Deep Learning Group

Addressing critical challenges for more robust models

We study the underlying principles that make deep learning systems effective, interpretable, and reliable. Our group takes a scientific approach to these challenges—the science of deep learning—centered around models, methods, and safety. Viewing neural networks through the lens of dynamical systems theory helps us understand and improve their behavior and inspire new architectures.

Our prior work has shown that analyzing recurrent neural networks as dynamical systems helps explain and mitigate problems like exploding and vanishing gradients. This perspective has also revealed structural biases in networks that model dynamical systems, motivating the use of physics-inspired inductive biases and higher-order integrator schemes for training continuous-time neural networks. Reformulating such networks as stochastic differential equations provides a natural framework for injecting perturbations during training, improving both stability and inference robustness.

Building on these foundations, our recent research focuses on large-scale generative diffusion models for spatio-temporal forecasting in scientific domains such as earth system modeling and fluid dynamics. We are also exploring how foundation models can combine reasoning and multimodal information to enable more adaptive and informed decision-making in complex environments. Beyond performance, we are deeply engaged in AI safety and reliability. We study vulnerabilities in large language models (LLMs), including jailbreaking and backdoor attacks. 

Together, these efforts aim to build a principled understanding of deep learning—pushing toward a future where AI systems are not only powerful and expressive, but also scientifically grounded, robust, and safe.

---