Abstract: Biological neural networks achieve something today's artificial intelligence  struggle to replicate: they learn from sparse data, adapt continuously, and deliver robust inference—all while operating at roughly 30 watts. The gap is not just quantitative—it points to missing principles. In this talk, I explore what it means to build efficient intelligence. I begin with event-based representations—fundamental to biological perception—and show how representations are the bottleneck to current event based systems. I then move to computation, showing activation pruning in FPGA-based neural networks reduces unnecessary MAC operations, unlocking higher throughput without increasing power. Finally, I turn to data efficiency. I present curriculum-driven learning rate schedules that adapt during training, and examine how generative models like Stable Diffusion can be used to synthesize prototypical, low-complexity data—aligning with an Occam’s Razor view of learning. The goal is not incremental optimization, but a reframing: if intelligence is to scale sustainably, it must become dramatically more efficient—across representation, computation, and data.

About the Speaker: Raghavendra Singh received his Ph.D. in Electrical Engineering from the University of Southern California (USA). He brings over 20 years of experience in designing, developing, and deploying computer vision algorithms to address real-world challenges. He has served as a Senior Researcher at IBM Research, co-founded a camera technology startup, and was Chief AI Architect at a sustainable energy solutions startup.  His technical interests span image and video processing, visual representation, computer vision, and computational neuroscience. He is currently a Professor (Off-Campus) in CSIS at BITS Pilani, where he collaborates with the Autonomous Vehicles Lab in Hyderabad.