Recent & Upcoming Seminars

This work proposed a transformer decoder-based Large Sensor Model (LSM) pretrained on 1.6 million CGM records from patients with different diabetes types and demonstrates improvement in near-future glucose prediction.

Unicast is a timeseries forecasting foundation model. It leverages multi-modal data (visual and textual information) from a timeseries through embedding generators and aligns the output of the embedders to forecasting task using parameter efficient soft prompting.

This study introduces SELF-CARE, a context-aware sensor fusion framework for stress detection using wearable devices.

Continual learning has always been central to intelligence, enabling systems to retain and refine knowledge over time. While foundation models and LLMs can handle diverse tasks, continual learning remains crucial. It prevents knowledge from becoming stale through continual pre-training, supports adaptation and personalization via continual fine-tuning, and enables scalable, modular intelligence through continual compositionality. Ultimately, the future of AI will be shaped not by static monolithic models, but by an evolving ecosystem of continually adapting and interacting models.

In this talk, Abdullah Mamun presents Nano Banana and its superiority in image editing over other models such as GPT-5. Additionally this talk provides insight into how Nano Banana performs in medical image segmentation compared to specialist models such as U-Net.

This presentation introduced a framework that incorporates clinical safety directly into the training of deep learning models for glucose prediction in diabetic patients. Instead of optimizing solely for accuracy, the approach uses a specialized loss function (gcMSE) that penalizes clinically risky errors more heavily. The authors also proposed the PICA algorithm, which gradually shifts training toward meeting clinical standards. Tested on type-1 and type-2 diabetes datasets, the method showed improved safety and reliability, especially in critical glycemic ranges like hypoglycemia.

The presentation reviews a benchmark study evaluating deep learning models for diagnosing viral infections using single-cell RNA sequencing (scRNA-seq) data. The study compared multiple models—including PCA, SAVER, scVI, scGPT, and contrastiveVI—on a large dataset of over 200,000 immune cells from patients with COVID-19, influenza, and dengue. Key challenges addressed were batch effects and reliable classification across diverse experimental conditions. Results showed that contrastiveVI outperformed others in classification accuracy, while scGPT contributed to effective batch harmonization. Evaluation metrics included ARI, NMI, k-nearest neighbor accuracy, and S-score, with UMAP used for visual confirmation. The findings demonstrate the importance of model choice, preprocessing, and batch control in applying scRNA-seq analysis for infection diagnosis.

TimesFM is a decoder-only foundation model trained in a supervised way. It has a very high out-of-the-box zero-shot performance and can be used with any context size.

This study addresses the challenge of detecting pilot stress, which is crucial for enhancing flight safety and performance.

This survey reviews continual learning (CL) with pre-trained models (PTMs), grouping methods into prompt-based, representation-based, and model mixture-based approaches. It finds representation-based baselines often outperform more complex prompt strategies and flags fairness issues in some evaluations. Key challenges include handling domain gaps, ensuring computational efficiency, expanding to multimodal and language models, and creating benchmarks beyond PTM pretraining data.

We present a structured overview of methods, challenges, and solutions, aiming to support researchers and practitioners in developing reliable and explainable AI solutions for digital health. This paper is further enriched with detailed discussions of the contributions toward robustness and explainability in digital health, the development of trustworthy AI systems in the era of LLMs, and various evaluation metrics for measuring trust and related parameters such as validity, fidelity, and diversity.

This paper proposes DT-based MIB optimization framework to assist clinicians during routine visits with pediatric T1D patients by optimizing insulin therapy on a biweekly basis

This paper proposes a novel approach for detecting insulin pump faults (IPFs) by combining the ability of a long short term memory (LSTM) autoencoder to extract features with the strength of random forest to distinguish between anomalous and normal patterns.

In this study, we developed an explainable machine learning solution, GlucoLens, that takes sensor-driven inputs and uses advanced data processing, large language models, and trainable machine learning models to predict postprandial AUC and hyperglycemia from diet, physical activity, and recent glucose patterns.

This paper runs a 48-week long clinical trial with Type 2 diabetic patients to assess the effectiveness of a digital health platform .

