Autonomous soaring – AI on the fly
The post Autonomous soaring – AI on the fly appeared first on The AI Blog.
Category: Global
AI of the Storm: Deep Learning Analyzes Atmospheric Events on Saturn
Saturn is many times the size of Earth, so it’s only natural that its storms are more massive — lasting months, covering thousands of miles, and producing lightning bolts thousands of times more powerful.
While scientists have access to galaxies of data on these storms, its sheer volume leaves traditional methods inadequate for studying the planet’s weather systems in their entirety.
Now AI is being used to launch into that trove of information. Researchers from University College London and the University of Arizona are working with data collected by NASA’s Cassini spacecraft, which spent 13 years studying Saturn before disintegrating in the planet’s atmosphere in 2017.
A recently published Nature Astronomy paper describes how the scientists’ deep learning model can reveal previously undetected atmospheric features on Saturn, and provide a clearer view of the planet’s storm systems at a global level.
In addition to providing new insights about Saturn, the AI can shed light on the behavior of planets both within and beyond our solar system.
“We have these missions that go around planets for many years now, and much of this data basically sits in an archive and it’s not being looked at,” said Ingo Waldmann, deputy director of the University College London’s Centre for Space Exoplanet Data. “It’s been difficult so far to look at the bigger picture of this global dataset because people have been analyzing data by hand.”
The researchers used an NVIDIA V100 GPU, the most advanced data center GPU, for both training and inference of their neural networks.
Parting the Clouds
Scientists studying the atmosphere of other planets take one of two strategies, Waldmann says. Either they conduct a detailed manual analysis of a small region of interest, which could take a doctoral student years — or they simplify the data, resulting in rough, low-resolution findings.
“The physics is quite complicated, so the data analysis has been either quite old-fashioned or simplistic,” Waldmann said. “There’s a lot of science one can do by using big data approaches on old problems.”
Thanks to the Cassini satellite, researchers have terabytes of data available to them. Primarily using unsupervised learning, Waldmann and Caitlin Griffith, his co-author from the University of Arizona, trained their deep learning model on data from the satellite’s mapping spectrometer.
This data is commonly collected on planetary missions, Waldmann said, making it easy to apply their AI model to study other planets.
The researchers saw speedups of 30x when training their deep learning models on a single V100 GPU compared to CPU. They’re now transitioning to using clusters of multiple GPUs. For inference, Waldmann said the GPU was around twice as fast as using a CPU.
Using the AI model, the researchers were able to analyze a months-long electrical storm that churned through Saturn’s southern hemisphere in 2008. Scientists had previously detected a bright ammonia cloud from satellite images of the storm — a feature more commonly spotted on Jupiter, but rarely seen on Saturn.
Waldmann and Griffith’s neural network found that the ammonia cloud visible by eye was just the tip of a “massive upwelling” of ammonia hidden under a thin layer of other clouds and gases.
“What you can see by eye is just the strongest bit of that ammonia feature,” Waldmann said. “It’s just the tip of the iceberg, literally. The rest is not visible by eye — but it’s definitely there.”
To Infinity and Beyond
For researchers like Waldmann, these findings are just the first step. Deep learning can provide planetary scientists for the first time with depth and breadth at once, producing detailed analyses that also cover vast geographic regions.
“It will tell you very quickly what the global picture is and how it all connects together,” said Waldmann. “Then researchers can go and look at individual spots that are interesting within a particular system, rather than blindly searching.”
A better understanding of Saturn’s atmosphere can help scientists analyze how our solar system behaves, and provide insights that can be extrapolated to planets around other stars.
Already, the researchers are extending their model to study features on Mars, Venus and Earth using transfer learning — which they were surprised to learn “works really well between planets.”
While Venus and Earth are almost identical in size, Venus has no global plate tectonics. In collaboration with the Observatoire de Paris, the team is starting a project to analyze Venus’s cloud structure and planetary surface to understand why the planet lacks tectonic plates.
Rather than atmospheric features, the researchers’ Mars project focuses on studying the planet’s surface. Data from the Mars Reconaissance Orbiter can create a global analysis that scientists can use to deduce where ancient water was most likely present, and to determine where the next Mars rover should land.
