Adaptive Gradient Harmonization: Mitigating Modality Dominance in Unified Representation Learning
Innovated the Modality Fairness Controller (MFC) to dynamically balance learning rates across Vision and Audio.
Lakshya
About Me
M.Tech AI & ML student specializing in Multimodal Representation Learning and Advanced Systems Engineering.
Education
M.Tech in AI & ML
2025-01-01
2027-01-01
Manipal Institute of Technology
B.Tech in CSE
2021-01-01
2025-01-01
SRMIST, Chennai
Focus Areas
Multimodal AI
Systems Engineering
GPU Optimization
Golang & PostgreSQL
🔬 Research Highlights
My research focuses on Multimodal Representation Learning, specifically addressing challenges like Modality Laziness and Gradient Harmonization.
Key Works:
- Adaptive Gradient Harmonization: Mitigating Modality Dominance in Unified Representation Learning. (Under Review)
- DFU Image Classification Cost Reduction: Knowledge Distillation & Network Pruning for efficient edge deployment.
I am also a strong proponent of High-Performance Engineering, building custom AI infrastructure and optimizing low-level GPU operations.
Featured Publications
•
•
1 min read
DFU Image Classification Cost Reduction
Knowledge Distillation & Network Pruning for efficient edge deployment, achieving 3.47x CPU speedup.
Lakshya
Recent Publications
(2025).
Adaptive Gradient Harmonization: Mitigating Modality Dominance in Unified Representation Learning.
Under Review.
(2025).
DFU Image Classification Cost Reduction.
Pipeline / In Progress.
Technical Projects
Cold Storage Management
Robust inventory tracking system. identified and patched a critical Race Condition bug involving inventory overselling during simultaneous gate pass approvals.
•
1 min read
Custom AI Infrastructure
Built custom TensorFlow wheels with CUDA support optimized for next-gen NVIDIA hardware (RTX 5070 Ti / Blackwell Architecture).
•
1 min read
GoBlog CMS
High-performance Content Management System built with Go and PostgreSQL. Engineered for concurrent content delivery and scalability.
•
1 min read