Babak Taati

Saeid Naeini, Raeid Saqur, Mozhgan Saeidi et al. · utoronto

The quest for human imitative AI has been an enduring topic in AI research since its inception. The technical evolution and emerging capabilities of the latest cohort of large language models (LLMs) have reinvigorated the subject beyond academia to the cultural zeitgeist. While recent NLP evaluation benchmark tasks test some aspects of human-imitative behaviour (e.g., BIG-bench's 'human-like behavior' tasks), few, if not none, examine creative problem solving abilities. Creative problem solving in humans is a well-studied topic in cognitive neuroscience with standardized tests that predominantly use the ability to associate (heterogeneous) connections among clue words as a metric for creativity. Exposure to misleading stimuli - distractors dubbed red herrings - impede human performance in such tasks via the fixation effect and Einstellung paradigm. In cognitive neuroscience studies, such fixations are experimentally induced by pre-exposing participants to orthographically similar incorrect words to subsequent word-fragments or clues. The popular British quiz show Only Connect's Connecting Wall segment essentially mimics Mednick's Remote Associates Test (RAT) formulation with built-in, deliberate red herrings, which makes it an ideal proxy dataset to explore and study fixation effect and Einstellung paradigm from cognitive neuroscience in LLMs. In this paper we present the novel Only Connect Wall (OCW) dataset and report results from our evaluation of selected pre-trained language models and LLMs on creative problem solving tasks like grouping clue words by heterogeneous connections, and identifying correct open knowledge domain connections in respective groups. We synthetically generate two additional datasets: OCW-Randomized, OCW-WordNet to further analyze our red-herrings hypothesis in language models. The code and link to the dataset are available at https://github.com/TaatiTeam/OCW.

Soroush Mehraban, Vida Adeli, Babak Taati

Recent transformer-based approaches have demonstrated excellent performance in 3D human pose estimation. However, they have a holistic view and by encoding global relationships between all the joints, they do not capture the local dependencies precisely. In this paper, we present a novel Attention-GCNFormer (AGFormer) block that divides the number of channels by using two parallel transformer and GCNFormer streams. Our proposed GCNFormer module exploits the local relationship between adjacent joints, outputting a new representation that is complementary to the transformer output. By fusing these two representation in an adaptive way, AGFormer exhibits the ability to better learn the underlying 3D structure. By stacking multiple AGFormer blocks, we propose MotionAGFormer in four different variants, which can be chosen based on the speed-accuracy trade-off. We evaluate our model on two popular benchmark datasets: Human3.6M and MPI-INF-3DHP. MotionAGFormer-B achieves state-of-the-art results, with P1 errors of 38.4mm and 16.2mm, respectively. Remarkably, it uses a quarter of the parameters and is three times more computationally efficient than the previous leading model on Human3.6M dataset. Code and models are available at https://github.com/TaatiTeam/MotionAGFormer.

Saeid Alavi Naeini, Leif Simmatis, Deniz Jafari et al.

Computer vision techniques can help automate or partially automate clinical examination of orofacial impairments to provide accurate and objective assessments. Towards the development of such automated systems, we evaluated two approaches to detect and temporally segment (parse) repetitions in orofacial assessment videos. Recorded videos of participants with amyotrophic lateral sclerosis (ALS) and healthy control (HC) individuals were obtained from the Toronto NeuroFace Dataset. Two approaches for repetition detection and parsing were examined: one based on engineered features from tracked facial landmarks and peak detection in the distance between the vermilion-cutaneous junction of the upper and lower lips (baseline analysis), and another using a pre-trained transformer-based deep learning model called RepNet (Dwibedi et al, 2020), which automatically detects periodicity, and parses periodic and semi-periodic repetitions in video data. In experimental evaluation of two orofacial assessments tasks, - repeating maximum mouth opening (OPEN) and repeating the sentence "Buy Bobby a Puppy" (BBP) - RepNet provided better parsing than the landmark-based approach, quantified by higher mean intersection-over-union (IoU) with respect to ground truth manual parsing. Automated parsing using RepNet also clearly separated HC and ALS participants based on the duration of BBP repetitions, whereas the landmark-based method could not.

Caroline Malin-Mayor, Vida Adeli, Andrea Sabo et al.

Video-based ambient monitoring of gait for older adults with dementia has the potential to detect negative changes in health and allow clinicians and caregivers to intervene early to prevent falls or hospitalizations. Computer vision-based pose tracking models can process video data automatically and extract joint locations; however, publicly available models are not optimized for gait analysis on older adults or clinical populations. In this work we train a deep neural network to map from a two dimensional pose sequence, extracted from a video of an individual walking down a hallway toward a wall-mounted camera, to a set of three-dimensional spatiotemporal gait features averaged over the walking sequence. The data of individuals with dementia used in this work was captured at two sites using a wall-mounted system to collect the video and depth information used to train and evaluate our model. Our Pose2Gait model is able to extract velocity and step length values from the video that are correlated with the features from the depth camera, with Spearman's correlation coefficients of .83 and .60 respectively, showing that three dimensional spatiotemporal features can be predicted from monocular video. Future work remains to improve the accuracy of other features, such as step time and step width, and test the utility of the predicted values for detecting meaningful changes in gait during longitudinal ambient monitoring.

Amirhossein Kazerouni, Maitreya Suin, Tristan Aumentado-Armstrong et al.

