It’s this week in machine learning
The list of papers accepted for this year’s International Conference on Learning Representations (ICLR) are out and they make interesting reading.
The conference will take place in Vancouver from the April 30 – May 3. It’ll showcase the latest trends and developments in AI and machine learning. It looks like General Adversarial Networks, often used for vision tasks, are still in fashion and have progressed enough to keep people’s interest.
There’s also a few adversarial learning papers, some reinforcement learning ones, natural language stuff, and complicated theoretical evaluating neural networks.
The event is split into oral presentations, poster sessions, and workshop tracks. The full list of papers that have been accepted, rejected, and withdrawn can be found here.
The Register will be combing through to try and pick any interesting ones, but if there are any you think are particularly exciting let us know.
New requests for research
OpenAI has released a list of seven unsolved problems for AI whizzes to have a crack at.
They’re all mostly related to the field of reinforcement learning. One involves training an algorithm to play a version of the hugely popular mobile phone game of snake, but with multiple snakes instead of one. Some are more technical such as investigating regularization in reinforcement learning to improve optimization and overfitting in algorithms.
It’s the second time OpenAI have published its requests for research. The first time resulted in a bunch of new research papers.
“We expect these problems to be a fun and meaningful way for new people to enter the field, as well as for practitioners to hone their skills (it’s also a great way to get a job at OpenAI). Many will require inventing new ideas,” it said in a blog post.
See the full list of problems here.
Machine learning for flight delays
Google’s Flights app can now, apparently, predict when a flight will be delayed before it’s formally announced by an airline.
The new feature only works for flights on American, Delta, and United airlines so far. “Using historic flight status data, our machine learning algorithms can predict some delays even when this information isn’t available from airlines yet—and delays are only flagged when we’re at least 80% confident in the prediction. We still recommend getting to the airport with enough time to spare, but hope this information can manage expectations and prevent surprises,” it said in a blog post.
Well, since Google hasn’t really bothered to explain how the technology works, we just have to take their word and it’s probably not very wise to rely on it. It obviously uses a bunch of data points taken from past flight statuses to predict common patterns that lead to delays, but we aren’t sure of any other details beyond that.
We did ask Google for comment, but as usual we didn’t really get anything useful back.
IBM Cloud joins the Volta party
It’s better late than never, but IBM Cloud announced it also now has Nvidia Volta GPU spot instances on its cloud service.
“Starting today, you can equip individual IBM Cloud bare metal servers with up to two NVIDIA Tesla V100 PCIe GPU accelerators — NVIDIA’s latest, fastest and most advanced GPU architecture,” it said in a blog post.
It also apparently “offers the industry’s only CPU-to-GPU Nvidia NVLink connection on [its] latest POWER9 servers”.
It’s geared towards attracting researchers and AI startups onto its cloud to carry out the heavy, number crunching workloads for high performance computing and machine learning.
Andrew Ng, a prominent AI spokesperson and ex-chief scientist Baidu, announced a new fund to support AI startups that “improve human life”.
The $175m (£124m) grant has been pooled together by investors from NEA, Sequoia, Greylock Partners, the SoftBank Group and others.
Ng announced he would be leading the AI Fund as its general partner. Eva Wang will be the partner and COO alongside Steven Syverud, who will be a partner.
Part of the money will also go into Landing.AI, a startup started by Ng focused on using AI in the manufacturing industry.
“In the early days of electricity, much of the innovation centered around slightly different improvements in lighting. While this was an important foundation, the really transformative applications, in which electric power spurred massive redesigns in multiple industries, took longer to be grasped. AI is the new electricity, and is at a similar inflection point,” he said in a Medium post. ®