Category Archives: Machine Learning

Google researchers recently published a paper describing a framework SEED RL that scales AI model training to thousands of machines. They say that it could facilitate training at millions of frames per second on a machine while reducing costs by up to 80%, potentially leveling the playing field for startups that couldnt previously compete with large AI labs.

Training sophisticated machine learning models in the cloud remains prohibitively expensive. According to a recent Synced report, the University of Washingtons Grover, which is tailored for both the generation and detection of fake news, cost $25,000 to train over the course of two weeks. OpenAI racked up $256 per hour to train its GPT-2 language model, and Google spent an estimated $6,912 training BERT, a bidirectional transformer model that redefined the state of the art for 11 natural language processing tasks.

SEED RL, which is based on Googles TensorFlow 2.0 framework, features an architecture that takes advantage of graphics cards and tensor processing units (TPUs) by centralizing model inference. To avoid data transfer bottlenecks, it performs AI inference centrally with a learner component that trains the model using input from distributed inference. The target models variables and state information are kept local, while observations are sent to the learner at every environment step and latency is kept to a minimum thanks to a network library based on the open source universal RPC framework.

SEED RLs learner component can be scaled across thousands of cores (e.g., up to 2,048 on Cloud TPUs), and the number of actors which iterate between taking steps in the environment and running inference on the model to predict the next action can scale up to thousands of machines. One algorithm V-trace predicts an action distribution from which an action can be sampled, while another R2D2 selects an action based on the predicted future value of that action.

To evaluate SEED RL, the research team benchmarked it on the commonly used Arcade Learning Environment, several DeepMind Lab environments, and the Google Research Football environment. They say that they managed to solve a previously unsolved Google Research Football task and that they achieved 2.4 million frames per second with 64 Cloud TPU cores, representing an improvement over the previous state-of-the-art distributed agent of 80 times.

This results in a significant speed-up in wall-clock time and, because accelerators are orders of magnitude cheaper per operation than CPUs, the cost of experiments is reduced drastically, wrote the coauthors of the paper. We believe SEED RL, and the results presented, demonstrate that reinforcement learning has once again caught up with the rest of the deep learning field in terms of taking advantage of accelerators.

Read more:
Google open-sources framework that reduces AI training costs by up to 80% - VentureBeat

Reader survey We hear a lot these days about IT automation. Yet whether it's labelled intelligent infrastructure, AIOps, self-driving IT, or even private cloud, the aim is the same.

And that aim is: to use the likes of machine learning, workflow automation, and infrastructure-as-code to automatically make changes in real-time, eliminating as much as possible of the manual drudgery associated with routine IT administration.

Are the latest AI/ML-powered intelligent automation solutions trustworthy and ready for mainstream deployment, particularly in areas such as storage management?

Should we go ahead and implement the technology now on offer?

This controversial topic is the subject of our latest reader survey, and we are eager to hear your views.

Please complete our short survey, here.

As always, your responses will be anonymous and your privacy assured.

Sponsored: Webcast: Why you need managed detection and response

The rest is here:
Are machine-learning-based automation tools good enough for storage management and other areas of IT? Let us know - The Register

In a briefing on Monday, research leaders across tech, academia and the government joined the White House to announce an open data set full of scientific literature on the novel coronavirus. The COVID-19 Open Research Dataset, known as CORD-19, will also add relevant new research moving forward, compiling it into one centralized hub. The new data set is machine readable, making it easily parsed for machine learning purposes a key advantage according to researchers involved in the ambitious project.

In a press conference, U.S. CTO Michael Kratsios called the new data set the most extensive collection of machine readable coronavirus literature to date. Kratsios characterized the project as a call to action for the AI community, which can employ machine learning techniques to surface unique insights in the body of data. To come up with guidance for researchers combing through the data, the National Academies of Sciences, Engineering, and Medicine collaborated with the World Health Organization to come up with high priority questions about the coronavirus related to genetics, incubation, treatment, symptoms and prevention.

The partnership, announced today by the White House Office of Science and Technology Policy, brings together the Chan Zuckerberg Initiative, Microsoft Research, the Allen Institute for Artificial Intelligence, the National Institutes of Healths National Library of Medicine, Georgetown Universitys Center for Security and Emerging Technology, Cold Spring Harbor Laboratory and the Kaggle AI platform, owned by Google.