The underlying pattern recognition algorithm can be extended even further, Waldmann said. On Earth, it can be repurposed to spot rogue fishing vessels to preserve protected environments. And across the solar system on Jupiter, a transfer learning approach can train an AI model to analyze how the planet’s storms change over time.
Waldmann says there’s relatively easy access to training data — creating an open field of opportunities for researchers.
“This is the beautiful thing about planetary science,” he said. “All of the data for all of the planets is publicly available.”
Main image, captured in 2011, shows the largest storm observed on Saturn by the Cassini spacecraft. (Image credit NASA/JPL-Caltech/SSI)
The post AI of the Storm: Deep Learning Analyzes Atmospheric Events on Saturn appeared first on The Official NVIDIA Blog.
Microsoft Build 2019: Vision Keynote Highlights
The post Microsoft Build 2019: Vision Keynote Highlights appeared first on The AI Blog.
What’s Microsoft’s vision for conversational AI? Computers that understand you
Today’s intelligent assistants are full of skills. They can check the weather, traffic and sports scores. They can play music, translate words and send text messages. They can even do math, tell jokes and read stories. But, when it comes to conversations that lead somewhere grander, the wheels fall off.
“You have to poke around for magic combinations of words to get various things to happen, and you find out that a lot of the functions that you expect the thing to do, it actually just can’t handle,” said Dan Roth, corporate vice president and former CEO of Semantic Machines, which Microsoft acquired in May 2018.
For example, he explained, systems today can add a new appointment to your calendar but not engage in a back-and-forth dialogue with you about how to juggle a high-priority meeting request. They are also unable to use contextual information from one skill to assist you in making decisions from another, such as checking the weather before scheduling an afternoon meeting on the patio of a nearby coffee shop.
The next generation of intelligent assistant technologies from Microsoft will be able to do this by leveraging breakthroughs in conversational artificial intelligence and machine learning pioneered by Semantic Machines.
The team unveiled its vision for the next leap in natural language interface technology today at Microsoft Build, an annual conference for developers, in Seattle, and announced plans to incorporate this technology into all of its conversational AI products and tools, including Cortana.
Teaching context and concepts
Natural language interfaces are technologies that aim to allow us to communicate with computers in the same way we talk with each other. When natural language interfaces work as Roth and his team envision, our computers will understand us, converse with us and do what we want them to do, much like most people can understand a complex request that requires a few actions.
“Being able to express ourselves in the way we have evolved to communicate and to be able to tie that into all of these really complicated systems without having to know how they work is the promise and vision of natural language interfaces,” said Roth.
The natural language technology in today’s intelligent assistants such as Cortana leverages machine learning to understand the intent of a user’s command. Once that intent is determined, a handwritten program – a skill – is triggered that follows a predetermined set of actions.
For example, the question, “Who won today’s football match between Liverpool and Barcelona?” prompts a sports skill that follows the rules of a pre-coded script to fill in slots for the type of sport, information requested, date and teams. “Will it rain this weekend?” prompts a weather skill and follows pre-scripted rules to get the weekend forecast.
Since the rules for these exchanges are handwritten, developers must anticipate all the ways the skill could be used and write a script to cover each scenario. The inability of humans to script every possible scenario limits the scope and functionality of skills, explained Roth.
The Semantic Machines technology extends the role of the machine learning beyond intents all the way through to enabling what the system does. Instead of a programmer trying to write a skill that plans for every context, the Semantic Machines system learns the functionality for itself from data.
In other words, the Semantic Machines technology learns how to map people’s words to the computational steps needed to carry out requested tasks.
For example, instead of executing a hand-coded program to get the score of the football match, the Semantic Machines approach starts with people who show the system how to get sports scores across a range of example contexts so that the system can learn to fetch sports scores itself.
What’s more, machine learning methods then enable the system to generalize from contexts it has seen to new contexts, learning to do more things in more ways. If it learns how to get sports scores, for example, it can also get weather forecasts and traffic reports. That’s because the system has learned not just a skill, but the concept of how to gather data from a service and present it back to the user.
That’s missing in today’s intelligent assistants, which are programmed to do a list of isolated things that a programmer anticipated. The machine learning in these systems primarily focuses on words that trigger a skill, explained Microsoft technical fellow Dan Klein, a recognized leader in the field of natural language processing and a professor of computer science at the University of California at Berkeley.