Recent advances in image restoration have enabled high-fidelity recovery of faces from degraded inputs using reference-based face restoration models (Ref-FR). However, such methods focus solely on facial regions, neglecting degradation across the full scene, including body and background, which limits practical usability. Meanwhile, full-scene restorers often ignore degradation cues entirely, leading to underdetermined predictions and visual artifacts. In this work, we propose Face2Scene, a two-stage restoration framework that leverages the face as a perceptual oracle to estimate degradation and guide the restoration of the entire image. Given a degraded image and one or more identity references, we first apply a Ref-FR model to reconstruct high-quality facial details. From the restored-degraded face pair, we extract a face-derived degradation code that captures degradation attributes (e.g., noise, blur, compression), which is then transformed into multi-scale degradation-aware tokens. These tokens condition a diffusion model to restore the full scene in a single step, including the body and background. Extensive experiments demonstrate the superior effectiveness of the proposed method compared to state-of-the-art methods.

Ahmadreza Jeddi, Hakki Can Karaimer, Hue Nguyen et al.

Recent reinforcement learning (RL) approaches like outcome-supervised GRPO have advanced chain-of-thought reasoning in Vision Language Models (VLMs), yet key issues linger: (i) reliance on costly and noisy hand-curated annotations or external verifiers; (ii) flat and sparse reward schemes in GRPO; and (iii) logical inconsistency between a chain's reasoning and its final answer. We present Puzzle Curriculum GRPO (PC-GRPO), a supervision-free recipe for RL with Verifiable Rewards (RLVR) that strengthens visual reasoning in VLMs without annotations or external verifiers. PC-GRPO replaces labels with three self-supervised puzzle environments: PatchFit, Rotation (with binary rewards) and Jigsaw (with graded partial credit mitigating reward sparsity). To counter flat rewards and vanishing group-relative advantages, we introduce a difficulty-aware curriculum that dynamically weights samples and peaks at medium difficulty. We further monitor Reasoning-Answer Consistency (RAC) during post-training: mirroring reports for vanilla GRPO in LLMs, RAC typically rises early then degrades; our curriculum delays this decline, and consistency-enforcing reward schemes further boost RAC. RAC correlates with downstream accuracy. Across diverse benchmarks and on Qwen-7B and Qwen-3B backbones, PC-GRPO improves reasoning quality, training stability, and end-task accuracy, offering a practical path to scalable, verifiable, and interpretable RL post-training for VLMs.

Javad Rajabi, Kimia Shaban, Koorosh Roohi et al.

Diffusion transformers (DiTs) have emerged as a dominant architecture for text-to-image generation, yet their performance drops when generating at resolutions beyond their training range. Existing training-free approaches mitigate this by modifying inference-time attention behavior, often through Rotary Position Embeddings (RoPE) extrapolation combined with attention scaling. However, these strategies apply a uniform and content-agnostic scaling across RoPE components with distinct frequency characteristics, inducing a trade-off between preserving global structure and recovering fine detail. We introduce SEGA, a training-free method that dynamically scales attention across RoPE components according to the latent's spatial-frequency structure at each denoising step. This adaptive scaling improves both structural coherence and fine-detail fidelity. Experiments show that SEGA consistently improves high-resolution synthesis across multiple target resolutions, outperforming state-of-the-art training-free baselines.

Alireza Morsali, MohammadJavad Vaez, Mohammadhossein Soltani et al.

Implicit Neural Representations (INRs) have emerged as a powerful framework for modeling continuous signals. The spectral bias of ReLU-based networks is a well-established limitation, restricting their ability to capture fine-grained details in target signals. While previous works have attempted to mitigate this issue through frequency-based encodings or architectural modifications, these approaches often introduce additional complexity and do not fully address the underlying challenge of learning high-frequency components efficiently. We introduce Sinusoidal Trainable Activation Functions (STAF), designed to directly tackle this limitation by enabling networks to adaptively learn and represent complex signals with higher precision and efficiency. STAF inherently modulates its frequency components, allowing for self-adaptive spectral learning. This capability significantly improves convergence speed and expressivity, making STAF highly effective for both signal representations and inverse problems. Through extensive evaluations across a range of tasks, including signal representation (shape, image, audio) and inverse problems (super-resolution, denoising), as well as neural radiance fields (NeRF), we demonstrate that STAF consistently outperforms state-of-the-art methods in accuracy and reconstruction fidelity. These results establish STAF as a robust solution to spectral bias and the capacity--convergence tradeoff, with broad applicability in computer vision and graphics. Our codebase is publicly accessible at https://github.com/AlirezaMorsali/STAF.

Emerson P. Grabke, Masoom A. Haider, Babak Taati

Objective: Latent diffusion models (LDM) could alleviate data scarcity challenges affecting machine learning development for medical imaging. However, medical LDM strategies typically rely on short-prompt text encoders, non-medical LDMs, or large data volumes. These strategies can limit performance and scientific accessibility. We propose a novel LDM conditioning approach to address these limitations. Methods: We propose Class-Conditioned Efficient Large Language model Adapter (CCELLA), a novel dual-head conditioning approach that simultaneously conditions the LDM U-Net with free-text clinical reports and radiology classification. We also propose a data-efficient LDM framework centered around CCELLA and a proposed joint loss function. We first evaluate our method on 3D prostate MRI against state-of-the-art. We then augment a downstream classifier model training dataset with synthetic images from our method. Results: Our method achieves a 3D FID score of 0.025 on a size-limited 3D prostate MRI dataset, significantly outperforming a recent foundation model with FID 0.071. When training a classifier for prostate cancer prediction, adding synthetic images generated by our method during training improves classifier accuracy from 69% to 74%. Training a classifier solely on our method's synthetic images achieved comparable performance to training on real images alone. Conclusion: We show that our method improved both synthetic image quality and downstream classifier performance using limited data and minimal human annotation. Significance: The proposed CCELLA-centric framework enables radiology report and class-conditioned LDM training for high-quality medical image synthesis given limited data volume and human data annotation, improving LDM performance and scientific accessibility. Code from this study will be available at https://github.com/grabkeem/CCELLA

Soroush Mehraban, Mohammad Javad Rajabi, Andrea Iaboni et al.

