Ali Etemad

Mohsen Zand, Ali Etemad, Michael Greenspan

We present ObjectBox, a novel single-stage anchor-free and highly generalizable object detection approach. As opposed to both existing anchor-based and anchor-free detectors, which are more biased toward specific object scales in their label assignments, we use only object center locations as positive samples and treat all objects equally in different feature levels regardless of the objects' sizes or shapes. Specifically, our label assignment strategy considers the object center locations as shape- and size-agnostic anchors in an anchor-free fashion, and allows learning to occur at all scales for every object. To support this, we define new regression targets as the distances from two corners of the center cell location to the four sides of the bounding box. Moreover, to handle scale-variant objects, we propose a tailored IoU loss to deal with boxes with different sizes. As a result, our proposed object detector does not need any dataset-dependent hyperparameters to be tuned across datasets. We evaluate our method on MS-COCO 2017 and PASCAL VOC 2012 datasets, and compare our results to state-of-the-art methods. We observe that ObjectBox performs favorably in comparison to prior works. Furthermore, we perform rigorous ablation experiments to evaluate different components of our method. Our code is available at: https://github.com/MohsenZand/ObjectBox.

Sotirios Vavaroutas, Yu Yvonne Wu, Ali Etemad et al.

Data samples used for training often differ from those encountered during fine-tuning and deployment, and while ML models show promise, their performance remains limited when only small annotated datasets are available. Performance often degrades under distribution shifts caused by diverse sensors, populations, and application settings. Although pre-training helps, models frequently encounter out-of-distribution (OOD) data in real-world settings, leading to reduced robustness. Existing adaptation methods usually assume fixed distribution shifts and struggle when multiple types or severities occur. In particular, they overlook shift severity, for example treating adaptation to a large familiar dataset the same as adaptation to a small dataset with a new task, which limits generalisation. To address this, we propose ADAPTOOD, a novel framework that leverages data uncertainty to quantify distribution shift severity and guide fine-tuning for time series. This uncertainty measures how strongly samples from the target deployment distribution deviate from the pre-training distribution, providing a direct signal of OOD severity. Our framework combines this uncertainty with low-rank model updates and adaptive hyperparameter optimisation to improve adaptation. We show that ADAPTOOD achieves up to 7% higher accuracy and 12.9% higher precision than existing methods in OOD tasks, maintaining strong performance as distribution shift severity increases.

Shuvendu Roy, Ali Etemad

We propose Consistency-guided Prompt learning (CoPrompt), a new fine-tuning method for vision-language models. Our approach improves the generalization of large foundation models when fine-tuned on downstream tasks in a few-shot setting. The basic idea of CoPrompt is to enforce a consistency constraint in the prediction of the trainable and pre-trained models to prevent overfitting on the downstream task. Additionally, we introduce the following two components into our consistency constraint to further boost the performance: enforcing consistency on two perturbed inputs and combining two dominant paradigms of tuning, prompting and adapter. Enforcing consistency on perturbed input serves to further regularize the consistency constraint, thereby improving generalization. Moreover, the integration of adapters and prompts not only enhances performance on downstream tasks but also offers increased tuning flexibility in both input and output spaces. This facilitates more effective adaptation to downstream tasks in a few-shot learning setting. Experiments show that CoPrompt outperforms existing methods on a range of evaluation suites, including base-to-novel generalization, domain generalization, and cross-dataset evaluation. On generalization, CoPrompt improves the state-of-the-art on zero-shot tasks and the overall harmonic mean over 11 datasets. Detailed ablation studies show the effectiveness of each of the components in CoPrompt. We make our code available at https://github.com/ShuvenduRoy/CoPrompt.

Mohsen Zand, Ali Etemad, Michael Greenspan

A new method is proposed for human motion prediction by learning temporal and spatial dependencies. Recently, multiscale graphs have been developed to model the human body at higher abstraction levels, resulting in more stable motion prediction. Current methods however predetermine scale levels and combine spatially proximal joints to generate coarser scales based on human priors, even though movement patterns in different motion sequences vary and do not fully comply with a fixed graph of spatially connected joints. Another problem with graph convolutional methods is mode collapse, in which predicted poses converge around a mean pose with no discernible movements, particularly in long-term predictions. To tackle these issues, we propose ResChunk, an end-to-end network which explores dynamically correlated body components based on the pairwise relationships between all joints in individual sequences. ResChunk is trained to learn the residuals between target sequence chunks in an autoregressive manner to enforce the temporal connectivities between consecutive chunks. It is hence a sequence-to-sequence prediction network which considers dynamic spatio-temporal features of sequences at multiple levels. Our experiments on two challenging benchmark datasets, CMU Mocap and Human3.6M, demonstrate that our proposed method is able to effectively model the sequence information for motion prediction and outperform other techniques to set a new state-of-the-art. Our code is available at https://github.com/MohsenZand/ResChunk.

Nishq Poorav Desai, Ali Etemad, Michael Greenspan

Self-driving research often underrepresents cyclist collisions and safety. To address this, we present CycleCrash, a novel dataset consisting of 3,000 dashcam videos with 436,347 frames that capture cyclists in a range of critical situations, from collisions to safe interactions. This dataset enables 9 different cyclist collision prediction and classification tasks focusing on potentially hazardous conditions for cyclists and is annotated with collision-related, cyclist-related, and scene-related labels. Next, we propose VidNeXt, a novel method that leverages a ConvNeXt spatial encoder and a non-stationary transformer to capture the temporal dynamics of videos for the tasks defined in our dataset. To demonstrate the effectiveness of our method and create additional baselines on CycleCrash, we apply and compare 7 models along with a detailed ablation. We release the dataset and code at https://github.com/DeSinister/CycleCrash/ .

Barys Liskavets, Maxim Ushakov, Shuvendu Roy et al.

Large language models (LLMs) have triggered a new stream of research focusing on compressing the context length to reduce the computational cost while ensuring the retention of helpful information for LLMs to answer the given question. Token-based removal methods are one of the most prominent approaches in this direction, but risk losing the semantics of the context caused by intermediate token removal, especially under high compression ratios, while also facing challenges in computational efficiency. In this work, we propose context-aware prompt compression (CPC), a sentence-level prompt compression technique where its key innovation is a novel context-aware sentence encoder that provides a relevance score for each sentence for a given question. To train this encoder, we generate a new dataset consisting of questions, positives, and negative pairs where positives are sentences relevant to the question, while negatives are irrelevant context sentences. We train the encoder in a contrastive setup to learn context-aware sentence representations. Our method considerably outperforms prior works on prompt compression on benchmark datasets and is up to 10.93x faster at inference compared to the best token-level compression method. We also find better improvement for shorter length constraints in most benchmarks, showing the effectiveness of our proposed solution in the compression of relevant information in a shorter context. Finally, we release the code and the dataset for quick reproducibility and further development: https://github.com/Workday/cpc.