The database brings together nearly 30,000 scientific articles about the virus known as SARS-CoV-2. as well as related viruses in the broader coronavirus group. Around half of those articles make the full text available. Critically, the database will include pre-publication research from resources like medRxiv and bioRxiv, open access archives for pre-print health sciences and biology research.

Sharing vital information across scientific and medical communities is key to accelerating our ability to respond to the coronavirus pandemic, Chan Zuckerberg Initiative Head of Science Cori Bargmann said of the project.

The Chan Zuckerberg Initiative hopes that the global machine learning community will be able to help the science community connect the dots on some of the enduring mysteries about the novel coronavirus as scientists pursue knowledge around prevention, treatment and a vaccine.

For updates to the CORD-19 data set, the Chan Zuckerberg Initiative will track new research on a dedicated page on Meta, the research search engine the organization acquired in 2017.

The CORD-19 data set announcement is certain to roll out more smoothly than the White Houses last attempt at a coronavirus-related partnership with the tech industry. The White House came under criticism last week for President Trumps announcement that Google would build a dedicated website for COVID-19 screening. In fact, the site was in development by Verily, Alphabets life science research group, and intended to serve California residents, beginning with San Mateo and Santa Clara County. (Alphabet is the parent company of Google.)

The site, now live, offers risk screening through an online questionnaire to direct high-risk individuals toward local mobile testing sites. At this time, the project has no plans for a nationwide rollout.

Google later clarified that the company is undertaking its own efforts to bring crucial COVID-19 information to users across its products, but that may have become conflated with Verilys much more limited screening site rollout. On Twitter, Googles comms team noted that Google is indeed working with the government on a website, but not one intended to screen potential COVID-19 patients or refer them to local testing sites.

In a partial clarification over the weekend, Vice President Pence, one of the Trump administrations designated point people on the pandemic, indicated that the White House is working with Google but also working with many other tech companies. Its not clear if that means a central site will indeed launch soon out of a White House collaboration with Silicon Valley, but Pence hinted that might be the case. If that centralized site will handle screening and testing location referral is not clear.

Our best estimate is that some point early in the week we will have a website that goes up, Pence said.

More:
With launch of COVID-19 data hub, the White House issues a call to action for AI researchers - TechCrunch

What Can Machine Learning Tell Us About WR Prospects?

One of my favorite parts of draft season is trying to model the incoming prospects. This year, I wanted to try something new, so I dove into the world of machine learning models. Using machine learning to detail the value of a WR prospect is very useful for dynasty fantasy football.

Machine learning leverages artificial intelligence to identify patterns (learn) from the data, and build an appropriate model. I took over 60 different variables and 366 receiving prospects between the 2004 and 2016 NFL Drafts, and let the machine do its thing. As with any machine, some human intervention is necessary, and I fine-tuned everything down to a 24-model ensemble built upon different logistic regressions.

Just like before, the model presents the likelihood of a WR hitting 200 or more PPR points in at least one of his first three seasons. Here are the nine different components featured, in order of significance:

This obviously represents a massive change from the original model, proving once again that machines are smarter than humans. I decided to move over to ESPN grades and ranks instead of NFL Draft Scout for a few reasons:

Those changes alone made strong improvements to the model, and it should be noted that the ESPN overall ranks have been very closely tied to actual NFL Draft position.

Having an idea of draft position will always help a model since draft position usually begets a bunch of opportunity at the NFL level.

Since the model is built on drafts up until 2016, I figured perhaps youd want to see the results from the last three drafts before seeing the 2020 outputs.

It is encouraging to see some hits towards the top of the model, but there are obviously some misses as well. Your biggest takeaway here should be just how difficult it is to hit that 200 point threshold. Only two prospects the last three years have even a 40% chance of success. The model is telling us not to be over-confident, and that is a good thing.

Now that youve already seen some results, here are the 2020 model outputs.

Tee Higgins as the top WR is likely surprising for a lot of people, but it shouldnt be. Higgins had a fantastic career at Clemson, arguably the best school in the country over the course of his career. He is a proven touchdown scorer, and is just over 21 years old with a prototypical body-type.