Self-supervised pretraining methods with masked prediction demonstrate remarkable within-dataset performance in skeleton-based action recognition. However, we show that, unlike contrastive learning approaches, they do not produce well-separated clusters. Additionally, these methods struggle with generalization in few-shot settings. To address these issues, we propose Self-supervised Tuning for 3D Action Recognition in Skeleton sequences (STARS). Specifically, STARS first uses a masked prediction stage using an encoder-decoder architecture. It then employs nearest-neighbor contrastive learning to partially tune the weights of the encoder, enhancing the formation of semantic clusters for different actions. By tuning the encoder for a few epochs, and without using hand-crafted data augmentations, STARS achieves state-of-the-art self-supervised results in various benchmarks, including NTU-60, NTU-120, and PKU-MMD. In addition, STARS exhibits significantly better results than masked prediction models in few-shot settings, where the model has not seen the actions throughout pretraining. Project page: https://soroushmehraban.github.io/stars/

Abhishek Moturu, Anna Goldenberg, Babak Taati

Synthetic data is useful only when the added samples fill missing parts of the training distribution that matter for the downstream task. We introduce LiBaGS, a lightweight, generator-agnostic method for targeted synthetic training data selection. LiBaGS scores candidate synthetic samples by combining decision-boundary proximity, predictive uncertainty, real-data density, and support validity, so that selected samples are both informative and likely to remain on the real data manifold. We then use a boundary-gap allocation rule that targets sparse but realistic decision-boundary neighborhoods, rather than simply adding more data or selecting only the most uncertain candidates. LiBaGS also learns when enough synthetic samples have been added through a marginal-value stopping rule, assigns softer labels near ambiguous boundaries, and uses a diversity objective to avoid redundant near-duplicate selections. Experiments show that LiBaGS improves accuracy over classical oversampling, hard augmentation, uncertainty and density ablations, and targeted-generation selection criteria.

Ahmadreza Jeddi, Kimia Shaban, Negin Baghbanzadeh et al.

Reinforcement learning (RL) is increasingly used to post-train medical Vision-Language Models (VLMs), yet it remains unclear whether RL improves medical visual reasoning or mainly sharpens behaviors already induced by supervised fine-tuning (SFT). We present a controlled study that disentangles these effects along three axes: vision, SFT, and RL. Using MedMNIST as a multi-modality testbed, we probe visual perception by benchmarking VLM vision towers against vision-only baselines, quantify reasoning support and sampling efficiency via Accuracy@1 versus Pass@K, and evaluate when RL closes the support gap and how gains transfer across modalities. We find that RL is most effective when the model already has non-trivial support (high Pass@K): it primarily sharpens the output distribution, improving Acc@1 and sampling efficiency, while SFT expands support and makes RL effective. Based on these findings, we propose a boundary-aware recipe and instantiate it by RL post-training an OctoMed-initialized model on a small, balanced subset of PMC multiple-choice VQA, achieving strong average performance across six medical VQA benchmarks.

Emerson P. Grabke, Babak Taati, Masoom A. Haider

Objective: Latent diffusion models (LDMs) could mitigate data scarcity challenges affecting machine learning development for medical image interpretation. The recent CCELLA LDM improved prostate cancer detection performance using synthetic MRI for classifier training but was limited to the axial T2-weighted (AxT2) sequence, did not investigate inter-institutional domain shift, and prioritized radiology over histopathology outcomes. We propose CCELLA++ to address these limitations and improve clinical utility. Methods: CCELLA++ expands CCELLA for simultaneous biparametric prostate MRI (bpMRI) generation, including the AxT2, high b-value diffusion series (HighB) and apparent diffusion coefficient map (ADC). Domain adaptation was investigated by pretraining classifiers on real or LDM-generated synthetic data from an internal institution, followed with fine-tuning on progressively smaller fractions of an out-of-distribution, external dataset. Results: CCELLA++ improved 3D FID for HighB and ADC but not AxT2 (0.013, 0.012, 0.063 respectively) sequences compared to CCELLA (0.060). Classifier pretraining with CCELLA++ bpMRI outperformed real bpMRI in AP and AUC for all domain adaptation scenarios. CCELLA++ pretraining achieved highest classifier performance below 50% (n=665) external dataset volume. Conclusion: Synthetic bpMRI generated by our method can improve downstream classifier generalization and performance beyond real bpMRI or CCELLA-generated AxT2-only images. Future work should seek to quantify medical image sample quality, balance multi-sequence LDM training, and condition the LDM with additional information. Significance: The proposed CCELLA++ LDM can generate synthetic bpMRI that outperforms real data for domain adaptation with a limited target institution dataset. Our code is available at https://github.com/grabkeem/CCELLA-plus-plus

Andrea Sabo, Sina Mehdizadeh, Andrea Iaboni et al.

Drug-induced parkinsonism affects many older adults with dementia, often causing gait disturbances. New advances in vision-based human pose-estimation have opened possibilities for frequent and unobtrusive analysis of gait in residential settings. This work leverages novel spatial-temporal graph convolutional network (ST-GCN) architectures and training procedures to predict clinical scores of parkinsonism in gait from video of individuals with dementia. We propose a two-stage training approach consisting of a self-supervised pretraining stage that encourages the ST-GCN model to learn about gait patterns before predicting clinical scores in the finetuning stage. The proposed ST-GCN models are evaluated on joint trajectories extracted from video and are compared against traditional (ordinal, linear, random forest) regression models and temporal convolutional network baselines. Three 2D human pose-estimation libraries (OpenPose, Detectron, AlphaPose) and the Microsoft Kinect (2D and 3D) are used to extract joint trajectories of 4787 natural walking bouts from 53 older adults with dementia. A subset of 399 walks from 14 participants is annotated with scores of parkinsonism severity on the gait criteria of the Unified Parkinson's Disease Rating Scale (UPDRS) and the Simpson-Angus Scale (SAS). Our results demonstrate that ST-GCN models operating on 3D joint trajectories extracted from the Kinect consistently outperform all other models and feature sets. Prediction of parkinsonism scores in natural walking bouts of unseen participants remains a challenging task, with the best models achieving macro-averaged F1-scores of 0.53 +/- 0.03 and 0.40 +/- 0.02 for UPDRS-gait and SAS-gait, respectively. Pre-trained model and demo code for this work is available: https://github.com/TaatiTeam/stgcn_parkinsonism_prediction.