Shuvendu Roy, Ali Etemad

Training deep neural networks for image recognition often requires large-scale human annotated data. To reduce the reliance of deep neural solutions on labeled data, state-of-the-art semi-supervised methods have been proposed in the literature. Nonetheless, the use of such semi-supervised methods has been quite rare in the field of facial expression recognition (FER). In this paper, we present a comprehensive study on recently proposed state-of-the-art semi-supervised learning methods in the context of FER. We conduct comparative study on eight semi-supervised learning methods, namely Pi-Model, Pseudo-label, Mean-Teacher, VAT, MixMatch, ReMixMatch, UDA, and FixMatch, on three FER datasets (FER13, RAF-DB, and AffectNet), when various amounts of labeled samples are used. We also compare the performance of these methods against fully-supervised training. Our study shows that when training existing semi-supervised methods on as little as 250 labeled samples per class can yield comparable performances to that of fully-supervised methods trained on the full labeled datasets. To facilitate further research in this area, we make our code publicly available at: https://github.com/ShuvenduRoy/SSL_FER

Shuvendu Roy, Ali Etemad

Deep learning has played a significant role in the success of facial expression recognition (FER), thanks to large models and vast amounts of labelled data. However, obtaining labelled data requires a tremendous amount of human effort, time, and financial resources. Even though some prior works have focused on reducing the need for large amounts of labelled data using different unsupervised methods, another promising approach called active learning is barely explored in the context of FER. This approach involves selecting and labelling the most representative samples from an unlabelled set to make the best use of a limited 'labelling budget'. In this paper, we implement and study 8 recent active learning methods on three public FER datasets, FER13, RAF-DB, and KDEF. Our findings show that existing active learning methods do not perform well in the context of FER, likely suffering from a phenomenon called 'Cold Start', which occurs when the initial set of labelled samples is not well representative of the entire dataset. To address this issue, we propose contrastive self-supervised pre-training, which first learns the underlying representations based on the entire unlabelled dataset. We then follow this with the active learning methods and observe that our 2-step approach shows up to 9.2% improvement over random sampling and up to 6.7% improvement over the best existing active learning baseline without the pre-training. We will make the code for this study public upon publication at: github.com/ShuvenduRoy/ActiveFER.

Shuvendu Roy, Ali Etemad

Deep learning-based methods have been the key driving force behind much of the recent success of facial expression recognition (FER) systems. However, the need for large amounts of labelled data remains a challenge. Semi-supervised learning offers a way to overcome this limitation, allowing models to learn from a small amount of labelled data along with a large unlabelled dataset. While semi-supervised learning has shown promise in FER, most current methods from general computer vision literature have not been explored in the context of FER. In this work, we present a comprehensive study on 11 of the most recent semi-supervised methods, in the context of FER, namely Pi-model, Pseudo-label, Mean Teacher, VAT, UDA, MixMatch, ReMixMatch, FlexMatch, CoMatch, and CCSSL. Our investigation covers semi-supervised learning from in-distribution, out-of-distribution, unconstrained, and very small unlabelled data. Our evaluation includes five FER datasets plus one large face dataset for unconstrained learning. Our results demonstrate that FixMatch consistently achieves better performance on in-distribution unlabelled data, while ReMixMatch stands out among all methods for out-of-distribution, unconstrained, and scarce unlabelled data scenarios. Another significant observation is that with an equal number of labelled samples, semi-supervised learning delivers a considerable improvement over supervised learning, regardless of whether the unlabelled data is in-distribution, out-of-distribution, or unconstrained. We also conduct sensitivity analyses on critical hyper-parameters for the two best methods of each setting. To facilitate reproducibility and further development, we make our code publicly available at: github.com/ShuvenduRoy/SSL_FER_OOD.

Prithila Angkan, Behnam Behinaein, Zunayed Mahmud et al.

Through this paper, we introduce a novel driver cognitive load assessment dataset, CL-Drive, which contains Electroencephalogram (EEG) signals along with other physiological signals such as Electrocardiography (ECG) and Electrodermal Activity (EDA) as well as eye tracking data. The data was collected from 21 subjects while driving in an immersive vehicle simulator, in various driving conditions, to induce different levels of cognitive load in the subjects. The tasks consisted of 9 complexity levels for 3 minutes each. Each driver reported their subjective cognitive load every 10 seconds throughout the experiment. The dataset contains the subjective cognitive load recorded as ground truth. In this paper, we also provide benchmark classification results for different machine learning and deep learning models for both binary and ternary label distributions. We followed 2 evaluation criteria namely 10-fold and leave-one-subject-out (LOSO). We have trained our models on both hand-crafted features as well as on raw data.

Sahar Soltanieh, Javad Hashemi, Ali Etemad

This paper presents a systematic investigation into the effectiveness of Self-Supervised Learning (SSL) methods for Electrocardiogram (ECG) arrhythmia detection. We begin by conducting a novel analysis of the data distributions on three popular ECG-based arrhythmia datasets: PTB-XL, Chapman, and Ribeiro. To the best of our knowledge, our study is the first to quantitatively explore and characterize these distributions in the area. We then perform a comprehensive set of experiments using different augmentations and parameters to evaluate the effectiveness of various SSL methods, namely SimCRL, BYOL, and SwAV, for ECG representation learning, where we observe the best performance achieved by SwAV. Furthermore, our analysis shows that SSL methods achieve highly competitive results to those achieved by supervised state-of-the-art methods. To further assess the performance of these methods on both In-Distribution (ID) and Out-of-Distribution (OOD) ECG data, we conduct cross-dataset training and testing experiments. Our comprehensive experiments show almost identical results when comparing ID and OOD schemes, indicating that SSL techniques can learn highly effective representations that generalize well across different OOD datasets. This finding can have major implications for ECG-based arrhythmia detection. Lastly, to further analyze our results, we perform detailed per-disease studies on the performance of the SSL methods on the three datasets.

Anubhav Bhatti, Behnam Behinaein, Paul Hungler et al.

We propose cross-modal attentive connections, a new dynamic and effective technique for multimodal representation learning from wearable data. Our solution can be integrated into any stage of the pipeline, i.e., after any convolutional layer or block, to create intermediate connections between individual streams responsible for processing each modality. Additionally, our method benefits from two properties. First, it can share information uni-directionally (from one modality to the other) or bi-directionally. Second, it can be integrated into multiple stages at the same time to further allow network gradients to be exchanged in several touch-points. We perform extensive experiments on three public multimodal wearable datasets, WESAD, SWELL-KW, and CASE, and demonstrate that our method can effectively regulate and share information between different modalities to learn better representations. Our experiments further demonstrate that once integrated into simple CNN-based multimodal solutions (2, 3, or 4 modalities), our method can result in superior or competitive performance to state-of-the-art and outperform a variety of baseline uni-modal and classical multimodal methods.