Nobody is surprised that the second WR on this list is from Alabama, but they are likely shocked to see that a data-based model has Henry Ruggs over Jerry Jeudy. The pair is honestly a lot closer that many people think in a lot of the peripheral statistics. The major edge for Ruggs comes on the ground. He had a 75 yard rushing touchdown, which really underlines his special athleticism and play-making ability.

The name that likely stands out the most is Geraud Sanders, who comes in ahead of Jerry Jeudy despite being a relative unknown out of Air Force. You can mentally bump him down a good bit. The academy schools are a bit of a glitch in the system, as their offensive approach usually yields some outrageous efficiency. Since 2015, 12 of the top 15 seasons in adjusted receiving yards per pass attempt came from either an academy school or Georgia Techs triple-option attack. Sanders isnt a total zero, his profile looks very impressive, but I would have him closer to a 10% chance of success given his likely Day 3 or undrafted outcome in the NFL Draft.

Read more here:
The Top Machine Learning WR Prospect Will Surprise You - RotoExperts

The work of MIT computer scientist Aleksander Madry is fueled by one core mission: doing machine learning the right way.

Madrys research centers largely on making machine learning a type of artificial intelligence more accurate, efficient, and robust against errors. In his classroom and beyond, he also worries about questions of ethical computing, as we approach an age where artificial intelligence will have great impact on many sectors of society.

I want society to truly embrace machine learning, says Madry, a recently tenured professor in the Department of Electrical Engineering and Computer Science. To do that, we need to figure out how to train models that people can use safely, reliably, and in a way that they understand.

Interestingly, his work with machine learning dates back only a couple of years, to shortly after he joined MIT in 2015. In that time, his research group has published several critical papers demonstrating that certain models can be easily tricked to produce inaccurate results and showing how to make them more robust.

In the end, he aims to make each models decisions more interpretable by humans, so researchers can peer inside to see where things went awry. At the same time, he wants to enable nonexperts to deploy the improved models in the real world for, say, helping diagnose disease or control driverless cars.

Its not just about trying to crack open the machine-learning black box. I want to open it up, see how it works, and pack it back up, so people can use it without needing to understand whats going on inside, he says.

For the love of algorithms

Madry was born in Wroclaw, Poland, where he attended the University of Wroclaw as an undergraduate in the mid-2000s. While he harbored interest in computer science and physics, I actually never thought Id become a scientist, he says.

An avid video gamer, Madry initially enrolled in the computer science program with intentions of programming his own games. But in joining friends in a few classes in theoretical computer science and, in particular, theory of algorithms, he fell in love with the material. Algorithm theory aims to find efficient optimization procedures for solving computational problems, which requires tackling difficult mathematical questions. I realized I enjoy thinking deeply about something and trying to figure it out, says Madry, who wound up double-majoring in physics and computer science.

When it came to delving deeper into algorithms in graduate school, he went to his first choice: MIT. Here, he worked under both Michel X. Goemans, who was a major figure in applied math and algorithm optimization, and Jonathan A. Kelner, who had just arrived to MIT as a junior faculty working in that field. For his PhD dissertation, Madry developed algorithms that solved a number of longstanding problems in graph algorithms, earning the 2011 George M. Sprowls Doctoral Dissertation Award for the best MIT doctoral thesis in computer science.

After his PhD, Madry spent a year as a postdoc at Microsoft Research New England, before teaching for three years at the Swiss Federal Institute of Technology Lausanne which Madry calls the Swiss version of MIT. But his alma mater kept calling him back: MIT has the thrilling energy I was missing. Its in my DNA.

Getting adversarial

Shortly after joining MIT, Madry found himself swept up in a novel science: machine learning. In particular, he focused on understanding the re-emerging paradigm of deep learning. Thats an artificial-intelligence application that uses multiple computing layers to extract high-level features from raw input such as using pixel-level data to classify images. MITs campus was, at the time, buzzing with new innovations in the domain.

But that begged the question: Was machine learning all hype or solid science? It seemed to work, but no one actually understood how and why, Madry says.