This paper proposes an efficient method for full causal graph discovery using large language models (LLMs). Unlike prior pairwise approaches that require O(n^2) queries, the authors use a breadth-first search (BFS)-based framework that reduces query complexity to O(n) by querying one node at a time. The method achieves state-of-the-art performance on small and large graphs, even without access to observational data.

This study proposes Transformer-based models for motor imagery EEG classification, leveraging their ability to capture spatial-temporal dependencies and visualize attention patterns. Using 3-second EEG data from the PhysioNet dataset with cross-individual validation, the models outperformed existing methods. Positional embeddings improved accuracy, and attention visualizations revealed neural patterns consistent with event-related desynchronization, highlighting the models’ potential for brain-computer interface applications.

This presentation introduced a framework that incorporates clinical safety directly into the training of deep learning models for glucose prediction in diabetic patients. Instead of optimizing solely for accuracy, the approach uses a specialized loss function (gcMSE) that penalizes clinically risky errors more heavily. The authors also proposed the PICA algorithm, which gradually shifts training toward meeting clinical standards. Tested on type-1 and type-2 diabetes datasets, the method showed improved safety and reliability, especially in critical glycemic ranges like hypoglycemia.

A system that generates personalized meal plans.

We introduce MetaBoost, a novel hybrid framework that integrates SMOTE, ADASYN, and CTGAN, optimizing synthetic data generation through weighted averaging and iterative weight tuning to enhance the model’s performance (achieving up to 1.87% accuracy improvement over individual balancing techniques).

LEAD is an explainable AI approach that perturbs synthetic critical samples to generate consistent, sparse and robust explanations for disease classification.

This paper is a search based but fast counterfactual generation method for textual data. It proposes three operators to conduct the search in an anytime algorithm. Therefore, the more time it gets the better qulaity countefactual it delivers.

This presentation covers an overview of perturbation-based approaches, as well as evaluation methods of explanations.

The paper introduces the Temporal Fusion Transformer (TFT), a novel attention-based deep learning architecture designed for interpretable multi-horizon time series forecasting that effectively handles a mix of static, known future, and observed past inputs. TFT combines gating mechanisms, variable selection networks, and interpretable attention layers to achieve state-of-the-art forecasting accuracy while offering insights into feature importance, temporal patterns, and regime changes.

The paper proposes PatchTST, a Transformer-based model that enhances long-term time series forecasting through two key innovations: patching of subseries-level time steps and channel-independence, where each univariate time series is processed independently with shared model weights. These techniques significantly improve both computational efficiency and forecasting accuracy, achieving state-of-the-art performance in supervised and self-supervised settings across multiple benchmark datasets.

The presentation reviews a one-year real-world study on a digital twin intervention for managing Type 2 Diabetes (T2D). The study involved 1,853 participants and used continuous glucose monitors, fitness trackers, smart scales, and an app-based digital twin system. The intervention targeted nutritional guidance while tracking sleep, stress, and activity. The program followed three phases: food restriction, reintroduction, and AI-guided normal diet. Results showed a 1.8% average HbA1c reduction, 83% of participants reversed their diabetic status, and over 58% discontinued medication like Metformin. Participants also lost an average of 10 pounds, demonstrating the digital twin’s potential for scalable, non-pharmaceutical T2D management.

RETFound is a self-supervised foundation model that achieves generalizable disease detection from retinal images. It performs competitively across tasks like diabetic retinopathy diagnosis and cardiovascular prediction, using minimal labeled data.

PPO and SAC Algorithms

The paper proposes an online continual learning method for food image classification, using clustering-based exemplar selection and balanced training with knowledge distillation. It outperforms existing methods on the Food-1K dataset.

This paper introduces new constraints in the optimization method to generate realistic and feasible counterfactual explanations.

Reinforcement Learning (RL) is a machine learning technique focused on optimizing decision-making through interaction with an environment, where an agent takes actions to maximize long-term rewards. RL involves a cycle of observation, action, reward, and state transition, requiring a balance between exploration and exploitation. The Markov Decision Process (MDP) provides a mathematical framework for RL, while Q-learning helps develop optimal policies using action-value updates. RL has applications in gaming, robotics, autonomous vehicles, healthcare, and finance, but challenges such as high data requirements, computational costs, safety concerns, and delayed rewards persist.

