3D visualizations of a patient's head, enabling malicious actors to possibly identify the subject by cross-referencing a database. Data anonymization and de-identification is concerned with ensuring the privacy and confidentiality of individuals' personal information. Traditional MRI de-identification methods remove privacy-sensitive parts (e.g. eyes, nose etc.) from a given scan. This comes at the expense of introducing a domain shift that can throw off downstream analyses. In this work, we propose CP-MAE, a model that de-identifies the face by remodeling it (e.g. changing the face) rather than by removing parts using masked autoencoders. CP-MAE outperforms all previous approaches in terms of downstream task performance as well as de-identification. With our method we are able to synthesize high-fidelity scans of resolution up to $256^3$ -- compared to $128^3$ with previous approaches -- which constitutes an eight-fold increase in the number of voxels.

Lennart Alexander Van der Goten, Kevin Smith

Artifacts on magnetic resonance scans are a serious challenge for both radiologists and computer-aided diagnosis systems. Most commonly, artifacts are caused by motion of the patients, but can also arise from device-specific abnormalities such as noise patterns. Irrespective of the source, artifacts can not only render a scan useless, but can potentially induce misdiagnoses if left unnoticed. For instance, an artifact may masquerade as a tumor or other abnormality. Retrospective artifact correction (RAC) is concerned with removing artifacts after the scan has already been taken. In this work, we propose a method capable of retrospectively removing eight common artifacts found in native-resolution MR imagery. Knowledge of the presence or location of a specific artifact is not assumed and the system is, by design, capable of undoing interactions of multiple artifacts. Our method is realized through the design of a novel volumetric transformer-based neural network that generalizes a \emph{window-centered} approach popularized by the Swin transformer. Unlike Swin, our method is (i) natively volumetric, (ii) geared towards dense prediction tasks instead of classification, and (iii), uses a novel and more global mechanism to enable information exchange between windows. Our experiments show that our reconstructions are considerably better than those attained by ResNet, V-Net, MobileNet-v2, DenseNet, CycleGAN and BicycleGAN. Moreover, we show that the reconstructed images from our model improves the accuracy of FSL BET, a standard skull-stripping method typically applied in diagnostic workflows.

Joana Palés Huix, Adithya Raju Ganeshan, Johan Fredin Haslum et al.

The deep learning field is converging towards the use of general foundation models that can be easily adapted for diverse tasks. While this paradigm shift has become common practice within the field of natural language processing, progress has been slower in computer vision. In this paper we attempt to address this issue by investigating the transferability of various state-of-the-art foundation models to medical image classification tasks. Specifically, we evaluate the performance of five foundation models, namely SAM, SEEM, DINOv2, BLIP, and OpenCLIP across four well-established medical imaging datasets. We explore different training settings to fully harness the potential of these models. Our study shows mixed results. DINOv2 consistently outperforms the standard practice of ImageNet pretraining. However, other foundation models failed to consistently beat this established baseline indicating limitations in their transferability to medical image classification tasks.

Johan Fredin Haslum, Christos Matsoukas, Karl-Johan Leuchowius et al.

High Content Imaging (HCI) plays a vital role in modern drug discovery and development pipelines, facilitating various stages from hit identification to candidate drug characterization. Applying machine learning models to these datasets can prove challenging as they typically consist of multiple batches, affected by experimental variation, especially if different imaging equipment have been used. Moreover, as new data arrive, it is preferable that they are analyzed in an online fashion. To overcome this, we propose CODA, an online self-supervised domain adaptation approach. CODA divides the classifier's role into a generic feature extractor and a task-specific model. We adapt the feature extractor's weights to the new domain using cross-batch self-supervision while keeping the task-specific model unchanged. Our results demonstrate that this strategy significantly reduces the generalization gap, achieving up to a 300% improvement when applied to data from different labs utilizing different microscopes. CODA can be applied to new, unlabeled out-of-domain data sources of different sizes, from a single plate to multiple experimental batches.

