Abhyarthana Bisoyi
Assistant Professor,
Odisha University of Technology and Research
About the Speaker
Mrs. Abhyarthana Bisoyi is an Assistant Professor in the School of Electronic Sciences at Odisha University of Technology and Research (OUTR). She has completed her Bachelors in Electronics and Communication Engineering in the year 2011 and Masters from National Institute of Technology Rourkela (NIT Rourkela) in Electronics and Instrumentation Engineering in the year 2013. From then to now, she has teaching experience of more than 12 years. She has guided 22 Masters students and more than 100 undergraduate students. She is currently pursuing her PhD from OUTR and her research area includes Digital VLSI, FPGA, CNN, Approximate and Neuromorphic Computing.
Co-Design and Verification of FPGA-Efficient CNNs and Next-Generation Deep Video Codecs
Overview
Deploying deep learning at the resource constrained edge requires a paradigm shift in hardware software codesign. This presentation details state of the art architectures for FPGA efficient convolutional neural networks and deep video codecs. We exploit hardware aware hybrid model compression, utilizing column-based bit level sparsity, mixed precision quantization, and sparsity aware multipliers to reduce on chip dynamic power. Furthermore, we explore scaling neural video coding models and leveraging multimodal large language models for semantic reconstruction. Finally, we establish rigorous verification and hardware emulation strategies to guarantee functional accuracy while maximizing energy efficiency for edge deployed autonomous vision systems.
Key Points
- Hardware-Aware CNN Compression and Co-Design on FPGAs
- Scaling Up Deep Video Codecs and Cross-Modality Semantic Compression
- Rigorous Quality Assessment and Verification Methodologies
