Deep neural networks (DNNs) are commonly used as universal function approximators where individual neurons learn to induce changes/non-liniearities in the function being learned. My research involves studying how the local geometry of non-linearities in the learned functions relate to memorization, generalization, biases, and robustness in DNNs. Our work has broad implications across domains, e.g., in interpreting/explaining DNN phenomenon, model auditing, providing provable robustness/safety guarantees and even prediciting downstream behavior of generative models for a given prompt/latent. Apart from this I like thinking about synthetic data training and how it can be used to obtain desired properties in DNN functions.