All attendees must register online before arriving to retrieve their badge.
From Visual Recognition to Embodied AI
Twenty years ago, at ECCV 2006, the challenge that captivated me was teaching machines to recognize objects in the visual world. Since then, computer vision has undergone an extraordinary transformation.
Through a journey from early visual recognition, through Kinect and HoloLens, to embodied AI at Wayve, this talk will reflect on some of the technical ideas and personal lessons behind that transformation: learning rather than engineering; betting on data and compute scaling; and being careful not to confuse today’s constraints with tomorrow’s limits.
Today, the opportunity is much bigger: building systems that can understand enough of the physical world to act intelligently within it. I’ll look at autonomous driving as an early proving ground for embodied AI, and at some of the open questions on the path toward more general embodied intelligence. If the past twenty years are any guide, the boundaries of what’s possible can move faster than we expect.
Can AI Build New Knowledge?
Modern AI systems are remarkably good at using knowledge acquired during large-scale training. But it is still unclear whether they can genuinely build new knowledge from experience: identify what they do not know, acquire the right evidence, form new abstractions, revise their understanding of the world, validate what they have learned, and retain it for future use. One way to view current AI progress is as a human-machine continual-improvement loop. A model is trained; humans inspect its failures; researchers design new datasets, benchmarks, architectures, losses, memory mechanisms, or post-training methods; the model is retrained and evaluated; and the cycle repeats. In this sense, AI systems are improving continually, but much of the knowledge-building still happens outside the model, through human diagnosis and design. The question for next-generation AI is whether more of this loop can be internalized. Can future systems identify their own knowledge gaps, decide what evidence they need, learn from video or interaction, store useful memories and exceptions, validate new knowledge, and update themselves without losing previous capabilities?
The Art Gallery Curator, Ioannis Siglidis, will moderate a panel discussion featuring the artists Sonia Bernac, George Cazenavette, Robin Champenois and Cezar Mocan
| ECCV uses cookies for essential functions only. We do not sell your personal information. Our Privacy Policy » |