Siavash Rezaei, Abhishek Moturu, Shun Zhao et al.

Although pain is frequent in old age, older adults are often undertreated for pain. This is especially the case for long-term care residents with moderate to severe dementia who cannot report their pain because of cognitive impairments that accompany dementia. Nursing staff acknowledge the challenges of effectively recognizing and managing pain in long-term care facilities due to lack of human resources and, sometimes, expertise to use validated pain assessment approaches on a regular basis. Vision-based ambient monitoring will allow for frequent automated assessments so care staff could be automatically notified when signs of pain are displayed. However, existing computer vision techniques for pain detection are not validated on faces of older adults or people with dementia, and this population is not represented in existing facial expression datasets of pain. We present the first fully automated vision-based technique validated on a dementia cohort. Our contributions are threefold. First, we develop a deep learning-based computer vision system for detecting painful facial expressions on a video dataset that is collected unobtrusively from older adult participants with and without dementia. Second, we introduce a pairwise comparative inference method that calibrates to each person and is sensitive to changes in facial expression while using training data more efficiently than sequence models. Third, we introduce a fast contrastive training method that improves cross-dataset performance. Our pain estimation model outperforms baselines by a wide margin, especially when evaluated on faces of people with dementia. Pre-trained model and demo code available at https://github.com/TaatiTeam/pain_detection_demo

Ahmadreza Jeddi, Minh Ngoc Le, Hakki C. Karaimer et al.

Autonomous research agents can already run machine learning experiments without human supervision, but many rely on a narrow search strategy: they repeatedly modify one program and keep changes only when they improve the current best result. This can cause them to discard useful partial ideas, alternative promising directions, and insights from failed or incomplete experiments. GEAR, or Genetic AutoResearch, replaces this single-path search with a population-based search over multiple research states. It keeps a set of strong candidate solutions, selects parents based on productivity, novelty, and coverage, and explores new ideas through mutation and crossover. Each research state stores its code changes, reflections, and performance data, allowing future decisions to build on past discoveries. The paper studies three versions of GEAR: one controlled through prompting, one using a fixed programmatic search controller, and one where the controller itself can evolve during the run. Under the same compute budget and environment, all three versions outperform the AutoResearch baseline. More importantly, while the baseline tends to settle into one local optimum, GEAR continues finding improvements over longer runs. Overall, the results suggest that autonomous research agents become more effective when they maintain multiple promising directions and can adapt their search strategy over time.

Ahmadreza Jeddi, Negin Baghbanzadeh, Elham Dolatabadi et al.

The computational demands of Vision Transformers (ViTs) and Vision-Language Models (VLMs) remain a significant challenge due to the quadratic complexity of self-attention. While token pruning offers a promising solution, existing methods often introduce training overhead or fail to adapt dynamically across layers. We present SAINT, a training-free token pruning framework that leverages token similarity and a graph-based formulation to dynamically optimize pruning rates and redundancy thresholds. Through systematic analysis, we identify a universal three-stage token evolution process (aligner-explorer-aggregator) in transformers, enabling aggressive pruning in early stages without sacrificing critical information. For ViTs, SAINT doubles the throughput of ViT-H/14 at 224px with only 0.6% accuracy loss on ImageNet-1K, surpassing the closest competitor by 0.8%. For VLMs, we apply SAINT in three modes: ViT-only, LLM-only, and hybrid. SAINT reduces LLaVA-13B's tokens by 75%, achieving latency comparable to LLaVA-7B with less than 1% performance loss across benchmarks. Our work establishes a unified, practical framework for efficient inference in ViTs and VLMs.

Javad Rajabi, Soroush Mehraban, Seyedmorteza Sadat et al.

Classifier-free guidance (CFG) has become an essential component of modern diffusion models to enhance both generation quality and alignment with input conditions. However, CFG requires specific training procedures and is limited to conditional generation. To address these limitations, we propose Token Perturbation Guidance (TPG), a novel method that applies perturbation matrices directly to intermediate token representations within the diffusion network. TPG employs a norm-preserving shuffling operation to provide effective and stable guidance signals that improve generation quality without architectural changes. As a result, TPG is training-free and agnostic to input conditions, making it readily applicable to both conditional and unconditional generation. We further analyze the guidance term provided by TPG and show that its effect on sampling more closely resembles CFG compared to existing training-free guidance techniques. Extensive experiments on SDXL and Stable Diffusion 2.1 show that TPG achieves nearly a 2$\times$ improvement in FID for unconditional generation over the SDXL baseline, while closely matching CFG in prompt alignment. These results establish TPG as a general, condition-agnostic guidance method that brings CFG-like benefits to a broader class of diffusion models.

Amirhossein Kazerouni, Soroush Mehraban, Michael Brudno et al.