Dustin Pulver, Prithila Angkan, Paul Hungler et al.

Cognitive load, the amount of mental effort required for task completion, plays an important role in performance and decision-making outcomes, making its classification and analysis essential in various sensitive domains. In this paper, we present a new solution for the classification of cognitive load using electroencephalogram (EEG). Our model uses a transformer architecture employing transfer learning between emotions and cognitive load. We pre-train our model using self-supervised masked autoencoding on emotion-related EEG datasets and use transfer learning with both frozen weights and fine-tuning to perform downstream cognitive load classification. To evaluate our method, we carry out a series of experiments utilizing two publicly available EEG-based emotion datasets, namely SEED and SEED-IV, for pre-training, while we use the CL-Drive dataset for downstream cognitive load classification. The results of our experiments show that our proposed approach achieves strong results and outperforms conventional single-stage fully supervised learning. Moreover, we perform detailed ablation and sensitivity studies to evaluate the impact of different aspects of our proposed solution. This research contributes to the growing body of literature in affective computing with a focus on cognitive load, and opens up new avenues for future research in the field of cross-domain transfer learning using self-supervised pre-training.

Pritam Sarkar, Ali Etemad

We present XKD, a novel self-supervised framework to learn meaningful representations from unlabelled videos. XKD is trained with two pseudo objectives. First, masked data reconstruction is performed to learn modality-specific representations from audio and visual streams. Next, self-supervised cross-modal knowledge distillation is performed between the two modalities through a teacher-student setup to learn complementary information. We introduce a novel domain alignment strategy to tackle domain discrepancy between audio and visual modalities enabling effective cross-modal knowledge distillation. Additionally, to develop a general-purpose network capable of handling both audio and visual streams, modality-agnostic variants of XKD are introduced, which use the same pretrained backbone for different audio and visual tasks. Our proposed cross-modal knowledge distillation improves video action classification by $8\%$ to $14\%$ on UCF101, HMDB51, and Kinetics400. Additionally, XKD improves multimodal action classification by $5.5\%$ on Kinetics-Sound. XKD shows state-of-the-art performance in sound classification on ESC50, achieving top-1 accuracy of $96.5\%$.

Setareh Rahimi Taghanaki, Michael Rainbow, Ali Etemad

In this paper, we propose a self-supervised learning solution for human activity recognition with smartphone accelerometer data. We aim to develop a model that learns strong representations from accelerometer signals, in order to perform robust human activity classification, while reducing the model's reliance on class labels. Specifically, we intend to enable cross-dataset transfer learning such that our network pre-trained on a particular dataset can perform effective activity classification on other datasets (successive to a small amount of fine-tuning). To tackle this problem, we design our solution with the intention of learning as much information from the accelerometer signals as possible. As a result, we design two separate pipelines, one that learns the data in time-frequency domain, and the other in time-domain alone. In order to address the issues mentioned above in regards to cross-dataset transfer learning, we use self-supervised contrastive learning to train each of these streams. Next, each stream is fine-tuned for final classification, and eventually the two are fused to provide the final results. We evaluate the performance of the proposed solution on three datasets, namely MotionSense, HAPT, and HHAR, and demonstrate that our solution outperforms prior works in this field. We further evaluate the performance of the method in learning generalized features, by using MobiAct dataset for pre-training and the remaining three datasets for the downstream classification task, and show that the proposed solution achieves better performance in comparison with other self-supervised methods in cross-dataset transfer learning.

Mojtaba Kolahdouzi, Alireza Sepas-Moghaddam, Ali Etemad

Prior work has shown that the order in which different components of the face are learned using a sequential learner can play an important role in the performance of facial expression recognition systems. We propose FaceTopoNet, an end-to-end deep model for facial expression recognition, which is capable of learning an effective tree topology of the face. Our model then traverses the learned tree to generate a sequence, which is then used to form an embedding to feed a sequential learner. The devised model adopts one stream for learning structure and one stream for learning texture. The structure stream focuses on the positions of the facial landmarks, while the main focus of the texture stream is on the patches around the landmarks to learn textural information. We then fuse the outputs of the two streams by utilizing an effective attention-based fusion strategy. We perform extensive experiments on four large-scale in-the-wild facial expression datasets - namely AffectNet, FER2013, ExpW, and RAF-DB - and one lab-controlled dataset (CK+) to evaluate our approach. FaceTopoNet achieves state-of-the-art performance on three of the five datasets and obtains competitive results on the other two datasets. We also perform rigorous ablation and sensitivity experiments to evaluate the impact of different components and parameters in our model. Lastly, we perform robustness experiments and demonstrate that FaceTopoNet is more robust against occlusions in comparison to other leading methods in the area.

Debaditya Shome, Pritam Sarkar, Ali Etemad

The high prevalence of cardiovascular diseases (CVDs) calls for accessible and cost-effective continuous cardiac monitoring tools. Despite Electrocardiography (ECG) being the gold standard, continuous monitoring remains a challenge, leading to the exploration of Photoplethysmography (PPG), a promising but more basic alternative available in consumer wearables. This notion has recently spurred interest in translating PPG to ECG signals. In this work, we introduce Region-Disentangled Diffusion Model (RDDM), a novel diffusion model designed to capture the complex temporal dynamics of ECG. Traditional Diffusion models like Denoising Diffusion Probabilistic Models (DDPM) face challenges in capturing such nuances due to the indiscriminate noise addition process across the entire signal. Our proposed RDDM overcomes such limitations by incorporating a novel forward process that selectively adds noise to specific regions of interest (ROI) such as QRS complex in ECG signals, and a reverse process that disentangles the denoising of ROI and non-ROI regions. Quantitative experiments demonstrate that RDDM can generate high-fidelity ECG from PPG in as few as 10 diffusion steps, making it highly effective and computationally efficient. Additionally, to rigorously validate the usefulness of the generated ECG signals, we introduce CardioBench, a comprehensive evaluation benchmark for a variety of cardiac-related tasks including heart rate and blood pressure estimation, stress classification, and the detection of atrial fibrillation and diabetes. Our thorough experiments show that RDDM achieves state-of-the-art performance on CardioBench. To the best of our knowledge, RDDM is the first diffusion model for cross-modal signal-to-signal translation in the bio-signal domain.

Haleh Damirchi, Michael Greenspan, Ali Etemad

We propose a novel solution for predicting future trajectories of pedestrians. Our method uses a multimodal encoder-decoder transformer architecture, which takes as input both pedestrian locations and ego-vehicle speeds. Notably, our decoder predicts the entire future trajectory in a single-pass and does not perform one-step-ahead prediction, which makes the method effective for embedded edge deployment. We perform detailed experiments and evaluate our method on two popular datasets, PIE and JAAD. Quantitative results demonstrate the superiority of our proposed model over the current state-of-the-art, which consistently achieves the lowest error for 3 time horizons of 0.5, 1.0 and 1.5 seconds. Moreover, the proposed method is significantly faster than the state-of-the-art for the two datasets of PIE and JAAD. Lastly, ablation experiments demonstrate the impact of the key multimodal configuration of our method.

