Inside NIPS 2016: Highlights, Papers, and Insights from Hulu’s Researchers

Conference Overview

Machine learning has two premier conferences: ICML (International Conference on Machine Learning) and NIPS (Neural Information Processing Systems). In 2016 NIPS broke tradition by moving from ski‑resort locations in North America to Barcelona, Spain, attracting over 6,000 attendees.

Hulu sent two researchers, senior researcher Tang Bangsheng and researcher Zheng Yin, to represent the company. Hulu also had two papers accepted at ICML and contributed reviewers for NIPS, including Zheng Yin who reviewed three submissions on recommender systems.

Key Statistics

More than 2,500 papers were submitted to NIPS 2016, with 568 accepted (45 oral presentations). A total of 3,242 reviewers evaluated the papers, each receiving roughly six reviews to reduce scoring bias.

Keynote and Tutorial Highlights

Yann LeCun’s keynote used a cake analogy to compare unsupervised learning (the cake body), supervised learning (the frosting), and reinforcement learning (the cherry). He emphasized the formula "Intelligence & Common Sense = Perception + Predictive Model + Memory + Reasoning & Planning" and discussed recent advances such as learning physics, entity RNNs, memory‑augmented networks, GANs, and video prediction.

The tutorial program featured leading experts:

David Blei on variational inference, covering mean‑field methods, variational auto‑encoders, and variance‑reduction techniques.

Andrew Ng on building deep‑learning AI systems for industry, covering dataset construction, evaluation metrics, and practical engineering considerations.

Ian Goodfellow on generative adversarial networks (GANs), including his paper "Improved Techniques for Training GANs" and the state‑of‑the‑art Plug & Play Generative Networks.

Award‑Winning Papers

The best paper award went to "Value Iteration Networks (VIN)", which integrates differentiable value‑iteration into neural networks for learning optimal policies.

The best student paper was "Matrix Completion has no Spurious Local Minima" by Rong Ge, Jason Lee, and Tengyu Ma, proving that under certain conditions all local minima of matrix factorization are global, a result with potential implications for deep learning optimization.

Invited Talks and Symposiums

DeepMind’s Drew Purves presented on the AI ecosystem, while Boston Dynamics showcased its latest robots, noting that current robot control still relies on pre‑programmed algorithms rather than learning.

Three symposiums were held: Deep Learning, Machine Learning & the Law, and Recurrent Neural Networks (which swapped venues due to high attendance). Hulu researchers attended the Deep Learning symposium, where Yoshua Bengio highlighted recent breakthroughs such as Real NVP, PixelCNN, WaveNet, InfoGAN, stochastic depth, layer normalization, seq2seq beam‑search optimization, counterfactual inference, and guided cost learning.

Conclusion

NIPS 2016 was the largest machine learning conference to date, reflecting the rapid growth of AI and deep learning. Breakthroughs like AlphaGo, generative models (PixelRNN/CNN, WaveNet), and the shift toward AI‑first strategies underscore the field’s momentum. Hulu continues to invest heavily in AI research across recommendation, advertising, computer vision, and video compression.