Kevin Smith, Hai Lin, Praveen Tiwari et al.

Property Specification Language (PSL) is a form of temporal logic that has been mainly used in discrete domains (e.g. formal hardware verification). In this paper, we show that by merging machine learning techniques with PSL monitors, we can extend PSL to work on continuous domains. We apply this technique in machine learning-based anomaly detection to analyze scenarios of real-time streaming events from continuous variables in order to detect abnormal behaviors of a system. By using machine learning with formal models, we leverage the strengths of both machine learning methods and formal semantics of time. On one hand, machine learning techniques can produce distributions on continuous variables, where abnormalities can be captured as deviations from the distributions. On the other hand, formal methods can characterize discrete temporal behaviors and relations that cannot be easily learned by machine learning techniques. Interestingly, the anomalies detected by machine learning and the underlying time representation used are discrete events. We implemented a temporal monitoring package (TEF) that operates in conjunction with normal data science packages for anomaly detection machine learning systems, and we show that TEF can be used to perform accurate interpretation of temporal correlation between events.

Moein Sorkhei, Emir Konuk, Kevin Smith et al.

Existing adaptation techniques typically require architectural modifications or added parameters, leading to high computational costs and complexity. We introduce Attention Projection Layer Adaptation (APLA), a simple approach to adapt vision transformers (ViTs) without altering the architecture or adding parameters. Through a systematic analysis, we find that the layer immediately after the attention mechanism is crucial for adaptation. By updating only this projection layer, or even just a random subset of this layer's weights, APLA achieves state-of-the-art performance while reducing GPU memory usage by up to 52.63% and training time by up to 43.0%, with no extra cost at inference. Across 46 datasets covering a variety of tasks including scene classification, medical imaging, satellite imaging, and fine-grained classification, APLA consistently outperforms 17 other leading adaptation methods, including full fine-tuning, on classification, segmentation, and detection tasks. The code is available at https://github.com/MoeinSorkhei/APLA.

Robert Welch, Emir Konuk, Kevin Smith

Vision-language models (VLMs) are increasingly deployed in high-stakes settings where reliable uncertainty quantification (UQ) is as important as predictive accuracy. Extended reasoning via chain-of-thought (CoT) prompting or reasoning-trained models has become ubiquitous in modern VLM pipelines, yet its effect on UQ reliability remains poorly understood. We show that reasoning consistently degrades the quality of most uncertainty estimates, even when it improves task accuracy. We identify implicit answer conditioning as the primary mechanism: as reasoning traces converge on a conclusion before the final answer is generated, token probabilities increasingly reflect consistency with the model's own reasoning trace rather than uncertainty about correctness. In effect, the model becomes overconfident in its answer. In contrast, agreement-based consistency remains robust and often improves under reasoning, making it a practical choice for uncertainty estimation in reasoning-enabled VLMs.

Ethan Li, Anders Boesen Lindbo Larsen, Chen Zhang et al. · apple-ml, cmu

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Tony Chen, Sam Cheyette, Kelsey Allen et al.

Probabilistic mental simulation is thought to play a key role in human reasoning, planning, and prediction, yet the demands of simulation in complex environments exceed realistic human capacity limits. A theory with growing evidence is that people simulate using simplified representations of the environment that abstract away from irrelevant details, but it is unclear how people determine these simplifications efficiently. Here, we present a "Just-in-Time" framework for simulation-based reasoning that demonstrates how such representations can be constructed online with minimal added computation. The model uses a tight interleaving of simulation, visual search, and representation modification, with the current simulation guiding where to look and visual search flagging objects that should be encoded for subsequent simulation. Despite only ever encoding a small subset of objects, the model makes high-utility predictions. We find strong empirical support for this account over alternative models in a grid-world planning task and a physical reasoning task across a range of behavioral measures. Together, these results offer a concrete algorithmic account of how people construct reduced representations to support efficient mental simulation.