Implicit Neural Representations (INRs) are proving to be a powerful paradigm in unifying task modeling across diverse data domains, offering key advantages such as memory efficiency and resolution independence. Conventional deep learning models are typically modality-dependent, often requiring custom architectures and objectives for different types of signals. However, existing INR frameworks frequently rely on global latent vectors or exhibit computational inefficiencies that limit their broader applicability. We introduce LIFT, a novel, high-performance framework that addresses these challenges by capturing multiscale information through meta-learning. LIFT leverages multiple parallel localized implicit functions alongside a hierarchical latent generator to produce unified latent representations that span local, intermediate, and global features. This architecture facilitates smooth transitions across local regions, enhancing expressivity while maintaining inference efficiency. Additionally, we introduce ReLIFT, an enhanced variant of LIFT that incorporates residual connections and expressive frequency encodings. With this straightforward approach, ReLIFT effectively addresses the convergence-capacity gap found in comparable methods, providing an efficient yet powerful solution to improve capacity and speed up convergence. Empirical results show that LIFT achieves state-of-the-art (SOTA) performance in generative modeling and classification tasks, with notable reductions in computational costs. Moreover, in single-task settings, the streamlined ReLIFT architecture proves effective in signal representations and inverse problem tasks.

Vahid Balazadeh, Michael Cooper, David Pellow et al. · utoronto

We present the design process and findings of the pre-conference workshop at the Machine Learning for Healthcare Conference (2024) entitled Red Teaming Large Language Models for Healthcare, which took place on August 15, 2024. Conference participants, comprising a mix of computational and clinical expertise, attempted to discover vulnerabilities -- realistic clinical prompts for which a large language model (LLM) outputs a response that could cause clinical harm. Red-teaming with clinicians enables the identification of LLM vulnerabilities that may not be recognised by LLM developers lacking clinical expertise. We report the vulnerabilities found, categorise them, and present the results of a replication study assessing the vulnerabilities across all LLMs provided.

Babak Taati, Muhammad Muzammil, Yasamin Zarghami et al. · utoronto

Accurate pain assessment in patients with limited ability to communicate, such as older adults with dementia, represents a critical healthcare challenge. Robust automated systems of pain detection may facilitate such assessments. Existing pain detection datasets, however, suffer from limited ethnic/racial diversity, privacy constraints, and underrepresentation of older adults who are the primary target population for clinical deployment. We present SynPAIN, a large-scale synthetic dataset containing 10,710 facial expression images (5,355 neutral/expressive pairs) across five ethnicities/races, two age groups (young: 20-35, old: 75+), and two genders. Using commercial generative AI tools, we created demographically balanced synthetic identities with clinically meaningful pain expressions. Our validation demonstrates that synthetic pain expressions exhibit expected pain patterns, scoring significantly higher than neutral and non-pain expressions using clinically validated pain assessment tools based on facial action unit analysis. We experimentally demonstrate SynPAIN's utility in identifying algorithmic bias in existing pain detection models. Through comprehensive bias evaluation, we reveal substantial performance disparities across demographic characteristics. These performance disparities were previously undetectable with smaller, less diverse datasets. Furthermore, we demonstrate that age-matched synthetic data augmentation improves pain detection performance on real clinical data, achieving a 7.0% improvement in average precision. SynPAIN addresses critical gaps in pain assessment research by providing the first publicly available, demographically diverse synthetic dataset specifically designed for older adult pain detection, while establishing a framework for measuring and mitigating algorithmic bias. The dataset is available at https://doi.org/10.5683/SP3/WCXMAP

Vida Adeli, Soroush Mehraban, Majid Mirmehdi et al.

Gait analysis is crucial for the diagnosis and monitoring of movement disorders like Parkinson's Disease. While computer vision models have shown potential for objectively evaluating parkinsonian gait, their effectiveness is limited by scarce clinical datasets and the challenge of collecting large and well-labelled data, impacting model accuracy and risk of bias. To address these gaps, we propose GAITGen, a novel framework that generates realistic gait sequences conditioned on specified pathology severity levels. GAITGen employs a Conditional Residual Vector Quantized Variational Autoencoder to learn disentangled representations of motion dynamics and pathology-specific factors, coupled with Mask and Residual Transformers for conditioned sequence generation. GAITGen generates realistic, diverse gait sequences across severity levels, enriching datasets and enabling large-scale model training in parkinsonian gait analysis. Experiments on our new PD-GaM (real) dataset demonstrate that GAITGen outperforms adapted state-of-the-art models in both reconstruction fidelity and generation quality, accurately capturing critical pathology-specific gait features. A clinical user study confirms the realism and clinical relevance of our generated sequences. Moreover, incorporating GAITGen-generated data into downstream tasks improves parkinsonian gait severity estimation, highlighting its potential for advancing clinical gait analysis.

Ahmadreza Jeddi, Marco Ciccone, Babak Taati

Looped Transformers have emerged as an efficient and powerful class of models for reasoning in the language domain. Recent studies show that these models achieve strong performance on algorithmic and reasoning tasks, suggesting that looped architectures possess an inductive bias toward latent reasoning. However, prior approaches fix the number of loop iterations during training and inference, leaving open the question of whether these models can flexibly adapt their computational depth under variable compute budgets. We introduce LoopFormer, a looped Transformer trained on variable-length trajectories to enable budget-conditioned reasoning. Our core contribution is a shortcut-consistency training scheme that aligns trajectories of different lengths, ensuring that shorter loops yield informative representations while longer loops continue to refine them. LoopFormer conditions each loop on the current time and step size, enabling representations to evolve consistently across trajectories of varying length rather than drifting or stagnating. Empirically, LoopFormer demonstrates robust performance on language modeling and reasoning benchmarks even under aggressive compute constraints, while scaling gracefully with additional budget. These results show that looped Transformers are inherently suited for adaptive language modeling, opening a path toward controllable and budget-aware large language models.