Pritam Sarkar, Ahmad Beirami, Ali Etemad

Video self-supervised learning (VSSL) has made significant progress in recent years. However, the exact behavior and dynamics of these models under different forms of distribution shift are not yet known. In this paper, we comprehensively study the behavior of six popular self-supervised methods (v-SimCLR, v-MoCo, v-BYOL, v-SimSiam, v-DINO, v-MAE) in response to various forms of natural distribution shift, i.e., (i) context shift, (ii) viewpoint shift, (iii) actor shift, (iv) source shift, (v) generalizability to unknown classes (zero-shot), and (vi) open-set recognition. To perform this extensive study, we carefully craft a test bed consisting of 17 in-distribution and out-of-distribution benchmark pairs using available public datasets and a series of evaluation protocols to stress-test the different methods under the intended shifts. Our study uncovers a series of intriguing findings and interesting behaviors of VSSL methods. For instance, we observe that while video models generally struggle with context shifts, v-MAE and supervised learning exhibit more robustness. Moreover, our study shows that v-MAE is a strong temporal learner, whereas contrastive methods, v-SimCLR and v-MoCo, exhibit strong performances against viewpoint shifts. When studying the notion of open-set recognition, we notice a trade-off between closed-set and open-set recognition performance if the pretrained VSSL encoders are used without finetuning. We hope that our work will contribute to the development of robust video representation learning frameworks for various real-world scenarios. The project page and code are available at: https://pritamqu.github.io/OOD-VSSL.

Zunayed Mahmud, Paul Hungler, Ali Etemad

We propose a novel neural pipeline, MSGazeNet, that learns gaze representations by taking advantage of the eye anatomy information through a multistream framework. Our proposed solution comprises two components, first a network for isolating anatomical eye regions, and a second network for multistream gaze estimation. The eye region isolation is performed with a U-Net style network which we train using a synthetic dataset that contains eye region masks for the visible eyeball and the iris region. The synthetic dataset used in this stage is procured using the UnityEyes simulator, and consists of 80,000 eye images. Successive to training, the eye region isolation network is then transferred to the real domain for generating masks for the real-world eye images. In order to successfully make the transfer, we exploit domain randomization in the training process, which allows for the synthetic images to benefit from a larger variance with the help of augmentations that resemble artifacts. The generated eye region masks along with the raw eye images are then used together as a multistream input to our gaze estimation network, which consists of wide residual blocks. The output embeddings from these encoders are fused in the channel dimension before feeding into the gaze regression layers. We evaluate our framework on three gaze estimation datasets and achieve strong performances. Our method surpasses the state-of-the-art by 7.57% and 1.85% on two datasets, and obtains competitive results on the other. We also study the robustness of our method with respect to the noise in the data and demonstrate that our model is less sensitive to noisy data. Lastly, we perform a variety of experiments including ablation studies to evaluate the contribution of different components and design choices in our solution.

Nishq Poorav Desai, Ali Etemad, Michael Greenspan

Time-to-Collision (TTC) forecasting is a critical task in collision prevention, requiring precise temporal prediction and comprehending both local and global patterns encapsulated in a video, both spatially and temporally. To address the multi-scale nature of video, we introduce a novel spatiotemporal hierarchical transformer-based architecture called CollideNet, specifically catered for effective TTC forecasting. In the spatial stream, CollideNet aggregates information for each video frame simultaneously at multiple resolutions. In the temporal stream, along with multi-scale feature encoding, CollideNet also disentangles the non-stationarity, trend, and seasonality components. Our method achieves state-of-the-art performance in comparison to prior works on three commonly used public datasets, setting a new state-of-the-art by a considerable margin. We conduct cross-dataset evaluations to analyze the generalization capabilities of our method, and visualize the effects of disentanglement of the trend and seasonality components of the video data. We release our code at https://github.com/DeSinister/CollideNet/.

Sahar Soltanieh, Ali Etemad, Javad Hashemi

This paper systematically investigates the effectiveness of various augmentations for contrastive self-supervised learning of electrocardiogram (ECG) signals and identifies the best parameters. The baseline of our proposed self-supervised framework consists of two main parts: the contrastive learning and the downstream task. In the first stage, we train an encoder using a number of augmentations to extract generalizable ECG signal representations. We then freeze the encoder and finetune a few linear layers with different amounts of labelled data for downstream arrhythmia detection. We then experiment with various augmentations techniques and explore a range of parameters. Our experiments are done on PTB-XL, a large and publicly available 12-lead ECG dataset. The results show that applying augmentations in a specific range of complexities works better for self-supervised contrastive learning. For instance, when adding Gaussian noise, a sigma in the range of 0.1 to 0.2 achieves better results, while poor training occurs when the added noise is too small or too large (outside of the specified range). A similar trend is observed with other augmentations, demonstrating the importance of selecting the optimum level of difficulty for the added augmentations, as augmentations that are too simple will not result in effective training, while augmentations that are too difficult will also prevent the model from effective learning of generalized representations. Our work can influence future research on self-supervised contrastive learning on bio-signals and aid in selecting optimum parameters for different augmentations.

Guangyi Zhang, Ali Etemad

Fully supervised learning has recently achieved promising performance in various electroencephalography (EEG) learning tasks by training on large datasets with ground truth labels. However, labeling EEG data for affective experiments is challenging, as it can be difficult for participants to accurately distinguish between similar emotions, resulting in ambiguous labeling (reporting multiple emotions for one EEG instance). This notion could cause model performance degradation, as the ground truth is hidden within multiple candidate labels. To address this issue, Partial Label Learning (PLL) has been proposed to identify the ground truth from candidate labels during the training phase, and has shown good performance in the computer vision domain. However, PLL methods have not yet been adopted for EEG representation learning or implemented for emotion recognition tasks. In this paper, we adapt and re-implement six state-of-the-art PLL approaches for emotion recognition from EEG on two large emotion datasets (SEED-IV and SEED-V). These datasets contain four and five categories of emotions, respectively. We evaluate the performance of all methods in classical, circumplex-based and real-world experiments. The results show that PLL methods can achieve strong results in affective computing from EEG and achieve comparable performance to fully supervised learning. We also investigate the effect of label disambiguation, a key step in many PLL methods. The results show that in most cases, label disambiguation would benefit the model when the candidate labels are generated based on their similarities to the ground truth rather than obeying a uniform distribution. This finding suggests the potential of using label disambiguation-based PLL methods for circumplex-based and real-world affective tasks.