Amir Mehrpanah, Matteo Gamba, Kevin Smith et al.

ReLU networks, while prevalent for visual data, have sharp transitions, sometimes relying on individual pixels for predictions, making vanilla gradient-based explanations noisy and difficult to interpret. Existing methods, such as GradCAM, smooth these explanations by producing surrogate models at the cost of faithfulness. We introduce a unifying spectral framework to systematically analyze and quantify smoothness, faithfulness, and their trade-off in explanations. Using this framework, we quantify and regularize the contribution of ReLU networks to high-frequency information, providing a principled approach to identifying this trade-off. Our analysis characterizes how surrogate-based smoothing distorts explanations, leading to an ``explanation gap'' that we formally define and measure for different post-hoc methods. Finally, we validate our theoretical findings across different design choices, datasets, and ablations.

Moein Sorkhei, Emir Konuk, Jingyu Guo et al.

Self-supervised adaptation (SSA) improves foundation model transfer to medical domains but is computationally prohibitive. Although parameter efficient fine-tuning methods such as LoRA have been explored for supervised adaptation, their effectiveness for SSA remains unknown. In this work, we introduce efficient self-supervised adaptation (ESSA), a framework that applies parameter-efficient fine-tuning techniques to SSA with the aim of reducing computational cost and improving adaptation performance. Among the methods tested, Attention Projection Layer Adaptation (APLA) sets a new state-of-the-art, consistently surpassing full-parameter SSA and supervised fine-tuning across diverse medical tasks, while reducing GPU memory by up to 40.1% and increasing training throughput by 25.2%, all while maintaining inference efficiency.

Moein Sorkhei, Christos Matsoukas, Johan Fredin Haslum et al.

How well can one expect transfer learning to work in a new setting where the domain is shifted, the task is different, and the architecture changes? Many transfer learning metrics have been proposed to answer this question. But how accurate are their predictions in a realistic new setting? We conducted an extensive evaluation involving over 42,000 experiments comparing 23 transferability metrics across 16 different datasets to assess their ability to predict transfer performance. Our findings reveal that none of the existing metrics perform well across the board. However, we find that a simple k-nearest neighbor evaluation -- as is commonly used to evaluate feature quality for self-supervision -- not only surpasses existing metrics, but also offers better computational efficiency and ease of implementation.

Leonardo Scabini, Kallil M. Zielinski, Emir Konuk et al.

Texture recognition has recently been dominated by ImageNet-pre-trained deep Convolutional Neural Networks (CNNs), with specialized modifications and feature engineering required to achieve state-of-the-art (SOTA) performance. However, although Vision Transformers (ViTs) were introduced a few years ago, little is known about their texture recognition ability. Therefore, in this work, we introduce VORTEX (ViTs with Orderless and Randomized Token Encodings for Texture Recognition), a novel method that enables the effective use of ViTs for texture analysis. VORTEX extracts multi-depth token embeddings from pre-trained ViT backbones and employs a lightweight module to aggregate hierarchical features and perform orderless encoding, obtaining a better image representation for texture recognition tasks. This approach allows seamless integration with any ViT with the common transformer architecture. Moreover, no fine-tuning of the backbone is performed, since they are used only as frozen feature extractors, and the features are fed to a linear SVM. We evaluate VORTEX on nine diverse texture datasets, demonstrating its ability to achieve or surpass SOTA performance in a variety of texture analysis scenarios. By bridging the gap between texture recognition with CNNs and transformer-based architectures, VORTEX paves the way for adopting emerging transformer foundation models. Furthermore, VORTEX demonstrates robust computational efficiency when coupled with ViT backbones compared to CNNs with similar costs. The method implementation and experimental scripts are publicly available in our online repository.

Jingyu Guo, Christos Matsoukas, Fredrik Strand et al.

