| Current Research Directions |
|
I am currently exploring several research directions that build upon my experience in graph neural networks,
high-dimensional algorithms, and machine learning systems.
Panorama Extensions: Building on my work at University of Copenhagen, I'm investigating
advanced techniques for high-dimensional nearest neighbor search, particularly applications to recommendation
systems and large-scale graph analytics.
Machine Unlearning: Following my internship at CMU, I'm working on formalizing unlearning
mechanisms with provable guarantees, exploring how machines can selectively forget information while
preserving essential knowledge.
Lifting & Bispace Optimization: Developing neurosymbolic architectures that combine the
expressiveness of neural networks with the interpretability of symbolic reasoning, particularly for
scientific applications.
These directions align with my goal of making AI systems more efficient, interpretable, and reliable
for real-world applications.
|
|
Panorama: Fast-Track Nearest Neighbors
UNDER REVIEW
Collaborators: Alexis Schlomer, Akash Nayar, Dr. Panagiotis Karras (University of Copenhagen),
Dr. Sayan Ranu (IIT Delhi),
Dr. Jignesh M. Patel (University of Wisconsin-Madison)
PANORAMA is a machine learning-driven approach that tackles the ANNS verification bottleneck through data-adaptive
learned orthogonal transforms that facilitate accretive refinement of distance bounds. Our transforms compact over 90%
of signal energy into the first half of dimensions, enabling early candidate pruning with partial distance computations.
Key Contributions:
• Data-adaptive learned orthogonal transforms for dimension reduction
• 2-30× end-to-end speedup with no recall loss across diverse datasets
• Integration with state-of-the-art ANNS methods (IVFPQ, HNSW, MRPT, Annoy)
• Evaluation on modern embedding spaces including OpenAI's Ada 2 and Large 3
• [arXiv Paper]
Status: Under peer review
|
| Recently Completed Projects |
|
Erudite: Agentic Knowledge Graph System
COMPLETED
Completion: July 2024 | Collaborators: IIT Delhi Research Team
Developed a multi-agent RAG pipeline for dynamic expanding knowledge graphs integrating 5+ sources
(Semantic Scholar/YouTube) using Claude Haiku, achieving interactive graph generation in 130s.
System Architecture:
• Fault-tolerant architecture with parallel agent execution
• 3 retries + exponential backoff for robustness
• Modular templates for easy integration of new data sources
• Real-time knowledge graph expansion and updating
Performance Metrics:
• Interactive graph generation in under 130 seconds
• Support for 5+ heterogeneous data sources
• 99.5% uptime with fault-tolerant design
• Scalable to millions of knowledge graph nodes
Repository: https://github.com/VanshRamani/ERUDITE
|
|
Optimised 2-D Bin Packing for VLSI Gate Design
COMPLETED
Completion: Dec 2023 | Collaborators: Dr. Preeti Panda
Optimized VLSI gate packing by benchmarking approximation methods and developing a visualization pipeline
for the 2D packing and wiring problem. Designed greedy and annealing algorithms achieving over 96% packing efficiency.
Technical Achievements:
• 96%+ packing efficiency while minimizing wire length
• Multiple sorting heuristics implementation and analysis
• Comprehensive algorithm complexity evaluation
• Edge case handling and performance optimization
Algorithms Implemented:
• Bottom-Left Fill (BLF) heuristic
• Best-Fit Decreasing Height (BFDH)
• Simulated Annealing optimization
• Custom genetic algorithm variant
Repository: https://github.com/VanshRamani/Gate-Packing
|
|
Comprehensive Deepfake Detection System with MTCNN and Fact-Checking
COMPLETED
Completion: Nov 2023 | Collaborators: Dhruv Joshi
Developed a robust Deepfake detection model by optimizing two key components: an MTCNN-based frame-by-frame
classifier with EfficientNet and a fact-checking approach that matches audio to visual features.
System Components:
• MTCNN for face detection and alignment
• EfficientNet for frame-by-frame classification
• Audio-visual feature matching for fact-checking
• Ensemble learning with weighted mean combination
Performance Results:
• 93% accuracy in deepfake detection
• Award-winning performance at CodeWars competition
• Real-time processing capability
• Robust against various deepfake generation methods
Repository: https://github.com/jdhruv1503/deepfake-detection
|
|
Looking ahead, I aim to broaden my understanding of machine learning fundamentals and explore new domains
that complement my current expertise in graph neural networks and high-dimensional algorithms.
Reinforcement Learning: I hope to get my hands into RL this semester, exploring how agents
learn optimal policies through interaction with environments. This connects naturally with my interest in
sequential decision-making and optimization.
Graph Signal Processing: Building on my graph learning background, I want to study how
signals propagate through networks and how classical signal processing techniques apply to graph-structured data.
Fundamental Probabilistic Models: I plan to study core probabilistic frameworks like
Markov Decision Processes (MDPs), Hidden Markov Models (HMMs), and Markov Logic Networks (MLNs) to
strengthen my theoretical foundations.
These areas will provide a solid mathematical foundation for tackling more complex problems in AI and
understanding the probabilistic underpinnings of modern machine learning systems.
|
|
