Hi I am Philip Huebner,
I am currently a PhD candidate at the UIUC Learning & Language Lab.
My research focuses on understanding how intelligent systems learn from and represent complex and evolving information in the context of large-scale and noisy environments.
I am particularly interested in interdisciplinary work at the intersection of language acquisition, and natural language processing. With one foot in each field, I seek productive exchanges and cross-pollination between researchers on both sides, by exchanging tools and ideas.
Recent advances in our ability to study large, naturalistic datasets, combined with advanced computational modeling techniques, have allowed us to ask ever bigger and more ambitious questions about the nature and statistics of children's language environment. One of the major insights from investigating the structure of experience is that it has forced us to radically re-evaluate traditional theories of learning and representation. Many models previously deemed insufficient (especially deep learning systems with non-linear transformations) perform qualitatively differently when faced with large amounts of data, often enabling them to solve complex tasks without being explicitly told what features to pay attention to.
Despite the success of "big data" approaches, the fast pace towards larger and more complex models, requiring more advanced hardware, model training time, and parameter tuning, there is an enormous need for natural language technology to incorporate inductive biases inspired by how humans learn to perform similar tasks with fewer computational resources and with less data. With this in mind, my work seeks to integrate deep-learning approaches based on "big data" with (human) inductive biases, to improve both the performance and efficiency of contemporary deep-learning models.