In multi-view medical diagnosis, deep learning-based models often fuse information from different imaging perspectives to improve diagnostic performance. However, existing approaches are prone to overfitting and rely heavily on view-specific features, which can lead to trivial solutions. In this work, we introduce Random Token Fusion (RTF), a novel technique designed to enhance multi-view medical image analysis using vision transformers. By integrating randomness into the feature fusion process during training, RTF addresses the issue of overfitting and enhances the robustness and accuracy of diagnostic models without incurring any additional cost at inference. We validate our approach on standard mammography and chest X-ray benchmark datasets. Through extensive experiments, we demonstrate that RTF consistently improves the performance of existing fusion methods, paving the way for a new generation of multi-view medical foundation models.

Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit et al.

Recent years have seen a significant progress in the general-purpose problem solving abilities of large vision and language models (LVLMs), such as ChatGPT, Gemini, etc.; some of these breakthroughs even seem to enable AI models to outperform human abilities in varied tasks that demand higher-order cognitive skills. Are the current large AI models indeed capable of generalized problem solving as humans do? A systematic analysis of AI capabilities for joint vision and text reasoning, however, is missing in the current scientific literature. In this paper, we make an effort towards filling this gap, by evaluating state-of-the-art LVLMs on their mathematical and algorithmic reasoning abilities using visuo-linguistic problems from children's Olympiads. Specifically, we consider problems from the Mathematical Kangaroo (MK) Olympiad, which is a popular international competition targeted at children from grades 1-12, that tests children's deeper mathematical abilities using puzzles that are appropriately gauged to their age and skills. Using the puzzles from MK, we created a dataset, dubbed SMART-840, consisting of 840 problems from years 2020-2024. With our dataset, we analyze LVLMs power on mathematical reasoning; their responses on our puzzles offer a direct way to compare against that of children. Our results show that modern LVLMs do demonstrate increasingly powerful reasoning skills in solving problems for higher grades, but lack the foundations to correctly answer problems designed for younger children. Further analysis shows that there is no significant correlation between the reasoning capabilities of AI models and that of young children, and their capabilities appear to be based on a different type of reasoning than the cumulative knowledge that underlies children's mathematics and logic skills.

Lennart Alexander Van der Goten, Jingyu Guo, Kevin Smith

The presence of motion artifacts in magnetic resonance imaging (MRI) scans poses a significant challenge, where even minor patient movements can lead to artifacts that may compromise the scan's utility.This paper introduces MAsked MOtion Correction (MAMOC), a novel method designed to address the issue of Retrospective Artifact Correction (RAC) in motion-affected MRI brain scans. MAMOC uses masked autoencoding self-supervision, transfer learning and test-time prediction to efficiently remove motion artifacts, producing high-fidelity, native-resolution scans. Until recently, realistic, openly available paired artifact presentations for training and evaluating retrospective motion correction methods did not exist, making it necessary to simulate motion artifacts. Leveraging the MR-ART dataset and bigger unlabeled datasets (ADNI, OASIS-3, IXI), this work is the first to evaluate motion correction in MRI scans using real motion data on a public dataset, showing that MAMOC achieves improved performance over existing motion correction methods.

Phillip Lee, Kevin Smith

Providing a provable privacy guarantees while maintaining the utility of data is a challenging task in many real-world applications. Recently, a new framework called One-Sided Differential Privacy (OSDP) was introduced that extends existing differential privacy approaches. OSDP increases the utility of the data by taking advantage of the fact that not all records are sensitive. However, the previous work assumed that all records are statistically independent from each other. Motivated by occupancy data in building management systems, this paper extends the existing one-sided differential privacy framework. In this paper, we quantify the overall privacy leakage when the adversary is given dependency information between the records. In addition, we show how an optimization problem can be constructed that efficiently trades off between the utility and privacy.

Moein Sorkhei, Yue Liu, Hossein Azizpour et al.