“They aren’t focused on learning how to do new things, or mixing and matching the things they already know in order to support new contexts,” said Klein, who was also a co-founder and chief scientist at Semantic Machines.
Dynamic conversation
Since the Semantic Machines system can learn how to do new things, it can more easily engage in a dynamic conversation with a person, accessing and stitching together relevant content, context and concepts from disparate sources to provide answers, present options and produce results.
The Semantic Machines system also has a memory to keep track of the context in a conversation and so-called full duplex capability to talk and listen at the same time in order to keep the dialogue flowing.
“Everything you say is contextualized by what has come before so you can do more complicated things: you can change your mind, you can explore,” said Klein. “Moreover, once things get contextual enough, the notion of a skill begins to dissolve.”
That’s because the notion of skills confines interactions to silos of data whereas true conversation relies on connecting data from all over the place. The Semantic Machines technology orchestrates gathering data and accomplishing tasks on the backend while maintaining a fluid, natural dialogue with the user on the frontend.
Reshuffling your schedule to accommodate a high-priority meeting, for example, requires calendar data and directory data to determine who is free, when, as well as contextually relevant data such as the weather, nearby coffee shops and traffic to figure out where to meet and sit, and when to leave to get there on time.
“Once you start letting things evolve and connect contextually, the notion of a skill is way too limiting,” said Klein. “Getting things done involves mixing and matching.”
Building with natural language
At Build, Microsoft showcased a calendaring application using Semantic Machines technology that can make organizing your day with an intelligent assistant a more fluid, natural and powerful experience. The same technology can be applied to any conversational experience and will eventually power conversations across all of Microsoft’s products and services.
That will build on Cortana’s existing capabilities such as providing answers to questions, offering previews of your day and helping you across your devices from phone to laptop and smart speaker.
Once the technology is incorporated into Cortana, for example, it could make getting things done in Office more about what you need to do and less about accomplishing tasks in certain applications.
“We want it to be less cognitive load, less feeling like I have to go to PowerPoint for this or Word for that, or Outlook for this and Teams for that, and more about personal preferences and intents,” said Andrew Shuman, Microsoft’s corporate vice president for Cortana.
What’s more, added Roth, the technology will be made available through the Microsoft Bot Framework. His team is currently engineering a way for developers working in the framework today to migrate their existing data to the Semantic Machines-powered conversational engine when it is ready.
“As a developer you can start building these experiences yourself,” he said. “We can collectively move, on the basis of this technology, past this notion of skills and silos and simple handwritten programs into the kind of fluid Star Trek-like natural language interfaces we all want.”
Microsoft Build 2019 – Related links to conversational AI:
- Read about Microsoft’s acquisition of Semantic Machines
- Learn more about Semantic Machines
John Roach writes about Microsoft research and innovation. Follow him on Twitter.
The post What’s Microsoft’s vision for conversational AI? Computers that understand you appeared first on The AI Blog.
How AI is making people’s workday more productive
Writing requires a dash of uniquely human creativity. Artificial intelligence alone cannot do it for us, at least not very well. But AI can – and already is – helping us do things like make sure we spell words correctly and use correct grammar, through the myriad ways it is infused across the suite of Microsoft 365 products. Some of them were even used to craft this story.
As the AI in these products is becoming more sophisticated, they are helping us do more than spot a misspelled word.
That includes new intelligent features in Microsoft Word that help us design our documents for maximum readability, along with other features in Microsoft Search and Microsoft Edge that aim to make everybody’s workday more productive. Microsoft showcased these intelligent features today at Microsoft Build, an annual conference for developers, in Seattle.
“Microsoft AI is all about amplifying human ingenuity with intelligent technologies,” said Malavika Rewari, a senior product marketing manager for Microsoft 365.
Microsoft 365 uses AI to help employees overcome some of the realities of modern work, including increasing time demands, overwhelming amounts of data and growing security threats, she noted.
Gathering knowledge
One modern reality of work is age old: a need for knowledge. The difference is that today’s workers turn to the internet to learn, and more than half start with a search engine.
Beginning on May 28, Microsoft Search will move to general availability, the company announced at Build. The technology brings access to the web and work into a single search experience.
Microsoft Search leverages the AI capabilities of Bing and Microsoft Graph, one of the largest collections of data about how people work ever created, enabling workers to find, command, navigate and discover items across their organization’s network of data.
