State space models (SSMs) are emerging as a compelling alternative to transformer-based architectures, providing a mathematically grounded and computationally efficient framework for modeling long-range dependencies in sequential data. While transformers rely on attention mechanisms that scale quadratically with sequence length, SSMs use linear recurrence relations and convolutional structures to capture temporal dynamics with far greater efficiency and scalability.

This session explores how state space models enable high-performance learning across modalities such as natural language, vision, and time series while reducing latency and memory overhead. Attendees will learn about the key innovations behind Structured State Space Sequence models (S4), Mamba, and Hyena, and how these architectures are redefining what is possible in long-context understanding. The talk will highlight why state space models may represent the next paradigm in AI, one that balances accuracy, efficiency, and scalability in the era beyond transformers.

What You Will Learn

• The core principles behind state space models and how they differ from transformers

• How architectures like S4, Mamba, and Hyena achieve efficient long-context reasoning

• Practical insights into the potential of SSMs as the foundation for next-generation AI models

Who Should Attend

AI researchers, ML engineers, and data scientists interested in cutting-edge model architectures, efficient sequence modeling, and the future of large-scale AI systems.