Interval and large invasive breast cancers, which are associated with worse prognosis than other cancers, are usually detected at a late stage due to false negative assessments of screening mammograms. The missed screening-time detection is commonly caused by the tumor being obscured by its surrounding breast tissues, a phenomenon called masking. To study and benchmark mammographic masking of cancer, in this work we introduce CSAW-M, the largest public mammographic dataset, collected from over 10,000 individuals and annotated with potential masking. In contrast to the previous approaches which measure breast image density as a proxy, our dataset directly provides annotations of masking potential assessments from five specialists. We also trained deep learning models on CSAW-M to estimate the masking level and showed that the estimated masking is significantly more predictive of screening participants diagnosed with interval and large invasive cancers -- without being explicitly trained for these tasks -- than its breast density counterparts.

Lennart Alexander Van der Goten, Tobias Hepp, Zeynep Akata et al.

Privacy protection of medical image data is challenging. Even if metadata is removed, brain scans are vulnerable to attacks that match renderings of the face to facial image databases. Solutions have been developed to de-identify diagnostic scans by obfuscating or removing parts of the face. However, these solutions either fail to reliably hide the patient's identity or are so aggressive that they impair further analyses. We propose a new class of de-identification techniques that, instead of removing facial features, remodels them. Our solution relies on a conditional multi-scale GAN architecture. It takes a patient's MRI scan as input and generates a 3D volume conditioned on the patient's brain, which is preserved exactly, but where the face has been de-identified through remodeling. We demonstrate that our approach preserves privacy far better than existing techniques, without compromising downstream medical analyses. Analyses were run on the OASIS-3 and ADNI corpora.

Christos Matsoukas, Johan Fredin Haslum, Magnus Söderberg et al.

Convolutional Neural Networks (CNNs) have reigned for a decade as the de facto approach to automated medical image diagnosis. Recently, vision transformers (ViTs) have appeared as a competitive alternative to CNNs, yielding similar levels of performance while possessing several interesting properties that could prove beneficial for medical imaging tasks. In this work, we explore whether it is time to move to transformer-based models or if we should keep working with CNNs - can we trivially switch to transformers? If so, what are the advantages and drawbacks of switching to ViTs for medical image diagnosis? We consider these questions in a series of experiments on three mainstream medical image datasets. Our findings show that, while CNNs perform better when trained from scratch, off-the-shelf vision transformers using default hyperparameters are on par with CNNs when pretrained on ImageNet, and outperform their CNN counterparts when pretrained using self-supervision.

Daniel M. Bear, Elias Wang, Damian Mrowca et al.

While current vision algorithms excel at many challenging tasks, it is unclear how well they understand the physical dynamics of real-world environments. Here we introduce Physion, a dataset and benchmark for rigorously evaluating the ability to predict how physical scenarios will evolve over time. Our dataset features realistic simulations of a wide range of physical phenomena, including rigid and soft-body collisions, stable multi-object configurations, rolling, sliding, and projectile motion, thus providing a more comprehensive challenge than previous benchmarks. We used Physion to benchmark a suite of models varying in their architecture, learning objective, input-output structure, and training data. In parallel, we obtained precise measurements of human prediction behavior on the same set of scenarios, allowing us to directly evaluate how well any model could approximate human behavior. We found that vision algorithms that learn object-centric representations generally outperform those that do not, yet still fall far short of human performance. On the other hand, graph neural networks with direct access to physical state information both perform substantially better and make predictions that are more similar to those made by humans. These results suggest that extracting physical representations of scenes is the main bottleneck to achieving human-level and human-like physical understanding in vision algorithms. We have publicly released all data and code to facilitate the use of Physion to benchmark additional models in a fully reproducible manner, enabling systematic evaluation of progress towards vision algorithms that understand physical environments as robustly as people do.

Yilun Du, Kevin Smith, Tomer Ulman et al.