Microsoft Graph includes data from the public internet as well as data available only to employees within an organization such as directories and policy manuals. What’s more, every employee’s graph is distinctive since it contains data that is available only to their specific team, such as documents, and data from their email and calendar.
“We are able to deliver a cohesive search experience that works across any endpoint in Microsoft 365,” said Bill Baer, a senior product marketing manager on the Microsoft Search team. “Whether you are searching in Bing or searching in the Windows 10 search bar, you’ll get a set of contextually relevant results.”
New intelligence in Microsoft Search includes a machine reading comprehension capability that can extract a paragraph from documents explicitly related to your question. For example, if an employee asks, “Can I bring my dog to work?” Microsoft Search will extract the relevant paragraph from the human resources manual and present it as a search result.
“It understands the question you are asking, and then it can find the answer within millions of words of text and give it to you in context,” said Baer.
Another new intelligent feature allows people within a company to conduct people searches with incomplete information. For example, consider being told, “Talk to Pat on the third floor,” and not knowing who Pat is. A search on “Pat, floor 3” uses intelligence from Microsoft Graph such as your immediate team and location to return the most likely Pat, including an office number and picture.
Working on Microsoft Edge
Microsoft, which recently announced plans to adopt the Chromium open source project in the development of Microsoft Edge on the desktop, also is working on ways to make the Edge browser a more natural extension of the Microsoft Search experience, noted Baer. That means users who are signed in to a Microsoft 365 account will be able to see related results within the Edge browser.
The Microsoft Edge team is also experimenting with a feature called Collections that allows users to compile and organize content as they browse the internet in their open browser window and intelligently share the compiled content via email or export it to Excel or Word.
For example, a person shopping for a new camera could visit several product websites and save each page in the Collections pane on the side of the browser. The underlying machine learning in Collections would intelligently display an image of each model along with relevant metadata such as price, user rating and the website where the data originated.
From there, a user could email the list to a friend, or copy and paste the collection elsewhere, maintaining the clean format of the content. Another option is to export to Excel, where the machine learning automatically populates a table organized with columns for brand, model, price, rating and so on based on the collected metadata.
“You can easily, at a glance, get the value and make your decision more quickly,” said Divya Kumar, group product marketing manager for Microsoft Edge. She added that the team is experimenting with similar functionality for exporting to Word, including the ability to compile a document with information such as images and text collected from several websites, citations included.
Better Word documents
Beginning this fall, people working in Word Online who are in search of inspiration and insights on how to make their document better will be able to receive intelligent suggestions with Ideas – a feature that is already making people more productive in PowerPoint and Excel.
The Ideas in Word feature uses machine learning and intelligence from Microsoft Graph to help users write polished prose, create more professional documents and efficiently navigate documents created by others.
For example, feedback and signals from Microsoft Graph indicate that workers generally ignore tools available in Word to structure their documents, such as section heads, but rather manually make some words bold and bigger to indicate a new section.
“Here’s something where we say, ‘Hey, we understand the structure of your document. We can make it navigable, or we could create a table of contents on your behalf,’” explained Kirk Gregersen, a partner director of program management in Microsoft’s Experiences and Devices group.
Other intelligent suggestions include recommended acronyms based on their usage in Microsoft Graph, calculated average time to read the document, highlight extraction, as well as familiar fixes for spelling and grammatical errors and advice on more concise and inclusive language such as “police officer” instead of “policeman.”
Neural rewrites
A recently available intelligent feature in Word is rewrite suggestions, which brings the power of deep learning to offer suggestions on different ways to write a phrase.
The technology builds on enhancements to the popular synonyms feature in Word that use machine learning to understand the context of the sentence the word appears in to offer alternative word choices that are more relevant.
“You don’t need to search online to find an alternative way to express a phrase,” said Zhang Li, a senior program manager in the Microsoft Office team, explaining that the intelligence service will surface suggestions within the document.
His team used similar technology to improve synonym ranking earlier this year, leading the synonym suggestion acceptance rate to double.
“We want to augment your skills,” said Rewari, the senior product marketing manager for Microsoft 365. “We want to help you communicate more efficiently, effectively and inclusively.”
Top video: The Ideas in Word feature uses machine learning and intelligence from Microsoft Graph to help a user style a table for a professional document.