Pritam Sarkar, Aaron Posen, Ali Etemad

We introduce AVCAffe, the first Audio-Visual dataset consisting of Cognitive load and Affect attributes. We record AVCAffe by simulating remote work scenarios over a video-conferencing platform, where subjects collaborate to complete a number of cognitively engaging tasks. AVCAffe is the largest originally collected (not collected from the Internet) affective dataset in English language. We recruit 106 participants from 18 different countries of origin, spanning an age range of 18 to 57 years old, with a balanced male-female ratio. AVCAffe comprises a total of 108 hours of video, equivalent to more than 58,000 clips along with task-based self-reported ground truth labels for arousal, valence, and cognitive load attributes such as mental demand, temporal demand, effort, and a few others. We believe AVCAffe would be a challenging benchmark for the deep learning research community given the inherent difficulty of classifying affect and cognitive load in particular. Moreover, our dataset fills an existing timely gap by facilitating the creation of learning systems for better self-management of remote work meetings, and further study of hypotheses regarding the impact of remote work on cognitive load and affective states.

Divij Gupta, Ali Etemad

Remote Photoplethysmography (rPPG) is the process of estimating PPG from facial videos. While this approach benefits from contactless interaction, it is reliant on videos of faces, which often constitutes an important privacy concern. Recent research has revealed that deep learning techniques are vulnerable to attacks, which can result in significant data breaches making deep rPPG estimation even more sensitive. To address this issue, we propose a data perturbation method that involves extraction of certain areas of the face with less identity-related information, followed by pixel shuffling and blurring. Our experiments on two rPPG datasets (PURE and UBFC) show that our approach reduces the accuracy of facial recognition algorithms by over 60%, with minimal impact on rPPG extraction. We also test our method on three facial recognition datasets (LFW, CALFW, and AgeDB), where our approach reduced performance by nearly 50%. Our findings demonstrate the potential of our approach as an effective privacy-preserving solution for rPPG estimation.

Mahdiyar Molahasani, Ali Etemad, Michael Greenspan

A continual learning solution is proposed to address the out-of-distribution generalization problem for pedestrian detection. While recent pedestrian detection models have achieved impressive performance on various datasets, they remain sensitive to shifts in the distribution of the inference data. Our method adopts and modifies Elastic Weight Consolidation to a backbone object detection network, in order to penalize the changes in the model weights based on their importance towards the initially learned task. We show that when trained with one dataset and fine-tuned on another, our solution learns the new distribution and maintains its performance on the previous one, avoiding catastrophic forgetting. We use two popular datasets, CrowdHuman and CityPersons for our cross-dataset experiments, and show considerable improvements over standard fine-tuning, with a 9% and 18% miss rate percent reduction improvement in the CrowdHuman and CityPersons datasets, respectively.

Leyla Khaleghi, Joshua Marshall, Ali Etemad

3D hand pose estimation (HPE) is the process of locating the joints of the hand in 3D from any visual input. HPE has recently received an increased amount of attention due to its key role in a variety of human-computer interaction applications. Recent HPE methods have demonstrated the advantages of employing videos or multi-view images, allowing for more robust HPE systems. Accordingly, in this study, we propose a new method to perform Sequential learning with Transformer for Hand Pose (SeTHPose) estimation. Our SeTHPose pipeline begins by extracting visual embeddings from individual hand images. We then use a transformer encoder to learn the sequential context along time or viewing angles and generate accurate 2D hand joint locations. Then, a graph convolutional neural network with a U-Net configuration is used to convert the 2D hand joint locations to 3D poses. Our experiments show that SeTHPose performs well on both hand sequence varieties, temporal and angular. Also, SeTHPose outperforms other methods in the field to achieve new state-of-the-art results on two public available sequential datasets, STB and MuViHand.

Debaditya Shome, Ali Etemad

We propose EmoDistill, a novel speech emotion recognition (SER) framework that leverages cross-modal knowledge distillation during training to learn strong linguistic and prosodic representations of emotion from speech. During inference, our method only uses a stream of speech signals to perform unimodal SER thus reducing computation overhead and avoiding run-time transcription and prosodic feature extraction errors. During training, our method distills information at both embedding and logit levels from a pair of pre-trained Prosodic and Linguistic teachers that are fine-tuned for SER. Experiments on the IEMOCAP benchmark demonstrate that our method outperforms other unimodal and multimodal techniques by a considerable margin, and achieves state-of-the-art performance of 77.49% unweighted accuracy and 78.91% weighted accuracy. Detailed ablation studies demonstrate the impact of each component of our method.

Divij Gupta, Ali Etemad

Recent advances in deep learning have made it increasingly feasible to estimate heart rate remotely in smart environments by analyzing videos. However, a notable limitation of deep learning methods is their heavy reliance on extensive sets of labeled data for effective training. To address this issue, self-supervised learning has emerged as a promising avenue. Building on this, we introduce a solution that utilizes self-supervised contrastive learning for the estimation of remote photoplethysmography (PPG) and heart rate monitoring, thereby reducing the dependence on labeled data and enhancing performance. We propose the use of 3 spatial and 3 temporal augmentations for training an encoder through a contrastive framework, followed by utilizing the late-intermediate embeddings of the encoder for remote PPG and heart rate estimation. Our experiments on two publicly available datasets showcase the improvement of our proposed approach over several related works as well as supervised learning baselines, as our results approach the state-of-the-art. We also perform thorough experiments to showcase the effects of using different design choices such as the video representation learning method, the augmentations used in the pre-training stage, and others. We also demonstrate the robustness of our proposed method over the supervised learning approaches on reduced amounts of labeled data.

Mahdiyar Molahasani, Michael Greenspan, Ali Etemad

Real-world datasets often exhibit long-tailed distributions, where a few dominant "Head" classes have abundant samples while most "Tail" classes are severely underrepresented, leading to biased learning and poor generalization for the Tail. We present a theoretical framework that reveals a previously undescribed connection between Long-Tailed Recognition (LTR) and Continual Learning (CL), the process of learning sequential tasks without forgetting prior knowledge. Our analysis demonstrates that, for models trained on imbalanced datasets, the weights converge to a bounded neighborhood of those trained exclusively on the Head, with the bound scaling as the inverse square root of the imbalance factor. Leveraging this insight, we introduce Continual Learning for Long-Tailed Recognition (CLTR), a principled approach that employs standard off-the-shelf CL methods to address LTR problems by sequentially learning Head and Tail classes without forgetting the Head. Our theoretical analysis further suggests that CLTR mitigates gradient saturation and improves Tail learning while maintaining strong Head performance. Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate our theoretical predictions, achieving strong results across various LTR benchmarks. Our work bridges the gap between LTR and CL, providing a principled way to tackle imbalanced data challenges with standard existing CL strategies.