We study the problem of unsupervised physical object discovery. While existing frameworks aim to decompose scenes into 2D segments based off each object's appearance, we explore how physics, especially object interactions, facilitates disentangling of 3D geometry and position of objects from video, in an unsupervised manner. Drawing inspiration from developmental psychology, our Physical Object Discovery Network (POD-Net) uses both multi-scale pixel cues and physical motion cues to accurately segment observable and partially occluded objects of varying sizes, and infer properties of those objects. Our model reliably segments objects on both synthetic and real scenes. The discovered object properties can also be used to reason about physical events.

Christos Matsoukas, Albert Bou I Hernandez, Yue Liu et al.

Evidence suggests that networks trained on large datasets generalize well not solely because of the numerous training examples, but also class diversity which encourages learning of enriched features. This raises the question of whether this remains true when data is scarce - is there an advantage to learning with additional labels in low-data regimes? In this work, we consider a task that requires difficult-to-obtain expert annotations: tumor segmentation in mammography images. We show that, in low-data settings, performance can be improved by complementing the expert annotations with seemingly uninformative labels from non-expert annotators, turning the task into a multi-class problem. We reveal that these gains increase when less expert data is available, and uncover several interesting properties through further studies. We demonstrate our findings on CSAW-S, a new dataset that we introduce here, and confirm them on two public datasets.

Yue Liu, Hossein Azizpour, Fredrik Strand et al.

The ability to accurately estimate risk of developing breast cancer would be invaluable for clinical decision-making. One promising new approach is to integrate image-based risk models based on deep neural networks. However, one must take care when using such models, as selection of training data influences the patterns the network will learn to identify. With this in mind, we trained networks using three different criteria to select the positive training data (i.e. images from patients that will develop cancer): an inherent risk model trained on images with no visible signs of cancer, a cancer signs model trained on images containing cancer or early signs of cancer, and a conflated model trained on all images from patients with a cancer diagnosis. We find that these three models learn distinctive features that focus on different patterns, which translates to contrasts in performance. Short-term risk is best estimated by the cancer signs model, whilst long-term risk is best estimated by the inherent risk model. Carelessly training with all images conflates inherent risk with early cancer signs, and yields sub-optimal estimates in both regimes. As a consequence, conflated models may lead physicians to recommend preventative action when early cancer signs are already visible.

Federico Baldassarre, Kevin Smith, Josephine Sullivan et al.

Visual relationship detection is fundamental for holistic image understanding. However, the localization and classification of (subject, predicate, object) triplets remain challenging tasks, due to the combinatorial explosion of possible relationships, their long-tailed distribution in natural images, and an expensive annotation process. This paper introduces a novel weakly-supervised method for visual relationship detection that relies on minimal image-level predicate labels. A graph neural network is trained to classify predicates in images from a graph representation of detected objects, implicitly encoding an inductive bias for pairwise relations. We then frame relationship detection as the explanation of such a predicate classifier, i.e. we obtain a complete relation by recovering the subject and object of a predicted predicate. We present results comparable to recent fully- and weakly-supervised methods on three diverse and challenging datasets: HICO-DET for human-object interaction, Visual Relationship Detection for generic object-to-object relations, and UnRel for unusual triplets; demonstrating robustness to non-comprehensive annotations and good few-shot generalization.

Emir Konuk, Kevin Smith

In this preregistration submission, we propose an empirical study of how networks handle changes in complexity of the data. We investigate the effect of network capacity on generalization performance in the face of increasing data complexity. For this, we measure the generalization error for an image classification task where the number of classes steadily increases. We compare a number of modern architectures at different scales in this setting. The methodology, setup, and hypotheses described in this proposal were evaluated by peer review before experiments were conducted.

Mattias Teye, Hossein Azizpour, Kevin Smith

We show that training a deep network using batch normalization is equivalent to approximate inference in Bayesian models. We further demonstrate that this finding allows us to make meaningful estimates of the model uncertainty using conventional architectures, without modifications to the network or the training procedure. Our approach is thoroughly validated by measuring the quality of uncertainty in a series of empirical experiments on different tasks. It outperforms baselines with strong statistical significance, and displays competitive performance with recent Bayesian approaches.