Soroush Mehraban, Andrea Iaboni, Babak Taati

Recent transformer-based models for 3D Human Mesh Recovery (HMR) have achieved strong performance but often suffer from high computational cost and complexity due to deep transformer architectures and redundant tokens. In this paper, we introduce two HMR-specific merging strategies: Error-Constrained Layer Merging (ECLM) and Mask-guided Token Merging (Mask-ToMe). ECLM selectively merges transformer layers that have minimal impact on the Mean Per Joint Position Error (MPJPE), while Mask-ToMe focuses on merging background tokens that contribute little to the final prediction. To further address the potential performance drop caused by merging, we propose a diffusion-based decoder that incorporates temporal context and leverages pose priors learned from large-scale motion capture datasets. Experiments across multiple benchmarks demonstrate that our method achieves up to 2.3x speed-up while slightly improving performance over the baseline.

Soroush Mehraban, Vida Adeli, Jacob Rommann et al.

We address the task of video style transfer with diffusion models, where the goal is to preserve the context of an input video while rendering it in a target style specified by a text prompt. A major challenge is the lack of paired video data for supervision. We propose PickStyle, a video-to-video style transfer framework that augments pretrained video diffusion backbones with style adapters and benefits from paired still image data with source-style correspondences for training. PickStyle inserts low-rank adapters into the self-attention layers of conditioning modules, enabling efficient specialization for motion-style transfer while maintaining strong alignment between video content and style. To bridge the gap between static image supervision and dynamic video, we construct synthetic training clips from paired images by applying shared augmentations that simulate camera motion, ensuring temporal priors are preserved. In addition, we introduce Context-Style Classifier-Free Guidance (CS-CFG), a novel factorization of classifier-free guidance into independent text (style) and video (context) directions. CS-CFG ensures that context is preserved in generated video while the style is effectively transferred. Experiments across benchmarks show that our approach achieves temporally coherent, style-faithful, and content-preserving video translations, outperforming existing baselines both qualitatively and quantitatively.

Vida Adeli, Ivan Klabucar, Javad Rajabi et al.

Objective gait assessment in Parkinson's Disease (PD) is limited by the absence of large, diverse, and clinically annotated motion datasets. We introduce CARE-PD, the largest publicly available archive of 3D mesh gait data for PD, and the first multi-site collection spanning 9 cohorts from 8 clinical centers. All recordings (RGB video or motion capture) are converted into anonymized SMPL meshes via a harmonized preprocessing pipeline. CARE-PD supports two key benchmarks: supervised clinical score prediction (estimating Unified Parkinson's Disease Rating Scale, UPDRS, gait scores) and unsupervised motion pretext tasks (2D-to-3D keypoint lifting and full-body 3D reconstruction). Clinical prediction is evaluated under four generalization protocols: within-dataset, cross-dataset, leave-one-dataset-out, and multi-dataset in-domain adaptation. To assess clinical relevance, we compare state-of-the-art motion encoders with a traditional gait-feature baseline, finding that encoders consistently outperform handcrafted features. Pretraining on CARE-PD reduces MPJPE (from 60.8mm to 7.5mm) and boosts PD severity macro-F1 by 17 percentage points, underscoring the value of clinically curated, diverse training data. CARE-PD and all benchmark code are released for non-commercial research at https://neurips2025.care-pd.ca/.

Abhishek Moturu, Anna Goldenberg, Babak Taati

Training deep neural networks in the presence of noisy labels and data heterogeneity is a major challenge. We introduce Lightweight Learnable Adaptive Weighting (LiLAW), a novel method that dynamically adjusts the loss weight of each training sample based on its evolving difficulty level, categorized as easy, moderate, or hard. Using only three learnable parameters, LiLAW adaptively prioritizes informative samples throughout training by updating these weights using a single mini-batch gradient descent step on the validation set after each training mini-batch, without requiring excessive hyperparameter tuning or a clean validation set. Extensive experiments across multiple general and medical imaging datasets, noise levels and types, loss functions, and architectures with and without pretraining demonstrate that LiLAW consistently enhances performance, even in high-noise environments. It is effective without heavy reliance on data augmentation or advanced regularization, highlighting its practicality. It offers a computationally efficient solution to boost model generalization and robustness in any neural network training setup.

Xin Lei Lin, Soroush Mehraban, Abhishek Moturu et al.

Automated pain assessment from facial expressions is crucial for non-communicative patients, such as those with dementia. Progress has been limited by two challenges: (i) existing datasets exhibit severe demographic and label imbalance due to ethical constraints, and (ii) current generative models cannot precisely control facial action units (AUs), facial structure, or clinically validated pain levels. We present 3DPain, a large-scale synthetic dataset specifically designed for automated pain assessment, featuring unprecedented annotation richness and demographic diversity. Our three-stage framework generates diverse 3D meshes, textures them with diffusion models, and applies AU-driven face rigging to synthesize multi-view faces with paired neutral and pain images, AU configurations, PSPI scores, and the first dataset-level annotations of pain-region heatmaps. The dataset comprises 82,500 samples across 25,000 pain expression heatmaps and 2,500 synthetic identities balanced by age, gender, and ethnicity. We further introduce ViTPain, a Vision Transformer based cross-modal distillation framework in which a heatmap-trained teacher guides a student trained on RGB images, enhancing accuracy, interpretability, and clinical reliability. Together, 3DPain and ViTPain establish a controllable, diverse, and clinically grounded foundation for generalizable automated pain assessment.

Alireza Hosseini, Amirhossein Kazerouni, Saeed Akhavan et al.