Mojtaba Kolahdouzi, Ali Etemad

We present a novel approach to mitigate bias in facial expression recognition (FER) models. Our method aims to reduce sensitive attribute information such as gender, age, or race, in the embeddings produced by FER models. We employ a kernel mean shrinkage estimator to estimate the kernel mean of the distributions of the embeddings associated with different sensitive attribute groups, such as young and old, in the Hilbert space. Using this estimation, we calculate the maximum mean discrepancy (MMD) distance between the distributions and incorporate it in the classifier loss along with an adversarial loss, which is then minimized through the learning process to improve the distribution alignment. Our method makes sensitive attributes less recognizable for the model, which in turn promotes fairness. Additionally, for the first time, we analyze the notion of attractiveness as an important sensitive attribute in FER models and demonstrate that FER models can indeed exhibit biases towards more attractive faces. To prove the efficacy of our model in reducing bias regarding different sensitive attributes (including the newly proposed attractiveness attribute), we perform several experiments on two widely used datasets, CelebA and RAF-DB. The results in terms of both accuracy and fairness measures outperform the state-of-the-art in most cases, demonstrating the effectiveness of the proposed method.

Vandad Davoodnia, Saeed Ghorbani, Ali Etemad

In-bed pose estimation has shown value in fields such as hospital patient monitoring, sleep studies, and smart homes. In this paper, we explore different strategies for detecting body pose from highly ambiguous pressure data, with the aid of pre-existing pose estimators. We examine the performance of pre-trained pose estimators by using them either directly or by re-training them on two pressure datasets. We also explore other strategies utilizing a learnable pre-processing domain adaptation step, which transforms the vague pressure maps to a representation closer to the expected input space of common purpose pose estimation modules. Accordingly, we used a fully convolutional network with multiple scales to provide the pose-specific characteristics of the pressure maps to the pre-trained pose estimation module. Our complete analysis of different approaches shows that the combination of learnable pre-processing module along with re-training pre-existing image-based pose estimators on the pressure data is able to overcome issues such as highly vague pressure points to achieve very high pose estimation accuracy.

Vandad Davoodnia, Ali Etemad

Despite the impressive performance of vision-based pose estimators, they generally fail to perform well under adverse vision conditions and often don't satisfy the privacy demands of customers. As a result, researchers have begun to study tactile sensing systems as an alternative. However, these systems suffer from noisy and ambiguous recordings. To tackle this problem, we propose a novel solution for pose estimation from ambiguous pressure data. Our method comprises a spatio-temporal vision transformer with an encoder-decoder architecture. Detailed experiments on two popular public datasets reveal that our model outperforms existing solutions in the area. Moreover, we observe that increasing the number of temporal crops in the early stages of the network positively impacts the performance while pre-training the network in a self-supervised setting using a masked auto-encoder approach also further improves the results.

Sayantan Das, Mojtaba Kolahdouzi, Levent Özparlak et al.

We present a novel approach for the detection of deepfake videos using a pair of vision transformers pre-trained by a self-supervised masked autoencoding setup. Our method consists of two distinct components, one of which focuses on learning spatial information from individual RGB frames of the video, while the other learns temporal consistency information from optical flow fields generated from consecutive frames. Unlike most approaches where pre-training is performed on a generic large corpus of images, we show that by pre-training on smaller face-related datasets, namely Celeb-A (for the spatial learning component) and YouTube Faces (for the temporal learning component), strong results can be obtained. We perform various experiments to evaluate the performance of our method on commonly used datasets namely FaceForensics++ (Low Quality and High Quality, along with a new highly compressed version named Very Low Quality) and Celeb-DFv2 datasets. Our experiments show that our method sets a new state-of-the-art on FaceForensics++ (LQ, HQ, and VLQ), and obtains competitive results on Celeb-DFv2. Moreover, our method outperforms other methods in the area in a cross-dataset setup where we fine-tune our model on FaceForensics++ and test on CelebDFv2, pointing to its strong cross-dataset generalization ability.

Debaditya Shome, Nasim Montazeri Ghahjaverestan, Ali Etemad

Sleep is known to be a key factor in emotional regulation and overall mental health. In this study, we explore the integration of sleep measures from the previous night into wearable-based mood recognition. To this end, we propose NapTune, a novel prompt-tuning framework that utilizes sleep-related measures as additional inputs to a frozen pre-trained wearable time-series encoder by adding and training lightweight prompt parameters to each Transformer layer. Through rigorous empirical evaluation, we demonstrate that the inclusion of sleep data using NapTune not only improves mood recognition performance across different wearable time-series namely ECG, PPG, and EDA, but also makes it more sample-efficient. Our method demonstrates significant improvements over the best baselines and unimodal variants. Furthermore, we analyze the impact of adding sleep-related measures on recognizing different moods as well as the influence of individual sleep-related measures.

Amirhossein Hajavi, Ali Etemad

Deep audio representation learning using multi-modal audio-visual data often leads to a better performance compared to uni-modal approaches. However, in real-world scenarios both modalities are not always available at the time of inference, leading to performance degradation by models trained for multi-modal inference. In this work, we propose a novel approach for deep audio representation learning using audio-visual data when the video modality is absent at inference. For this purpose, we adopt teacher-student knowledge distillation under the framework of learning using privileged information (LUPI). While the previous methods proposed for LUPI use soft-labels generated by the teacher, in our proposed method we use embeddings learned by the teacher to train the student network. We integrate our method in two different settings: sequential data where the features are divided into multiple segments throughout time, and non-sequential data where the entire features are treated as one whole segment. In the non-sequential setting both the teacher and student networks are comprised of an encoder component and a task header. We use the embeddings produced by the encoder component of the teacher to train the encoder of the student, while the task header of the student is trained using ground-truth labels. In the sequential setting, the networks have an additional aggregation component that is placed between the encoder and task header. We use two sets of embeddings produced by the encoder and aggregation component of the teacher to train the student. Similar to the non-sequential setting, the task header of the student network is trained using ground-truth labels. We test our framework on two different audio-visual tasks, namely speaker recognition and speech emotion recognition and show considerable improvements over sole audio-based recognition as well as prior works that use LUPI.

Shuvendu Roy, Ali Etemad

We propose UnMixMatch, a semi-supervised learning framework which can learn effective representations from unconstrained unlabelled data in order to scale up performance. Most existing semi-supervised methods rely on the assumption that labelled and unlabelled samples are drawn from the same distribution, which limits the potential for improvement through the use of free-living unlabeled data. Consequently, the generalizability and scalability of semi-supervised learning are often hindered by this assumption. Our method aims to overcome these constraints and effectively utilize unconstrained unlabelled data in semi-supervised learning. UnMixMatch consists of three main components: a supervised learner with hard augmentations that provides strong regularization, a contrastive consistency regularizer to learn underlying representations from the unlabelled data, and a self-supervised loss to enhance the representations that are learnt from the unlabelled data. We perform extensive experiments on 4 commonly used datasets and demonstrate superior performance over existing semi-supervised methods with a performance boost of 4.79%. Extensive ablation and sensitivity studies show the effectiveness and impact of each of the proposed components of our method.