Answering that question set his group on a long journey, running experiment after experiment on deep-learning models to understand the underlying principles. A major milestone in this journey was an influential paper they published in 2018, developing a methodology for making machine-learning models more resistant to adversarial examples. Adversarial examples are slight perturbations to input data that are imperceptible to humans such as changing the color of one pixel in an image but cause a model to make inaccurate predictions. They illuminate a major shortcoming of existing machine-learning tools.

Continuing this line of work, Madrys group showed that the existence of these mysterious adversarial examples may contribute to how machine-learning models make decisions. In particular, models designed to differentiate images of, say, cats and dogs, make decisions based on features that do not align with how humans make classifications. Simply changing these features can make the model consistently misclassify cats as dogs, without changing anything in the image thats really meaningful to humans.

Results indicated some models which may be used to, say, identify abnormalities in medical images or help autonomous cars identify objects in the road arent exactly up to snuff. People often think these models are superhuman, but they didnt actually solve the classification problem we intend them to solve, Madry says. And their complete vulnerability to adversarial examples was a manifestation of that fact. That was an eye-opening finding.

Thats why Madry seeks to make machine-learning models more interpretable to humans. New models hes developed show how much certain pixels in images the system is trained on can influence the systems predictions. Researchers can then tweak the models to focus on pixels clusters more closely correlated with identifiable features such as detecting an animals snout, ears, and tail. In the end, that will help make the models more humanlike or superhumanlike in their decisions. To further this work, Madry and his colleagues recently founded the MIT Center for Deployable Machine Learning, a collaborative research effort within the MIT Quest for Intelligence that is working toward building machine-learning tools ready for real-world deployment.

We want machine learning not just as a toy, but as something you can use in, say, an autonomous car, or health care. Right now, we dont understand enough to have sufficient confidence in it for those critical applications, Madry says.

Shaping education and policy

Madry views artificial intelligence and decision making (AI+D is one of the three new academic units in the Department of Electrical Engineering and Computer Science) as the interface of computing thats going to have the biggest impact on society.

In that regard, he makes sure to expose his students to the human aspect of computing. In part, that means considering consequences of what theyre building. Often, he says, students will be overly ambitious in creating new technologies, but they havent thought through potential ramifications on individuals and society. Building something cool isnt a good enough reason to build something, Madry says. Its about thinking about not if we can build something, but if we should build something.

Madry has also been engaging in conversations about laws and policies to help regulate machine learning. A point of these discussions, he says, is to better understand the costs and benefits of unleashing machine-learning technologies on society.

Sometimes we overestimate the power of machine learning, thinking it will be our salvation. Sometimes we underestimate the cost it may have on society, Madry says. To do machine learning right, theres still a lot still left to figure out.

Read the rest here:
Doing machine learning the right way - MIT News

Chinese ecommerce giant Alibaba has announced a breakthrough in natural language processing (NLP) through machine learning.

NLP is a key technology in the field of speech technologies such as machine translation and automatic speech recognition. The companys DAMO academy, a global research program, has made a breakthrough in machine reading techniques with applications in the fight against coronavirus.

Alibaba not only topped the GLUE Benchmark rankings, a table measuring the performance of competing NLP models, despite competition from the likes of Google, Facebook and Microsoft, but beat human baselines, signifying that its model could even outperform a human at understanding language. Applications include sentiment analysis, textual entailment (i.e. understanding the correct chronology of sentences) and question-answering.

SEE ALSO:

With the solution already deployed in technologies ranging from AI chatbots to search engines, it is now finding use in the analysis of healthcare records by centers for disease control in cities across China.

We are excited to achieve a new breakthrough in driving research of the NLP development, said Si Luo, head of NLP Research at Alibaba DAMO Academy. Not only NLP as a core technology underpinning Alibabas various businesses, which serve hundreds of millions of customers, but it also becomes a critical technology now in fighting the coronavirus. We hope we can continue to leverage our leading technologies and contribute to the community during this difficult time.

Other AI initiatives put forth by the company for use in containing the coronavirus epidemic include technology to assist in the diagnosis of the virus. The company also made its Alibaba Cloud computing platform free for research organisations seeking to sequence the virus genome.

Read more:
Alibaba using machine learning to fight coronavirus with AI - Gigabit Magazine - Technology News, Magazine and Website