A collection of research, machine learning projects, and experiments that taught me one important lesson: the model is usually innocent until proven guilty by the data.
Going Beyond Explainability: Learning Justifiable Vision Transformers Guided by Flow Matching
Authors: Thomas John, Akshay Krishna, Jotipriya Das, Mrinal Das
Published at ECML PKDD 2026. The work explores learning vision transformers that are not only accurate but also provide more justifiable reasoning by leveraging flow matching techniques.
Explainable AI • Vision Transformers • Flow Matching • Deep LearningDesigned a heart disease prediction framework combining Vertical Federated Learning with Zero-Knowledge Proofs and Differential Privacy to enable collaborative learning without exposing sensitive data.
Federated Learning • Zero-Knowledge Proofs • Differential PrivacyBuilt a multi-output machine learning system capable of predicting diagnosis, management strategy, and severity simultaneously using the Regensburg Pediatric Appendicitis dataset.
Random Forest • XGBoost • Multi-Output LearningDeveloped convolutional neural network models with skip connections and evaluated different architectures for robust image classification.
CNN • TensorFlow • Keras • Computer VisionTrained machine learning models to recognize music genres from audio features and explored how machines can learn patterns that humans simply call “good taste.”
Audio Processing • Random Forest • SVM • Machine Learning