Dynamics and Learning in Deep Neural Networks

The unexpected and unprecedented performance of modern AI systems—such as deep convolutional networks, diffusion models, and large language models—is raising fundamental questions about the principles underlying their success. These architectures rely on global gradient-based optimization, which is energy intensive and not directly aligned with how biological systems learn. In this talk, I will first review the state of the art in AI and neural learning models, highlighting recent efforts to develop local and dynamical learning mechanisms inspired by biological computation. I will then discuss a new framework based on asymmetric deep recurrent networks, where learning arises through the stabilization of dense connected clusters of stable fixed points—the Representation Manifold. This approach connects deep learning with ideas from statistical physics and neural dynamical systems, offering a physically grounded view of how neural networks can learn and represent information efficiently without explicit gradient computations.