Across the tasks of summarization, helpful dialogue generation, and harmless dialogue generation, this papers show that RLAIF achieves comparable performance to RLHF.

We evaluate how LLMs like GPT-4 can detect early cognitive decline from EHR notes and show that an ensemble with traditional models boosts diagnostic accuracy, demonstrating the value of hybrid AI approaches in early dementia screening.

This work explores how large language models (LLMs) can be used to answer questions about CGM data in diabetes management and produces an open-source benchmark of time-series question-answering tasks for CGM data in diabetes management.

This paper introduces DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL.

This introduces a three step algorithm that first detects the food items inside the image, second it classifies the food items, and finally it outputs the nutritional facts about the food items in the image.

The Capacitated Set Cover Problem with Exact Coverage involves selecting a minimal number of agents to cover all items such that: Each item is covered by exactly one agent who is eligible (i.e., included in their subset), No agent covers more than m items (capacity constraint).

Since evaluation of counterfactual explanations is a difficult task, his paper tries to automate the process by fine-tuning large language models (LLMs) on human ratings for counterfactuals.

This paper investigates the following three strategies: (1) fusing convolutional feature maps at differing network depths; (2) fusion of bottleneck latent representations prior to classification; and (3) score fusion.

This presentation examines the extension of the PCS (Predictability, Computability, Stability) framework for Reinforcement Learning (RL) algorithm design in digital interventions. Since RL models must be finalized before deployment, robust pre-implementation planning is critical. A case study on Oralytics, a digital oral health intervention, demonstrates how PCS principles enhance RL design. Using a contextual bandit algorithm with Bayesian Linear Regression (BLR), the system optimized intervention effectiveness while ensuring stability. Results showed that BLR performed best, larger clusters improved learning, and simulations required greater diversity. Overall, PCS offers a systematic approach for RL-driven interventions, balancing personalization, computational feasibility, and stability. Further refinements are needed to enhance adaptability and generalization.

This presentation discusses a novel VLM framework aligning ingredient recognition and calorie estimation via training with visual text pairs.

The study presents CXR-LLAVA, an open-source multimodal large language model designed to interpret chest X-rays (CXRs) and generate radiology reports, aiming to alleviate the workload of radiologists and enhance diagnostic accuracy. The model combines a Vision Transformer (ViT-L/16) for image encoding with the LLAMA-2 language model for text generation.

This paper begins with an innovative taxonomy of hallucination in the era of LLM and then delve into the factors contributing to hallucinations. Subsequently, it presents a thorough overview of hallucination detection methods and benchmarks

This paper proposes a framework to build a more generic ConvNet-Boosting (ConvBoost) framework for mainstream HAR ConvNets.

the paper modifies the architecture of the two of the most popular methods in self-supervised continual learning (CaSSLe and Kaizen) and applies them on time-serries data for human activity recognition.

In this work, the authors proposed a VMD-IPSO-LSTM model combining Variational Mode Decomposition and improved Particle Swarm Optimization with LSTM to enhance short-term blood glucose prediction accuracy.

Using transformer based variational autoencoder and Gumbrl-Softmax, TABCF produces counterdactuals the are not biased towards changing the continuous features more often than the categorical features.

This presentation discusses a case study for diabetes pateints with the help of Knowledge-Infused LLM-Powered Conversational Health Agent.

The paper shows that LLMs are decent zero-shot reasoners by simply adding “Let’s think step by step” before each answer. Experimental results demonstrate that the Zero-shot-CoT, using the same single prompt template, significantly outperforms zero-shot LLM performances on diverse benchmark reasoning tasks

Bayesian optimization provides a solution to finding optimal parameters efficiently, especially when dealing with functions where traditional gradient-based methods are infeasible. This technique begins with a probabilistic model (Gaussian process) that represents our uncertainty about the function. As data points are sampled, this model is updated, gradually refining our understanding. The acquisition function then guides the search, balancing between exploring uncertain regions and exploiting known promising areas. This structured, adaptive search enables efficient navigation through high-dimensional spaces, even when some variables are discrete or conditional. Key acquisition functions, such as the upper confidence bound and probability of improvement, are discussed to illustrate how Bayesian optimization outperforms simpler search strategies by integrating prior knowledge and sequential learning.

