Date: July 17 / 10:30am followed by a buffet lunch
Location: B612, salle 6
Program :
Scaling Laws vs. Neural Laws: Toward More Natural Artificial Vision by Thomas Serre
The remarkable progress of modern computer vision has been propelled by the relentless logic of scaling laws: bigger models, more data, more compute, predictably better performance. On benchmarks like ImageNet, deep networks now match or even surpass human accuracy. Yet beneath these headline results, the alignment with human vision is fragile: on deceptively simple probes from the cognitive sciences, even the largest models drop to near-chance, and on ImageNet itself, models that reach human accuracy do so by strikingly different visual strategies — a divergence that, troublingly, widens with scale.
In this talk, I will argue that the path to more natural artificial vision lies not in pushing scaling laws further, but in a deeper engagement with the neural laws of biological vision: developmental principles that shape how brains learn to see, and architectural constraints that impose strong inductive biases on cortical processing. I will share recent work from my lab on two such laws. On the learning side, I will present preliminary evidence that pairing the right learning objectives with naturalistic video — sequences of object transformations like those the developing brain encounters — can pull deep networks toward markedly more human-like visual strategies. On the architectural side, I will show how recent advances in state space models can scale cortical recurrent feedback into a brain-inspired alternative to transformer self-attention, one that closes the gap on cognitive probes where transformers fail, and on ImageNet traces more favorable scaling laws than transformers.
Together, these results point toward a future in which scaling laws and neural laws are in agreement rather than in tension, and in which computer vision, in dialogue with the brain sciences, helps build AI systems that are not only more capable but more aligned with the kind of intelligence we ultimately seek to understand and emulate.
Learning robotic behaviors with optimal control and world-models by Ludovic Righetti
Sensor-driven simulators, or “world-models”, offer a unique opportunity to design planning and control algorithms working as close as possible to a robot’s embodiment. By enabling optimization directly in multi-modal sensor space, such as vision and touch, these models can capture complex dynamics that are difficult to model, including soft bodies and deformable objects. However, optimization algorithms designed for classical simulators often fail to transfer to world-models due to the high dimensionality of sensor data and the difficulty of defining physical goals in latent space. In this presentation I will discuss our recent work in learning world models and designing planning and reinforcement learning algorithms tailored to work with them. Since the algorithms we design are intended for real applications that could change how we organize our societies, I will end the presentation with a broader discussion on the impacts of robotics research on society and the role engineers ought to play.
Short bio:
Ludovic Righetti is the Glenn Y. Louie Professor of Engineering at New York University and founding co-director of the NYU Center for Robotics and Embodied Intelligence. He also holds an international chair at the Artificial and Natural Intelligence Toulouse Institute. He has an Engineering Diploma in Computer Science and a Doctorate in Science from the Ecole Polytechnique Fédérale de Lausanne. He was previously a postdoctoral fellow at the University of Southern California and a group leader at the Max-Planck Institute for Intelligent Systems. His work has received several awards including the Georges Giralt PhD Award, IROS and RA-L Best Paper Awards, IEEE RAS Early Career Award and NYU Jacobs Excellence in Education Innovation Award. His research focuses on the planning, control and learning of movements for autonomous robots, with a special emphasis on legged locomotion and manipulation. He is also interested in the broader societal impacts of robotics and AI and regularly works with international organizations on the topic, especially on issues related to peace and security.