Microsoft Build 2019 – Related links to Microsoft 365 and AI
- Read: New, people centered experiences in Microsoft 365, the world’s productivity cloud
- Check out Microsoft Search and Microsoft Graph
- Read: Microsoft creates AI that can read a document and answer questions about it as well as a person
- Learn more about Microsoft Edge on Chromium
- Check out Word Online and enroll in the Office Insider program for early access to Ideas in Word
John Roach writes about Microsoft research and innovation. Follow him on Twitter.
The post How AI is making people’s workday more productive appeared first on The AI Blog.
Google at ICLR 2019
This week, New Orleans, LA hosts the 7th International Conference on Learning Representations (ICLR 2019), a conference focused on how one can learn meaningful and useful representations of data for machine learning. ICLR offers conference and workshop tracks, both of which include invited talks along with oral and poster presentations of some of the latest research on deep learning, metric learning, kernel learning, compositional models, non-linear structured prediction and issues regarding non-convex optimization.
At the forefront of innovation in neural networks and deep learning, Google focuses on on both theory and application, developing learning approaches to understand and generalize. As Platinum Sponsor of ICLR 2019, Google will have a strong presence with over 200 researchers attending, contributing to and learning from the broader academic research community by presenting papers and posters, in addition to participating on organizing committees and in workshops.
If you are attending ICLR 2019, we hope you’ll stop by our booth and chat with our researchers about the projects and opportunities at Google that go into solving interesting problems for billions of people. You can also learn more about our research being presented at ICLR 2019 in the list below (Googlers highlighted in blue).
Officers and Board Members
Hugo Larochelle, Samy Bengio, Tara Sainath
General Chair
Tara Sainath
Workshop Chairs
Been Kim, Graham Taylor
Program Committee includes:
Chelsea Finn, Dale Schuurmans, Dumitru Erhan, Katherine Heller, Lihong Li, Samy Bengio, Rohit Prabhavalkar, Alex Wiltschko, Slav Petrov, George Dahl
Oral Contributions
Generating High Fidelity Images with Subscale Pixel Networks and Multidimensional Upscaling
Jacob Menick, Nal Kalchbrenner
Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset
Curtis Hawthorne, Andrew Stasyuk, Adam Roberts, Ian Simon, Anna Huang, Sander Dieleman, Erich Elsen, Jesse Engel, Douglas Eck
Meta-Learning Update Rules for Unsupervised Representation Learning
Luke Metz, Niru Maheswaranathan, Brian Cheung, Jascha Sohl-Dickstein
Posters
A Data-Driven and Distributed Approach to Sparse Signal Representation and Recovery
Ali Mousavi, Gautam Dasarathy, Richard G. Baraniuk
Bayesian Deep Convolutional Networks with Many Channels are Gaussian Processes
Roman Novak, Lechao Xiao, Yasaman Bahri, Jaehoon Lee, Greg Yang, Jiri Hron, Daniel A. Abolafia, Jeffrey Pennington, Jascha Sohl-Dickstein
Diversity-Sensitive Conditional Generative Adversarial Networks
Dingdong Yang, Seunghoon Hong, Yunseok Jang, Tianchen Zhao, Honglak Lee
Diversity and Depth in Per-Example Routing Models
Prajit Ramachandran, Quoc V. Le
Eidetic 3D LSTM: A Model for Video Prediction and Beyond
Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, Li Fei-Fei
GANSynth: Adversarial Neural Audio Synthesis
Jesse Engel, Kumar Krishna Agrawal, Shuo Chen, Ishaan Gulrajani, Chris Donahue, Adam Roberts
K for the Price of 1: Parameter-efficient Multi-task and Transfer Learning
Pramod Kaushik Mudrakarta, Mark Sandler, Andrey Zhmoginov, Andrew Howard