This paper proposes aa framework to reprogram an existing LLM to perform time series forecasting.

In this work, a novel approach based on DL has been proposed to classify four types of seizures - complex partial seizure, generalized non-specific seizure, simple partial seizure, tonic-clonic seizure, and seizure-free.

The paper addresses the limitations of recurrent neural networks (RNNs) in sequence transduction tasks, specifically their sequential nature hindering parallelization and efficiency, especially with long sequences. The authors aim to overcome these limitations and propose a novel architecture that can learn long-range dependencies more effectively.

This paper proposes a DAG-GNN based VAE to reflect the inter-feature relationships in the produced counterfactuals.

In this work, we introduce a framework for this optimization problem, mathematically formulate the minimum cost channel selection (MCCS), and propose two novel algorithms to solve the MCCS problem.

This presentation discusses the NLGraph benchmark, which was created to evaluate LLMs on graph problems.

A reinforcement learning approach to solve the vehicle routing problem.

This study aims to quantify the influence of various meal-related factors on predicting postprandial blood glucose levels (BGLs) at different time intervals (15 min, 60 min, and 120 min) after meals by using deep neural network (DNN) models.

This paper proposes a series of adaptations and enhancements around BERT to best work with IMU data in mobile sensing applications.

This paper study provides a novel contribution to the field of signal processing and DL for ECG signal analysis by introducing a new feature representation method for ECG signals.

DualNet is a state-of-the-art neural net architecture that uses the idea of self-supervised learning to learn general features of image data to prevent catastrophic forgetting.

The authors aimed to fit a simple mathematical model on dynamic postprandial glucose data from repeated MMTTs among participants with elevated T2DM risk to identify response clusters and investigate their association with T2DM risk factors and gut microbiota.

This paper introduces a technique to produce synthetic counterfactual images to train a model free of bias.

This presentation discusses label-free tunning of zero-shot classifiers using language and unlabeled image collections.

In this paper, the authors introduce Random Erasing, a new data augmentation method for training the convolutional neural network (CNN). In training, Random Erasing randomly selects a rectangle region in an image and erases its pixels with random values.

In this work, the authors present self-supervision strategies that can be used to learn informative representations from multivariate time series.

This paper proposes Health-LLM, a framework in the healthcare domain that aims to bridge the gap between pre-trained knowledge in current LLMs and consumer health problems.

This study leveraged advances in continuous glucose monitoring (CGM) and cognitive ecological momentary assessment (EMA) to characterize dynamic, within-person associations between glucose and cognition in naturalistic environments.

This presentation discusses a new way to encrypt data with the help of LLMs.

In this work, the authors revisit the most fundamental building block in deep learning, the multi-layer perceptron (MLP), and study the limits of its performance on vision tasks.

CountER tells us under what conditions a chosen item would not be recommended to a person anymore.

This paper presents a framework for learning representations from EEG signals via contrastive learning.

This paper presents a robotic object hitting task to quantify sensorimotor impairments in participants with stroke

This presentation introduced the latest AI assistance from Meta company.

This presentation presents self-supervised method that learns invariant representations rather than covariant ones.

This paper proposes a framework for better understanding what is happening behind the training of over-parameterized neural nets, for which the objective function is not convex, not even locally

Adversarial Counterfactual Explanations propose to develop a filter that robustifies a model against adversarial attacks and helps to transform them into counterfactual explanations.

This presentation discusses solving complex AI tasks with a new model named HuggingGPT.

This paper proposes a contrastive learning based approach to solve the problem of causal representation learning, under the assumption of linear-gaussian representation, and a non-linear mixing function

In this paper, the authors present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently.

This paper proposes a novel approach for learning such representations from multi-channel EEG time-series, and demonstrate its advantages in the context of mental load classification task.

This presentation presents semi-supervised neural network for freezing of gait classification.

This presentation aims understanding the relationship between HbA1c and brain health across the entire glycaemic spectrum.

CounterNet proposes an end-to-end training of the predictor and explanation generator to solve some key issues associated with post-hoc/model agnostic explainers.

This presentation discusses missing data imputation through fuzzy c-means clustering.

