This work tackles the question: can a model point to exactly where a face has been forged, without ever being trained on pixel-level fake/real labels? By combining handcrafted forensic cues with self-attention, we get localization without pixel-level supervision, along with explanations without any extra supervision other than the normal binary classification labels.
We are working on an explainable molecular classification task that enhances standard Graph Neural Networks (GNNs) by incorporating 3D structural information. While typical GNNs only look at 2D chemical bonds, we add 3D atomic coordinates—like distances and angles—directly into the network's learning process to better capture a molecule's true shape.
Open world learning is a branch of learning where tasks from different distributions arrive in a stream and the number of distributions and the point of distribution change are unknown. This project aims to design a mixture-of-experts method for open world regression where tasks belonging to similar distributions are learned by the same expert, resulting in positive transfer.
Vision Transformers (ViTs), despite their success in image classification, suffers from spurious correlations in the dataset. Techniques to mitigate this often rely heavily on pixel-level human annotations. We develop a technique to generate annotation-style masks using Flow-Matching training on a very limited budget of annotations, greatly reducing the manual burden in creating annotations.
Paper →
Neural Processes are a deep learning based version of Gaussian Processes, and work very well as meta-learning models. However they require a large number of tasks to reach a good level of performance contrary to the expectation in meta-learning. We developed our method of pseudo-tasks to reduce this task requirement and accelerate generalization for Neural Process models.
Paper →
An unsupervised Time Series Anomaly Detection method which learns relationships among groups of variables using a Hierarchical Graph Neural Network (H-GNN). This multi-view information, along with temporal and structural information, helped to improve detection performance and make the model more robust to anomalies during training.
Paper →
We address the challenge of incorporating 3-dimensional structural information in graph kernels for machine learning-based virtual screening, a crucial task in drug discovery. Existing kernels that capture 3D information often suffer from high computational complexity, which limits their scalability. To overcome this, we propose the 3-dimensional chain motif graph kernel (c-MGK), which effectively integrates essential 3D structural properties—bond length, bond angle, and torsion angle—within the three-hop neighborhood of each atom in a molecule. In addition, we introduce a more computationally efficient variant, the 3-dimensional graph hopper kernel (3DGHK) which reduces the complexity of c-MGK.
Paper →
Vision Transformers (ViTs) learn highly label-correlated features, but may rely on coincidental cues that are not justifiable from a human perspective. We propose incorporating minimal human guidance directly into ViT feature extraction to bridge this gap between model learning and human justification. Our approach improves justification without compromising accuracy, while also reducing data and training costs and improving robustness to contextual variations.
Paper →
Warping Resilient Scalable Anomaly Detection model learns warp resilient embeddings for time series data.
Paper →
Deep Extreme Mixture Model for Time Series Forecasting.
Creating diverse novel compounds with optimal drug-like qualities while addressing long-term toxicity assessments is challenging. Target-based drug design, especially for new proteins, adds complexity. We propose a model which minimizes the generation of toxic molecules for both known and novel molecules.
Deep Learning based android app to help farmers to save time by counting crops in seconds.
GitHub →
This image represents our project on occlusion-invariant face detection for an automatic attendance system. The system efficiently captures and processes classroom images to accurately record student attendance, even when faces are occluded.
While DNN models demonstrate impressive performance, their adoption in critical tasks like the medical domain, self-driving cars, etc. is challenged by the lack of explainability on its predictions. Our model is trained with pretrained black box models providing explanations in the form of text.
Our study proposes bottlenecked backpropagation, a novel method to enhance privacy in deep learning with sensitive data. Unlike adding noise to gradients, our method prioritizes gradient direction, reducing noise's impact on magnitude. This leads to faster convergence, higher accuracy, and improved protection against information leakage.
Explainable artificial intelligence (XAI) is a set of processes and methods that allows human users to comprehend and trust the results and output created by machine learning algorithms. Our work attempts to bring explainability in deep learning classification models using the intervention of textual concepts instead of visual ones.
This work focuses on predicting disease categories automatically from symptoms documented in the Afaan Oromo language, employing various classification algorithms such as support vector machines, random forests, logistic regression, and Naïve Bayes, as well as deep learning approaches including LSTM, GRU, and BiLSTM.
Paper →