Vijay Srinivas Agneeswaran

Short CV: Vijay Srinivas Agneeswaran has a Bachelor’s degree in Computer Science & Engineering from SVCE, Madras University (1998), an MS (By Research) from IIT Madras in 2001, a PhD from IIT Madras (2008) and a post-doctoral research fellowship in the LSIR Labs, Swiss Federal Institute of Technology, Lausanne (EPFL). He has spent the last twenty plus years creating intellectual property and building AI/ML based products in Industry and academia. He is currently head of cloud + E AI leader and architect at Microsoft. Vijay has led AI efforts in areas such as forecasting, anomaly detection, recommenders etc. and productionized several models for business consumption. He and his team have built spectral transformers and patented and published these in top conference such as NeurIPS 2023, AAAI 2024 (Industry track) and ICASSP 2025. In addition, he is the responsible AI champion for the C + E team, reviewing 50+ models for responsible AI standards compliance. He was heading the ML platform and the data sciences foundations teams at Walmart in his previous role. He has seven granted US patents as well as numerous other disclosures that have been filed in Indian and US Patent offices.

Title of Talk 1:  Recent Advances in Computer Vision: Convolutional Networks to Transformers

Synopsis: The talk details recent advances in computer vision, starting from how convolutional networks were used in CV and how transformers revolutionised computer vision. It also explains how the Vision Transformer (ViT) was a seminal work. It also explains issues in ViT and subsequent optimizations made on ViT such as Pyramid vision transformer, Swin transformer, CSWin, LiTv2, etc. It motivates need for spectral transformers and outlines GFNet, HiloAttention and WaveViT as prominent spectral transformers. This sets the stage for our NeurIPS 2023 publication, the Patro, Batro N., & Agneeswaran, Vijay.S. (2023). Scattering Vision Transformer: Spectral Mixing Matters. NeurIPS 2023. A brief view of SVT and future research directions complete the talk.

Title of Talk 2:  Comparing Transformers and State Space Models for Computer Vision Use Cases: Image Classification to Segmentation

Synopsis: This talk introduces the basics of transformers and how they have revolutionised the field of computer vision as well as how Vision Transformer (ViT) was a seminal work. It also explains issues in ViT and explores subsequent optimizations made on ViT such as Pyramid vision transformer, Swin transformer, CSWin, LiTv2, etc. It motivates need for spectral transformers and outlines GFNet, HiloAttention and WaveViT as well as SVT (our NeurIPS 2023 publication) as prominent spectral transformers. The talk also outlines state space models with examples of prominent state space models such as S4 and outlines a brief survey of state space models. It also gives detailed comparisons of state space models and transformers including theoretical analysis and performance metrics such as accuracy, latency and throughput. Finally, common computer vision use cases such as image classification and segmentation are discussed along with performance comparison of transformers and state space models for each use case.

Title of Talk 3:  Research Topics in Deep Neural Networks

Synopsis: This research focused talk provides the history of development of deep neural networks starting from convolutional networks, recurrent networks and transformers. It explains the fundamentals of transformers and also lays out a 360 degree survey of transformers from different perspectives including robustness, efficiency (both number of FLOPS and parameters), transparency, inclusiveness, spectral complexity etc. – this provides a comprehensive view of improvements made to transformers in the last few years. This talk also highlights the key state space models and how they are different from transformers and are able to handle longer sequence lengths. It also provides a performance comparison of state-of-art transformers with recent state space models highlighting the open areas of research in both transformers and state space models. The talk also delves into other areas of AI such as privacy to highlight open research questions.