This paper proposes a new self-supervised learning based approach for industrial time-series analysis which uses data-augmentation techniques

This paper proposes two binary imbalanced data classification methods with extreme learning machine autoencoders and generative models. Their methods outperform SMOTE and similar data balancing tools on most benchmark datasets.

This presentation describes and explains Fourier Transform, Short-Time Fourier Transform, and Wavelet Transform.

In this paper, various data augmentation methods for wearable sensor data are proposed. The proposed methods and CNNs are applied to the classification of the motor state of Parkinson’s Disease patients.

inability to recognize counterrfactual diagnosis from associative diagnosis can lead to disastrous outcomes.

This article demonstrate a new approach to implement counterfactual explanations (CE) without necessarily employing any generative model. As generative models come with several drawbacks in CE generation, the approach in this paper is based on Epistemic and Aleatoric Uncertainty reduction.

This paper discusses the available methods to detect change points in time-series data.

This paper presents SimCLR, a simple framework for contrastive learning of visual representations.

A brief introduction to causality with the emphasis on the book “The Book of Why”, by Judea Pearl.

This article focuses on defining what counterfactuals mean in the context of timeseries data and proposed an algorithm that generates plausible and minimally distant counterfactual explanations.

A fundamental limitation of applying semi-supervised learning in real-world settings is the assumption that unlabeled test data contains only classes previously encountered in the labeled training data. However, this assumption rarely holds for data in-the-wild, where instances belonging to novel classes may appear at testing time. Here, we introduce a novel open-world semi-supervised learning setting that formalizes the notion that novel classes may appear in the unlabeled test data.

Conducting a causal inference study with observational data is a difficult endeavor that necessitates a slew of assumptions. One of the most common assumptions is “ignorability,” which argues that given a patient (X), the pair of outcomes (Y0, Y1) is independent of the actual treatment received (T). This assumption is used in this paper to develop an AI model for calculating the Average Treatment Effect (ATE).

An introduction to counterfactual explanations using the novel approach developed by Microsoft Corporation, India.

In this position paper, the authors define interpretability and describe when interpretability is needed (and when it is not). Next, they suggest a taxonomy for rigorous evaluation and expose open questions towards a more rigorous science of interpretable machine learning.

Propensity score is the probability of a patient receiving the treatment. Given the propensity score, the covariates are independent from the treatment variable. Therefore, it can be used to turn an observational data into a data collected using a randomized trial. This paper talks about different methods of applying propensity scores to causal studies.

This study implemented SimCLR on EHR data to detect rare dseases. In general, rare diseases are underrepresentative classes in any clinical dataset. Therefore, using SimCLR (contrastive learning) boosts their classification accuracy.

We propose an automated risk prediction system that can make recommendations to clinicians in real-time with machine learning classifiers that predict the risk of developing neurological impairment.

Neural ODEs describe a flow from the input to the output. They utilize the traditional ODE solvers to reach this goal. By doing so, they introduce a trade-off between accuracy and computational cost.

This paper reduces drug-drug interaction in automated medication recommendation with help of self-supervised adversarial distribution regularization

TIME is a great tool for explaining the predictions made by a model. It is very similar to Grad-CAM, however, it can work with tabular data as well. Also, TIME can provide global expainability as well.

This study used machine learning methods to identify and detect the moments of cannabis use in participants. The model was trained using a leave-one-subject-out method, and later personalized using transfer learning

This research paper aims at building an entropy layer for end-to-end explainable AI framework. Unlike LIME and Grad-CAM, Entropy Net does not rely on any auxuliary tool/model to explain its predictions. Also, Entropy Net focuses on downsizing the concepts required to explain predictions.

This research aimed to develop a lifestyle intervention system with an emphasis on time series forecasting and also to develop a generic multimodal activity forecasting scheme that can be used with either early fusion or late fusion mechanisms.

REP-Net is a counter to traditional transfer learning for On-board model training with more focus on memory efficiency.

TDragonnet is an end-to-end neural network architecture that uses propensity score to estimate the average treatment effect.

This paper provides an overview of basic and advanced approaches of time-series data augmentation.

Training GANs require large amount of data. If tried with small datasets, the discriminator often times overfit producing meaningless feedback to the generator. One solution to training GANs with smaller dta could be using adaptive data augmentation.

