NeurIPS 2018 Conference Review

The 32nd Conference on Neural Information Processing Systems took place at the convention center of Montreal, Canada, from Dec. 3 to Dec. 8, 2018. This year, the acronym of the conference changed from NIPS to NeurIPS.  The conference received a total of unprecedented 4854 submissions this year among which 1010 were accepted. The acceptance rate was 20.8%. As a top AI conference with focus on multi-track machine learning and computational neuroscience, this year's NeurIPS attracted a huge number of participants from both industry and academia.  8,300 registered attendees also hit a record high. The conference started with a tutorial day followed by the main conference in the next three days. Around 40 parallel workshops were held after the main conference for another two days.

This year's Test of Time Award went to Léon Bottou and Olivier Bousquet for their 2007 NIPS paper "The Tradeoffs of Large Scale Learning".  The paper analyzed distinct trade-off patterns of small-scale and large-scale learning problems and showed that stochastic gradient descent (SGD), which has been dominantly used in deep learning these days, could outperform more complex optimization algorithms under the budget constraints on the number of samples and computation time in large-scale optimization.

More than one thousand conference papers covered a broad variety of AI fields from learning theory and statistics to applications such as computer vision, speech and language.  Reinforcement learning, generative adversarial learning and optimization were among the most active research topics which represented a significant fraction of the presented papers.  Transfer learning, unsupervised learning, causal learning and network interpretability, among others, were also strong topics.

In terms of application areas, the speech community had a descent presence while computer vision was the strongest. There were a number of papers on speech and audio processing including time-domain modeling for realistic music generation, multi-speaker text-to-speech synthesis and on-device automatic speech recognition. Most notably, the paper from Baidu "Neural Voice Cloning with a Few Samples" was selected as a NeurIPS spotlight.  There were also two speech-related workshops -- "Interpretability and Robustness in Audio, Speech, and Language" and "The second Conversational AI workshop – today's practice and tomorrow's potential" -- which brought together researchers from various research communities to share their work and create new ideas in the domains of speech, audio, NLP and dialogue systems.

This year's NeurIPS was a mirror of the current AI world.  Whether the story of the 12-minutes it took registrations to sell out or the amazing demos played at the conference expo, they all marked a vibrating and fast-developing world for AI.  With this excitement, we are looking forward to meeting again next year in Vancouver, Canada.