Shuvendu Roy, Ali Etemad

Although there has been much progress in the area of facial expression recognition (FER), most existing methods suffer when presented with images that have been captured from viewing angles that are non-frontal and substantially different from those used in the training process. In this paper, we propose ViewFX, a novel view-invariant FER framework based on contrastive learning, capable of accurately classifying facial expressions regardless of the input viewing angles during inference. ViewFX learns view-invariant features of expression using a proposed self-supervised contrastive loss which brings together different views of the same subject with a particular expression in the embedding space. We also introduce a supervised contrastive loss to push the learnt view-invariant features of each expression away from other expressions. Since facial expressions are often distinguished with very subtle differences in the learned feature space, we incorporate the Barlow twins loss to reduce the redundancy and correlations of the representations in the learned representations. The proposed method is a substantial extension of our previously proposed CL-MEx, which only had a self-supervised loss. We test the proposed framework on two public multi-view facial expression recognition datasets, KDEF and DDCF. The experiments demonstrate that our approach outperforms previous works in the area and sets a new state-of-the-art for both datasets while showing considerably less sensitivity to challenging angles and the number of output labels used for training. We also perform detailed sensitivity and ablation experiments to evaluate the impact of different components of our model as well as its sensitivity to different parameters.

Amirhossein Hajavi, Ali Etemad

With the ubiquity of smart devices that use speaker recognition (SR) systems as a means of authenticating individuals and personalizing their services, fairness of SR systems has becomes an important point of focus. In this paper we study the notion of fairness in recent SR systems based on 3 popular and relevant definitions, namely Statistical Parity, Equalized Odds, and Equal Opportunity. We examine 5 popular neural architectures and 5 commonly used loss functions in training SR systems, while evaluating their fairness against gender and nationality groups. Our detailed experiments shed light on this concept and demonstrate that more sophisticated encoder architectures better align with the definitions of fairness. Additionally, we find that the choice of loss functions can significantly impact the bias of SR models.

Shuvendu Roy, Ali Etemad

We present ConCur, a contrastive video representation learning method that uses curriculum learning to impose a dynamic sampling strategy in contrastive training. More specifically, ConCur starts the contrastive training with easy positive samples (temporally close and semantically similar clips), and as the training progresses, it increases the temporal span effectively sampling hard positives (temporally away and semantically dissimilar). To learn better context-aware representations, we also propose an auxiliary task of predicting the temporal distance between a positive pair of clips. We conduct extensive experiments on two popular action recognition datasets, UCF101 and HMDB51, on which our proposed method achieves state-of-the-art performance on two benchmark tasks of video action recognition and video retrieval. We explore the impact of encoder backbones and pre-training strategies by using R(2+1)D and C3D encoders and pre-training on Kinetics-400 and Kinetics-200 datasets. Moreover, a detailed ablation study shows the effectiveness of each of the components of our proposed method.

Shuvendu Roy, Ali Etemad

In semi-supervised representation learning frameworks, when the number of labelled data is very scarce, the quality and representativeness of these samples become increasingly important. Existing literature on semi-supervised learning randomly sample a limited number of data points for labelling. All these labelled samples are then used along with the unlabelled data throughout the training process. In this work, we ask two important questions in this context: (1) does it matter which samples are selected for labelling? (2) does it matter how the labelled samples are used throughout the training process along with the unlabelled data? To answer the first question, we explore a number of unsupervised methods for selecting specific subsets of data to label (without prior knowledge of their labels), with the goal of maximizing representativeness w.r.t. the unlabelled set. Then, for our second line of inquiry, we define a variety of different label injection strategies in the training process. Extensive experiments on four popular datasets, CIFAR-10, CIFAR-100, SVHN, and STL-10, show that unsupervised selection of samples that are more representative of the entire data improves performance by up to ~2% over the existing semi-supervised frameworks such as MixMatch, ReMixMatch, FixMatch and others with random sample labelling. We show that this boost could even increase to 7.5% for very few-labelled scenarios. However, our study shows that gradually injecting the labels throughout the training procedure does not impact the performance considerably versus when all the existing labels are used throughout the entire training.

Prithila Angkan, Paul Hungler, Ali Etemad

Understanding how driver mental states differ between active and autonomous driving is critical for designing safe human-vehicle interfaces. This paper presents the first EEG-based comparison of cognitive load, fatigue, valence, and arousal across the two driving modes. Using data from 31 participants performing identical tasks in both scenarios of three different complexity levels, we analyze temporal patterns, task-complexity effects, and channel-wise activation differences. Our findings show that although both modes evoke similar trends across complexity levels, the intensity of mental states and the underlying neural activation differ substantially, indicating a clear distribution shift between active and autonomous driving. Transfer-learning experiments confirm that models trained on active driving data generalize poorly to autonomous driving and vice versa. We attribute this distribution shift primarily to differences in motor engagement and attentional demands between the two driving modes, which lead to distinct spatial and temporal EEG activation patterns. Although autonomous driving results in lower overall cortical activation, participants continue to exhibit measurable fluctuations in cognitive load, fatigue, valence, and arousal associated with readiness to intervene, task-evoked emotional responses, and monotony-related passive fatigue. These results emphasize the need for scenario-specific data and models when developing next-generation driver monitoring systems for autonomous vehicles.

Prithila Angkan, Amin Jalali, Paul Hungler et al.

We present a novel graph-based learning of EEG representations with gradient alignment (GEEGA) that leverages multi-domain information to learn EEG representations for brain-computer interfaces. Our model leverages graph convolutional networks to fuse embeddings from frequency-based topographical maps and time-frequency spectrograms, capturing inter-domain relationships. GEEGA addresses the challenge of achieving high inter-class separability, which arises from the temporally dynamic and subject-sensitive nature of EEG signals by incorporating the center loss and pairwise difference loss. Additionally, GEEGA incorporates a gradient alignment strategy to resolve conflicts between gradients from different domains and the fused embeddings, ensuring that discrepancies, where gradients point in conflicting directions, are aligned toward a unified optimization direction. We validate the efficacy of our method through extensive experiments on three publicly available EEG datasets: BCI-2a, CL-Drive and CLARE. Comprehensive ablation studies further highlight the impact of various components of our model.

John Turnbull, Shivam Grover, Amin Jalali et al.