This paper proposes GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, and test its performance in the few-shot setting.

Including domain knowledge can increase explainability without compromising on model performance. This paper mainly discusses a preprocessing technique to incorporate domain knowledge in a way they become very useful for explaining model’s predictions.

This paper proposes a new deep learning framework for irregularly sampled time-series data that is called Multi-Time Attention Networks.

This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision.

Incorporating domain knowledge to neural networks is a creative and case specific approach. This paper modifies the loss function of a fully-connected network with domain knowledge from kinetics which helped the model make precise prediction in its regression task.

An algorithm for meta-learning that is model-agnostic, in the sense that it is compatible with any model trained with gradient descent and applicable to a variety of different learning problems, including classification, regression, and reinforcement learning.

Prelim Exam Practice Presentation of Ramesh Sah.

A reinforcement learning approach to solve the vehicle routing problem.

Optimizes change in consumed nutrients while driving users to their desired diet.

Practice presentation for Master defense of Ramesh Sah.

This paper proposes a new channel pruning method based on artificial bee colony algorithm (ABC), dubbed as ABCPruner, which aims to efficiently find optimal pruned structure.

LIME is a great tool for explaining the predictions made by a model. LIME can explain any model regardless of their type, it works by building a linear model on vicinity of the sample intended to be explained.

This paper addresses the generalization gap on large neural networks with adaptive gradient-based optimizers and proposes a partially adaptive solution to overcome the problem.

In this talk, we present and discuss different strategies for data augmentation.

Decision boundaries are imposible to be visualized in hogh dimensional feature sets. Instead of visualizing them, we can characterize them and make them useful

We focus on devising algorithms that combine data about physical activity and engagement with the app to predict future physical activity performance.

In this talk, Ramesh explained and highlighted different methods to deal with imbalanced dataset. He shared his experience working with imbalance datasets and partical tips and tricks.

Neural networks have a bad reputations as they are treated like black boxes and lack interpretations on the results they make. Grad-CAM can slightly interprete what is driving the model to make a decision.

Enhancement of accuracy for wrist-worn sensor based lying posture classification via transfer learning

Optimization Algorithms to begin with Sparse Coding.

Continous detection of stress using external heart rate monitor to drive inter beat interval feature on a smartphone.

Introduction to variational autoencoders.

Techniques for training GANs using Semi-supervised learning.

Examine the relationship between stress and physical activity and healthy eating in both direction - negative and positive - using ecological momentary assessment (EMA).

A novel framework called SubNetwork Routing (SNR) to achieve more flexible parameter sharing while maintaining the computational advantage of the classic multi-task neural-network model.

Introduction to binary neural networks with details about training and evaluation of these networks.

Introduction to the statistical learning theory, and a prelude to PAC learning

Adding Online-Learning ability to TinyML.

Review of the “Any-Width Networks” from CVPR 2020 workshops.

How does the brain prevent interference of memory and sensory data?

Emotion classification using self-supervised learning and six signal transformations as pretext tasks.

Review of the paper “Attention Is All You Need”

Review of the recent research on bringing the transformer architecture to the computer vision field.

Quantifying user burden for EMA studies with frequent and one-question EMA assessment.

Review of the AI Economist paper by Salesforce Research.

Overview of two visions systems by Open AI

Picking phase and detecting earthquakes using encoder-decoder transformer model.

Review of ICLR'18 paper by Novak et.al.

Understanding why transfer learning is successful

Practical tips for researchers for managing ML experiments

General high-level introduction to Gaussian Processes for time-series data modelling.

Introduction to neural tangent kernel

Review of self-supervised learning for human activity recognition

Review of The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks

Re-presentation of Yann LeCun’s AAAI talk and discussion on semi-supervised learning methods.

Paper Review: Elsayed, G.F., Goodfellow, I.J., & Sohl-Dickstein, J. (2019). Adversarial Reprogramming of Neural Networks. ICLR.

Paper Review: Kornblith, S., Norouzi, M., Lee, H., & Hinton, G.E. (2019). Similarity of Neural Network Representations Revisited. ICML

An introduction to Differential Privacy and Federated Learning concepts.

Practice presentation for SMARTCOMP conference.