Visual attention modeling, important for interpreting and prioritizing visual stimuli, plays a significant role in applications such as marketing, multimedia, and robotics. Traditional saliency prediction models, especially those based on Convolutional Neural Networks (CNNs) or Transformers, achieve notable success by leveraging large-scale annotated datasets. However, the current state-of-the-art (SOTA) models that use Transformers are computationally expensive. Additionally, separate models are often required for each image type, lacking a unified approach. In this paper, we propose Saliency Unification through Mamba (SUM), a novel approach that integrates the efficient long-range dependency modeling of Mamba with U-Net to provide a unified model for diverse image types. Using a novel Conditional Visual State Space (C-VSS) block, SUM dynamically adapts to various image types, including natural scenes, web pages, and commercial imagery, ensuring universal applicability across different data types. Our comprehensive evaluations across five benchmarks demonstrate that SUM seamlessly adapts to different visual characteristics and consistently outperforms existing models. These results position SUM as a versatile and powerful tool for advancing visual attention modeling, offering a robust solution universally applicable across different types of visual content.

Diego L. Guarin, Aidan Dempster, Andrea Bandini et al.

Orofacial deficits are common in people with Parkinson's disease (PD) and their evolution might represent an important biomarker of disease progression. We are developing an automated system for assessment of orofacial function in PD that can be used in-home or in-clinic and can provide useful and objective clinical information that informs disease management. Our current approach relies on color and depth cameras for the estimation of 3D facial movements. However, depth cameras are not commonly available, might be expensive, and require specialized software for control and data processing. The objective of this paper was to evaluate if depth cameras are needed to differentiate between healthy controls and PD patients based on features extracted from orofacial kinematics. Results indicate that 2D features, extracted from color cameras only, are as informative as 3D features, extracted from color and depth cameras, differentiating healthy controls from PD patients. These results pave the way for the development of a universal system for automatic and objective assessment of orofacial function in PD.

Diego L. Guarin, Yana Yunusova, Babak Taati et al.

Importance: Machine learning (ML) approaches to facial landmark localization carry great clinical potential for quantitative assessment of facial function as they enable high-throughput automated quantification of relevant facial metrics from photographs. However, translation from research settings to clinical applications requires important improvements. Objective: To develop an ML algorithm for accurate facial landmarks localization in photographs of facial palsy patients, and use it as part of an automated computer-aided diagnosis system. Design, Setting, and Participants: Facial landmarks were manually localized in portrait photographs of eight expressions obtained from 200 facial palsy patients and 10 controls. A novel ML model for automated facial landmark localization was trained using this disease-specific database. Model output was compared to manual annotations and the output of a model trained using a larger database consisting only of healthy subjects. Model accuracy was evaluated by the normalized root mean square error (NRMSE) between algorithms' prediction and manual annotations. Results: Publicly available algorithms provide poor results when applied to patients compared to healthy controls (NRMSE, 8.56 +/- 2.16 vs. 7.09 +/- 2.34, p << 0.01). We found significant improvement in facial landmark localization accuracy for the clinical population when using a model trained with a relatively small number patients' photographs (1440) compared to a model trained using several thousand more images of healthy faces (NRMSE, 6.03 +/- 2.43 vs. 8.56 +/- 2.16, p << 0.01). Conclusions: Retraining a landmark detection model with a small number of clinical images significantly improved landmark detection performance in frontal view photographs of the clinical population. These results represent the first steps towards an automatic system for computer-aided assessment in facial palsy.

Azin Asgarian, Shun Zhao, Ahmed B. Ashraf et al.

Accurate facial expression analysis is an essential step in various clinical applications that involve physical and mental health assessments of older adults (e.g. diagnosis of pain or depression). Although remarkable progress has been achieved toward developing robust facial landmark detection methods, state-of-the-art methods still face many challenges when encountering uncontrolled environments, different ranges of facial expressions, and different demographics of the population. A recent study has revealed that the health status of individuals can also affect the performance of facial landmark detection methods on front views of faces. In this work, we investigate this matter in a much greater context using seven facial landmark detection methods. We perform our evaluation not only on frontal faces but also on profile faces and in various regions of the face. Our results shed light on limitations of the existing methods and challenges of applying these methods in clinical settings by indicating: 1) a significant difference between the performance of state-of-the-art when tested on the profile or frontal faces of individuals with vs. without dementia; 2) insights on the existing bias for all regions of the face; and 3) the presence of this bias despite re-training/fine-tuning with various configurations of six datasets.

Azin Asgarian, Parinaz Sobhani, Ji Chao Zhang et al.

In recent years, supervised machine learning models have demonstrated tremendous success in a variety of application domains. Despite the promising results, these successful models are data hungry and their performance relies heavily on the size of training data. However, in many healthcare applications it is difficult to collect sufficiently large training datasets. Transfer learning can help overcome this issue by transferring the knowledge from readily available datasets (source) to a new dataset (target). In this work, we propose a hybrid instance-based transfer learning method that outperforms a set of baselines including state-of-the-art instance-based transfer learning approaches. Our method uses a probabilistic weighting strategy to fuse information from the source domain to the model learned in the target domain. Our method is generic, applicable to multiple source domains, and robust with respect to negative transfer. We demonstrate the effectiveness of our approach through extensive experiments for two different applications.

Ahmed Ashraf, Shehroz Khan, Nikhil Bhagwat et al.