Learning to Describe Scenes with Programs
Yunchao Liu, Zheng Wu, Daniel Ritchie, William Freeman, Joshua B Tenenbaum, Jiajun Wu
Learning to Infer and Execute 3D Shape Programs
Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William Freeman, Joshua B Tenenbaum, Jiajun Wu
The Singular Values of Convolutional Layers
Hanie Sedghi, Vineet Gupta, Philip M. Long
Unsupervised Discovery of Parts, Structure, and Dynamics
Zhenjia Xu, Zhijian Liu, Chen Sun, Kevin Murphy, William Freeman, Joshua B Tenenbaum, Jiajun Wu
Adversarial Reprogramming of Neural Networks
Gamaleldin Elsayed, Ian Goodfellow (no longer at Google), Jascha Sohl-Dickstein
Discriminator Rejection Sampling
Ian Goodfellow (no longer at Google), Jascha Sohl-Dickstein
On Self Modulation for Generative Adversarial Networks
Ting Chen, Mario Lucic, Neil Houlsby, Sylvain Gelly
Towards GAN Benchmarks Which Require Generalization
Ishaan Gulrajani, Colin Raffel, Luke Metz
Understanding and Improving Interpolation in Autoencoders via an Adversarial Regularizer
David Berthelot, Colin Raffel, Aurko Roy, Ian Goodfellow (no longer at Google)
A new dog learns old tricks: RL finds classic optimization algorithms
Weiwei Kong, Christopher Liaw, Aranyak Mehta, D. Sivakumar
Contingency-Aware Exploration in Reinforcement Learning
Jongwook Choi, Yijie Guo, Marcin Moczulski, Junhyuk Oh, Neal Wu, Mohammad Norouzi, Honglak Lee
Discriminator-Actor-Critic: Addressing Sample Inefficiency and Reward Bias in Adversarial Imitation Learning
Ilya Kostrikov, Kumar Krishna Agrawal, Debidatta Dwibedi, Sergey Levine, Jonathan Tompson
Diversity is All You Need: Learning Skills without a Reward Function
Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, Sergey Levine
Episodic Curiosity through Reachability
Nikolay Savinov, Anton Raichuk, Raphael Marinier, Damien Vincent, Marc Pollefeys, Timothy Lillicrap, Sylvain Gelly
Learning to Navigate the Web
Izzeddin Gur, Ulrich Rueckert, Aleksandra Faust, Dilek Hakkani-Tur
Meta-Learning Probabilistic Inference for Prediction
Jonathan Gordon, John Bronskill, Matthias Bauer, Sebastian Nowozin, Richard E. Turner
Multi-step Retriever-Reader Interaction for Scalable Open-domain Question Answering
Rajarshi Das, Shehzaad Dhuliawala, Manzil Zaheer, Andrew McCallum
Near-Optimal Representation Learning for Hierarchical Reinforcement Learning
Ofir Nachum, Shixiang Gu, Honglak Lee, Sergey Levine
Neural Logic Machines
Honghua Dong, Jiayuan Mao, Tian Lin, Chong Wang, Lihong Li, Dengyong Zhou
Neural Program Repair by Jointly Learning to Localize and Repair
Marko Vasic, Aditya Kanade, Petros Maniatis, David Bieber, Rishabh Singh
Optimal Completion Distillation for Sequence Learning
Sara Sabour, William Chan, Mohammad Norouzi
Recall Traces: Backtracking Models for Efficient Reinforcement Learning
Anirudh Goyal, Philemon Brakel, William Fedus, Soumye Singhal, Timothy Lillicrap, Sergey Levine, Hugo Larochelle, Yoshua Bengio
Sample Efficient Adaptive Text-to-Speech
Yutian Chen, Yannis M Assael, Brendan Shillingford, David Budden, Scott Reed, Heiga Zen, Quan Wang, Luis C. Cobo, Andrew Trask, Ben Laurie, Caglar Gulcehre, Aaron van den Oord, Oriol Vinyals, Nando de Freitas
Synthetic Datasets for Neural Program Synthesis
Richard Shin, Neel Kant, Kavi Gupta, Chris Bender, Brandon Trabucco, Rishabh Singh, Dawn Song
The Laplacian in RL: Learning Representations with Efficient Approximations
Yifan Wu, George Tucker, Ofir Nachum
A Mean Field Theory of Batch Normalization
Greg Yang, Jeffrey Pennington, Vinay Rao, Jascha Sohl-Dickstein, Samuel S Schoenholz
Efficient Training on Very Large Corpora via Gramian Estimation
Walid Krichene, Nicolas Mayoraz, Steffen Rendle, Li Zhang, Xinyang Yi, Lichan Hong, Ed Chi, John Anderson