Deep learning models often struggle under natural distribution shifts, a common challenge in real-world deployments. Test-Time Adaptation (TTA) addresses this by adapting models during inference without labeled source data. We present the first evaluation of TTA methods for FER under natural domain shifts, performing cross-dataset experiments with widely used FER datasets. This moves beyond synthetic corruptions to examine real-world shifts caused by differing collection protocols, annotation standards, and demographics. Results show TTA can boost FER performance under natural shifts by up to 11.34\%. Entropy minimization methods such as TENT and SAR perform best when the target distribution is clean. In contrast, prototype adjustment methods like T3A excel under larger distributional distance scenarios. Finally, feature alignment methods such as SHOT deliver the largest gains when the target distribution is noisier than our source. Our cross-dataset analysis shows that TTA effectiveness is governed by the distributional distance and the severity of the natural shift across domains.

Mahdiyar Molahasani, Azadeh Motamedi, Michael Greenspan et al.

We introduce Projection-based Reduction of Implicit Spurious bias in vision-language Models (PRISM), a new data-free and task-agnostic solution for bias mitigation in VLMs like CLIP. VLMs often inherit and amplify biases in their training data, leading to skewed predictions. PRISM is designed to debias VLMs without relying on predefined bias categories or additional external data. It operates in two stages: first, an LLM is prompted with simple class prompts to generate scene descriptions that contain spurious correlations. Next, PRISM uses our novel contrastive-style debiasing loss to learn a projection that maps the embeddings onto a latent space that minimizes spurious correlations while preserving the alignment between image and text embeddings.Extensive experiments demonstrate that PRISM outperforms current debiasing methods on the commonly used Waterbirds and CelebA datasets We make our code public at: https://github.com/MahdiyarMM/PRISM.

Negin Baghbanzadeh, Adibvafa Fallahpour, Yasaman Parhizkar et al.

Despite the growing scale of medical Vision-Language datasets, the impact of dataset quality on model performance remains under-explored. We introduce Open-PMC, a high-quality medical dataset from PubMed Central, containing 2.2 million image-text pairs, enriched with image modality annotations, subfigures, and summarized in-text references. Notably, the in-text references provide richer medical context, extending beyond the abstract information typically found in captions. Through extensive experiments, we benchmark Open-PMC against larger datasets across retrieval and zero-shot classification tasks. Our results show that dataset quality-not just size-drives significant performance gains. We complement our benchmark with an in-depth analysis of feature representation. Our findings highlight the crucial role of data curation quality in advancing multimodal medical AI. We release Open-PMC, along with the trained models and our codebase.

Shuvendu Roy, Franklin Ogidi, Ali Etemad et al.

Multimodal representation learning has demonstrated remarkable potential in enabling models to process and integrate diverse data modalities, such as text and images, for improved understanding and performance. While the medical domain can benefit significantly from this paradigm, the scarcity of paired multimodal data and reliance on proprietary or pretrained encoders pose significant challenges. In this work, we present a shared encoder framework for multimodal representation learning tailored to the medical domain. Our approach employs a single set of encoder parameters shared across modalities, augmented with learnable modality features. Empirical results demonstrate that our shared encoder idea achieves superior performance compared to separate modality-specific encoders, demonstrating improved generalization in data-constrained settings. Notably, the performance gains are more pronounced with fewer training examples, underscoring the efficiency of our shared encoder framework for real-world medical applications with limited data. Our code and experiment setup are available at https://github.com/VectorInstitute/shared_encoder.

Mojtaba Kolahdouzi, Hatice Gunes, Ali Etemad

We study whether and how the choice of optimization algorithm can impact group fairness in deep neural networks. Through stochastic differential equation analysis of optimization dynamics in an analytically tractable setup, we demonstrate that the choice of optimization algorithm indeed influences fairness outcomes, particularly under severe imbalance. Furthermore, we show that when comparing two categories of optimizers, adaptive methods and stochastic methods, RMSProp (from the adaptive category) has a higher likelihood of converging to fairer minima than SGD (from the stochastic category). Building on this insight, we derive two new theoretical guarantees showing that, under appropriate conditions, RMSProp exhibits fairer parameter updates and improved fairness in a single optimization step compared to SGD. We then validate these findings through extensive experiments on three publicly available datasets, namely CelebA, FairFace, and MS-COCO, across different tasks as facial expression recognition, gender classification, and multi-label classification, using various backbones. Considering multiple fairness definitions including equalized odds, equal opportunity, and demographic parity, adaptive optimizers like RMSProp and Adam consistently outperform SGD in terms of group fairness, while maintaining comparable predictive accuracy. Our results highlight the role of adaptive updates as a crucial yet overlooked mechanism for promoting fair outcomes. We release the source code at: https://github.com/Mkolahdoozi/Some-Optimizers-Are-More-Equal.

Mohsen Zand, Ali Etemad, Michael Greenspan

For faster sampling and higher sample quality, we propose DiNof ($\textbf{Di}$ffusion with $\textbf{No}$rmalizing $\textbf{f}$low priors), a technique that makes use of normalizing flows and diffusion models. We use normalizing flows to parameterize the noisy data at any arbitrary step of the diffusion process and utilize it as the prior in the reverse diffusion process. More specifically, the forward noising process turns a data distribution into partially noisy data, which are subsequently transformed into a Gaussian distribution by a nonlinear process. The backward denoising procedure begins with a prior created by sampling from the Gaussian distribution and applying the invertible normalizing flow transformations deterministically. To generate the data distribution, the prior then undergoes the remaining diffusion stochastic denoising procedure. Through the reduction of the number of total diffusion steps, we are able to speed up both the forward and backward processes. More importantly, we improve the expressive power of diffusion models by employing both deterministic and stochastic mappings. Experiments on standard image generation datasets demonstrate the advantage of the proposed method over existing approaches. On the unconditional CIFAR10 dataset, for example, we achieve an FID of 2.01 and an Inception score of 9.96. Our method also demonstrates competitive performance on CelebA-HQ-256 dataset as it obtains an FID score of 7.11. Code is available at https://github.com/MohsenZand/DiNof.

Mohsen Zand, Haleh Damirchi, Andrew Farley et al.

We propose a multitask approach for crowd counting and person localization in a unified framework. As the detection and localization tasks are well-correlated and can be jointly tackled, our model benefits from a multitask solution by learning multiscale representations of encoded crowd images, and subsequently fusing them. In contrast to the relatively more popular density-based methods, our model uses point supervision to allow for crowd locations to be accurately identified. We test our model on two popular crowd counting datasets, ShanghaiTech A and B, and demonstrate that our method achieves strong results on both counting and localization tasks, with MSE measures of 110.7 and 15.0 for crowd counting and AP measures of 0.71 and 0.75 for localization, on ShanghaiTech A and B respectively. Our detailed ablation experiments show the impact of our multiscale approach as well as the effectiveness of the fusion module embedded in our network. Our code is available at: https://github.com/RCVLab-AiimLab/crowd_counting.