Medical imaging machine learning algorithms are usually evaluated on a single dataset. Although training and testing are performed on different subsets of the dataset, models built on one study show limited capability to generalize to other studies. While database bias has been recognized as a serious problem in the computer vision community, it has remained largely unnoticed in medical imaging research. Transfer learning thus remains confined to the re-use of feature representations requiring re-training on the new dataset. As a result, machine learning models do not generalize even when trained on imaging datasets that were captured to study the same variable of interest. The ability to transfer knowledge gleaned from one study to another, without the need for re-training, if possible, would provide reassurance that the models are learning knowledge fundamental to the problem under study instead of latching onto the idiosyncracies of a dataset. In this paper, we situate the problem of dataset bias in the context of medical imaging studies. We show empirical evidence that such a problem exists in medical datasets. We then present a framework to unlearn study membership as a means to handle the problem of database bias. Our main idea is to take the data from the original feature space to an intermediate space where the data points are indistinguishable in terms of which study they come from, while maintaining the recognition capability with respect to the variable of interest. This will promote models which learn the more general properties of the etiology under study instead of aligning to dataset-specific peculiarities. Essentially, our proposed model learns to unlearn the dataset bias.

Azin Asgarian, Ahmed Bilal Ashraf, David Fleet et al.

Active appearance models (AAMs) are a class of generative models that have seen tremendous success in face analysis. However, model learning depends on the availability of detailed annotation of canonical landmark points. As a result, when accurate AAM fitting is required on a different set of variations (expression, pose, identity), a new dataset is collected and annotated. To overcome the need for time consuming data collection and annotation, transfer learning approaches have received recent attention. The goal is to transfer knowledge from previously available datasets (source) to a new dataset (target). We propose a subspace transfer learning method, in which we select a subspace from the source that best describes the target space. We propose a metric to compute the directional similarity between the source eigenvectors and the target subspace. We show an equivalence between this metric and the variance of target data when projected onto source eigenvectors. Using this equivalence, we select a subset of source principal directions that capture the variance in target data. To define our model, we augment the selected source subspace with the target subspace learned from a handful of target examples. In experiments done on six publicly available datasets, we show that our approach outperforms the state of the art in terms of the RMS fitting error as well as the percentage of test examples for which AAM fitting converges to the ground truth.

Michael H. Li, Tiago A. Mestre, Susan H. Fox et al.

Objective: To apply deep learning pose estimation algorithms for vision-based assessment of parkinsonism and levodopa-induced dyskinesia (LID). Methods: Nine participants with Parkinson's disease (PD) and LID completed a levodopa infusion protocol, where symptoms were assessed at regular intervals using the Unified Dyskinesia Rating Scale (UDysRS) and Unified Parkinson's Disease Rating Scale (UPDRS). A state-of-the-art deep learning pose estimation method was used to extract movement trajectories from videos of PD assessments. Features of the movement trajectories were used to detect and estimate the severity of parkinsonism and LID using random forest. Communication and drinking tasks were used to assess LID, while leg agility and toe tapping tasks were used to assess parkinsonism. Feature sets from tasks were also combined to predict total UDysRS and UPDRS Part III scores. Results: For LID, the communication task yielded the best results for dyskinesia (severity estimation: r = 0.661, detection: AUC = 0.930). For parkinsonism, leg agility had better results for severity estimation (r = 0.618), while toe tapping was better for detection (AUC = 0.773). UDysRS and UPDRS Part III scores were predicted with r = 0.741 and 0.530, respectively. Conclusion: This paper presents the first application of deep learning for vision-based assessment of parkinsonism and LID and demonstrates promising performance for the future translation of deep learning to PD clinical practices. Significance: The proposed system provides insight into the potential of computer vision and deep learning for clinical application in PD.

Babak Taati, Pranay Lohia, Avril Mansfield et al.

Because falls are funny, YouTube and other video sharing sites contain a large repository of real-life falls. We propose extracting gait and balance information from these videos to help us better understand some of the factors that contribute to falls. Proof-of-concept is explored in a single video containing multiple (n=14) falls/non-falls in the presence of an unexpected obstacle. The analysis explores: computing spatiotemporal parameters of gait in a video captured from an arbitrary viewpoint; the relationship between parameters of gait from the last few steps before the obstacle and falling vs. not falling; and the predictive capacity of a multivariate model in predicting a fall in the presence of an unexpected obstacle. Homography transformations correct the perspective projection distortion and allow for the consistent tracking of gait parameters as an individual walks in an arbitrary direction in the scene. A synthetic top view allows for computing the average stride length and a synthetic side view allows for measuring up and down motions of the head. In leave-one-out cross-validation, we were able to correctly predict whether a person would fall or not in 11 out of the 14 cases (78.6%), just by looking at the average stride length and the range of vertical head motion during the 1-4 most recent steps prior to reaching the obstacle.

Shehroz S. Khan, Babak Taati

A fall is an abnormal activity that occurs rarely, so it is hard to collect real data for falls. It is, therefore, difficult to use supervised learning methods to automatically detect falls. Another challenge in using machine learning methods to automatically detect falls is the choice of engineered features. In this paper, we propose to use an ensemble of autoencoders to extract features from different channels of wearable sensor data trained only on normal activities. We show that the traditional approach of choosing a threshold as the maximum of the reconstruction error on the training normal data is not the right way to identify unseen falls. We propose two methods for automatic tightening of reconstruction error from only the normal activities for better identification of unseen falls. We present our results on two activity recognition datasets and show the efficacy of our proposed method against traditional autoencoder models and two standard one-class classification methods.

Yani Ioannou, Babak Taati, Robin Harrap et al.

A novel multi-scale operator for unorganized 3D point clouds is introduced. The Difference of Normals (DoN) provides a computationally efficient, multi-scale approach to processing large unorganized 3D point clouds. The application of DoN in the multi-scale filtering of two different real-world outdoor urban LIDAR scene datasets is quantitatively and qualitatively demonstrated. In both datasets the DoN operator is shown to segment large 3D point clouds into scale-salient clusters, such as cars, people, and lamp posts towards applications in semi-automatic annotation, and as a pre-processing step in automatic object recognition. The application of the operator to segmentation is evaluated on a large public dataset of outdoor LIDAR scenes with ground truth annotations.