Predicting the Generalization Gap in Deep Networks with Margin Distributions
Yiding Jiang, Dilip Krishnan, Hossein Mobahi, Samy Bengio
InfoBot: Transfer and Exploration via the Information Bottleneck
Anirudh Goyal, Riashat Islam, DJ Strouse, Zafarali Ahmed, Hugo Larochelle, Matthew Botvinick, Sergey Levine, Yoshua Bengio
AntisymmetricRNN: A Dynamical System View on Recurrent Neural Networks
Bo Chang, Minmin Chen, Eldad Haber, Ed H. Chi
Complement Objective Training
Hao-Yun Chen, Pei-Hsin Wang, Chun-Hao Liu, Shih-Chieh Chang, Jia-Yu Pan, Yu-Ting Chen, Wei Wei, Da-Cheng Juan
DOM-Q-NET: Grounded RL on Structured Language
Sheng Jia, Jamie Kiros, Jimmy Ba
From Language to Goals: Inverse Reinforcement Learning for Vision-Based Instruction Following
Justin Fu, Anoop Korattikara Balan, Sergey Levine, Sergio Guadarrama
Harmonic Unpaired Image-to-image Translation
Rui Zhang, Tomas Pfister, Li-Jia Li
Hierarchical Generative Modeling for Controllable Speech Synthesis
Wei-Ning Hsu, Yu Zhang, Ron Weiss, Heiga Zen, Yonghui Wu, Yuxuan Wang, Yuan Cao, Ye Jia, Zhifeng Chen, Jonathan Shen, Patrick Nguyen, Ruoming Pang
Learning Finite State Representations of Recurrent Policy Networks
Anurag Koul, Alan Fern, Samuel Greydanus
Learning to Screen for Fast Softmax Inference on Large Vocabulary Neural Networks
Patrick Chen, Si Si, Sanjiv Kumar, Yang Li, Cho-Jui Hsieh
Music Transformer: Generating Music with Long-Term Structure
Chen-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Ian Simon, Curtis Hawthorne, Noam Shazeer, Andrew Dai, Matthew D Hoffman, Monica Dinculescu, Douglas Eck
Universal Transformers
Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, Lukasz Kaiser
What do you learn from context? Probing for sentence structure in contextualized word representations
Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, Tom McCoy, Najoung Kim, Benjamin Van Durme, Samuel R. Bowman, Dipanjan Das, Ellie Pavlick
Doubly Reparameterized Gradient Estimators for Monte Carlo Objectives
George Tucker, Dieterich Lawson, Shixiang Gu, Chris J. Maddison
How Important Is a Neuron?
Kedar Dhamdhere, Mukund Sundararajan, Qiqi Yan
Integer Networks for Data Compression with Latent-Variable Models
Johannes Ballé, Nick Johnston, David Minnen
Modeling Uncertainty with Hedged Instance Embeddings
Seong Joon Oh, Andrew Gallagher, Kevin Murphy, Florian Schroff, Jiyan Pan, Joseph Roth
Preventing Posterior Collapse with delta-VAEs
Ali Razavi, Aaron van den Oord, Ben Poole, Oriol Vinyals
Spectral Inference Networks: Unifying Deep and Spectral Learning
David Pfau, Stig Petersen, Ashish Agarwal, David GT Barrett, Kimberly L Stachenfeld
Spreading vectors for similarity search
Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, Hervé Jégou
Stochastic Prediction of Multi-Agent Interactions from Partial Observations
Chen Sun, Per Karlsson, Jiajun Wu, Joshua B Tenenbaum, Kevin Murphy
Workshops
Learning from Limited Labeled Data
Sponsored by Google
Deep Reinforcement Learning Meets Structured Prediction
Organizing Committee includes: Chen Liang
Invited Speaker: Mohammad Norouzi
Debugging Machine Learning Models
Organizing Committee includes: D. Sculley
Invited Speaker: Dan Moldovan
Structure & Priors in Reinforcement Learning (SPiRL)
Organizing Committee includes: Chelsea Finn
Task-Agnostic Reinforcement Learning (TARL)
Sponsored by Google
Organizing Committee includes: Danijar Hafner, Marc G. Bellemare
Invited Speaker: Chelsea Finn
AI for Social Good
Program Committee includes: Ernest Mwebaze
Safe Machine Learning Specification, Robustness and Assurance
Program Committee includes: Nicholas Carlini
Representation Learning on Graphs and Manifolds
Program Committee includes: Bryan Perozzi
