Drew A. Hudson (original) (raw)

dorarad [at] cs.stanford.edu [scholar] [github] [linkedin] [twitter]

Hi! My name is Drew, I am a Research Scientist at DeepMind, and have recently completed my PhD in Computer Science at Stanford University. I was fortunate to work with my advisor Prof. Christopher D. Manning and collaborate with Dr. Larry Zitnick from FAIR, Meta AI and with Prof. James L. McClelland. I was a member of the Stanford AI Lab and the Stanford NLP group. My research focuses on reasoning, compositionality, and representation learning, at the intersection of vision and language.

I explore structural principles and inductive biases for making neural networks more interpretable, robust and data-efficient, and allow them to generalize effectively and systematically from a few samples only. I believe in the importance of multi-disciplinary both within the AI field and across domains, and draw high-level inspiration from the feats of the human mind, including its structural properties as well as cognitive capabilities.

I believe that compositionality is a key ingredient that, if incorporated successfully into neural models, may help bridging the gap between machine intelligence and natural intelligence. I explore ways to achieve compositionality both in terms of computation and representation.

Towards the former, I introduced, together with my advisor, models such as MAC and the Neural State Machine that perform transparent step-by-step reasoning, as well as the GQA dataset for real-world visual question answering.
Towards the latter, I began more recently to explore ways to learn compositional scene representations, and along with Larry, presented the Generative Adversarial Transformers, for fast, data-efficient and high-resolution image synthesis. I am actively researching this subject further and hope to present new findings on this exciting direction in the near future!

Papers

spotlight_nscl

Compositional Transformers for Scene Generation

We propose a new model for sequential image generation, to explicitly account for differnet objects for enhanced controllability, disentanglement and interpretability.

spotlight_nscl

On the Opportunities and Risks of Foundation Models

Rishi Bommasani Drew A. Hudson,et al. (the Center for Research on Foundation Models)

We thoroughly discuss the emergent paradigm shift of scalable self-supervision, and explore its potential benefits, technical innovations and societal impact.

spotlight_nscl

Generative Adversarial Transformers

Drew A. Hudson,C. Lawrence Zitnick

Spotlight presentation

We introduce the Generative Adversarial Transformer model, a linearly efficient bipartite transformer, and combine it with the GAN framework for high-resolution scene generation.

SLM: Learning a Discourse Language Representation with Sentence Unshuffling

Haejun Lee,Drew A. Hudson,Kangwook Lee, Christopher D. Manning

We introduce a hierarchical transformer that is aware both semantics at the word and the sentence levels, allowing it to acquire better understanding of global properties and discourse relations.

spotlight_nscl

Learning by Abstraction: The Neural State Machine

Drew A. Hudson,Christopher D. Manning

Spotlight presentation, top 3%

We introduce a Neuro-Symbolic model that represent semantic knowledge in the form of scene graph to support iterative reasoning for the task of compositional visual question answering.

spotlight_nscl

GQA: A new Dataset for Real-World Visual Reasoning and Compositional Question Answering

Drew A. Hudson,Christopher D. Manning

Oral Presentation, top 5%

We introduce GQA, a large-scale dataset for real-world visual reasoning and compositional question answering, that focuses on biases reduction and full grounding of each object and entity in a provided scene graph.

spotlight_nscl

Compositional Attention Networks for Machine Reasoning

Drew A. Hudson,Christopher D. Manning

We present the MAC network, a fully differentiable neural network for compositional reasoning, that achieved state-of-the-art 98.9% accuracy on the CLEVR dataset.

spotlight_nscl

Tighter Bounds for Makespan Minimization on Unrelated Machines

Dor Arad,Yael Mordechai, Hadas Shachnai

We obtain tight bounds for the problem of scheduling n jobs to minimize the makespan on m unrelated machines.

Videos

Selected Talks

Activities, Associations & Community

Internships, Work and Awards

Workshops and Conferences

Teaching, Seminars and Mentorship

Hobbies and Fun Facts