Google gives its AI the reins over its data center cooling systems

The inside of data centers is loud and hot — and keeping servers from overheating is a major factor in the cost of running them. It’s no surprise then that the big players in this space, including Facebook, Microsoft and Google, all look for different ways of saving cooling costs. Facebook uses cool outside air when possible, Microsoft is experimenting with underwater data centers and Google is being Google and looking to its AI models for some extra savings.

A few years ago, Google, through its DeepMind affiliate, started looking into how it could use machine learning to provide its operators some additional guidance on how to best cool its data centers. At the time, though, the system only made recommendations and the human operators decided whether to implement them. Those humans can now take longer naps during the afternoon, because the team has decided the models are now good enough to give the AI-powered system full control over the cooling system. Operators can still intervene, of course, but as long as the AI doesn’t decide to burn the place down, the system runs autonomously.

The new cooling system is now in place in a number of Google’s data centers. Every five minutes, the system polls thousands of sensors inside the data center and chooses the optimal actions based on this information. There are all kinds of checks and balances here, of course, so the chances of one of Google’s data centers going up in flames because of this is low.

Like most machine learning models, this one also became better as it gathered more data. It’s now delivering energy savings of 30 percent on average, compared to the data centers’ historical energy usage.

One thing that’s worth noting here is that Google is obviously trying to save a few bucks, but in many ways, the company is also looking at this as a way of promoting its own machine learning services. What works in a data center, after all, should also work in a large office building. “In the long term, we think there’s potential to apply this technology in other industrial settings and help tackle climate change on an even grander scale,” DeepMind writes in today’s announcement.

Amazon starts shipping its $249 DeepLens AI camera for developers

Back at its re:Invent conference in November, AWS announced its $249 DeepLens, a camera that’s specifically geared toward developers who want to build and prototype vision-centric machine learning models. The company started taking pre-orders for DeepLens a few months ago, but now the camera is actually shipping to developers.

Ahead of today’s launch, I had a chance to attend a workshop in Seattle with DeepLens senior product manager Jyothi Nookula and Amazon’s VP for AI Swami Sivasubramanian to get some hands-on time with the hardware and the software services that make it tick.

DeepLens is essentially a small Ubuntu- and Intel Atom-based computer with a built-in camera that’s powerful enough to easily run and evaluate visual machine learning models. In total, DeepLens offers about 106 GFLOPS of performance.

The hardware has all of the usual I/O ports (think Micro HDMI, USB 2.0, Audio out, etc.) to let you create prototype applications, no matter whether those are simple toy apps that send you an alert when the camera detects a bear in your backyard or an industrial application that keeps an eye on a conveyor belt in your factory. The 4 megapixel camera isn’t going to win any prizes, but it’s perfectly adequate for most use cases. Unsurprisingly, DeepLens is deeply integrated with the rest of AWS’s services. Those include the AWS IoT service Greengrass, which you use to deploy models to DeepLens, for example, but also SageMaker, Amazon’s newest tool for building machine learning models.

These integrations are also what makes getting started with the camera pretty easy. Indeed, if all you want to do is run one of the pre-built samples that AWS provides, it shouldn’t take you more than 10 minutes to set up your DeepLens and deploy one of these models to the camera. Those project templates include an object detection model that can distinguish between 20 objects (though it had some issues with toy dogs, as you can see in the image above), a style transfer example to render the camera image in the style of van Gogh, a face detection model and a model that can distinguish between cats and dogs and one that can recognize about 30 different actions (like playing guitar, for example). The DeepLens team is also adding a model for tracking head poses. Oh, and there’s also a hot dog detection model.

But that’s obviously just the beginning. As the DeepLens team stressed during our workshop, even developers who have never worked with machine learning can take the existing templates and easily extend them. In part, that’s due to the fact that a DeepLens project consists of two parts: the model and a Lambda function that runs instances of the model and lets you perform actions based on the model’s output. And with SageMaker, AWS now offers a tool that also makes it easy to build models without having to manage the underlying infrastructure.

You could do a lot of the development on the DeepLens hardware itself, given that it is essentially a small computer, though you’re probably better off using a more powerful machine and then deploying to DeepLens using the AWS Console. If you really wanted to, you could use DeepLens as a low-powered desktop machine as it comes with Ubuntu 16.04 pre-installed.

For developers who know their way around machine learning frameworks, DeepLens makes it easy to import models from virtually all the popular tools, including Caffe, TensorFlow, MXNet and others. It’s worth noting that the AWS team also built a model optimizer for MXNet models that allows them to run more efficiently on the DeepLens device.

So why did AWS build DeepLens? “The whole rationale behind DeepLens came from a simple question that we asked ourselves: How do we put machine learning in the hands of every developer,” Sivasubramanian said. “To that end, we brainstormed a number of ideas and the most promising idea was actually that developers love to build solutions as hands-on fashion on devices.” And why did AWS decide to build its own hardware instead of simply working with a partner? “We had a specific customer experience in mind and wanted to make sure that the end-to-end experience is really easy,” he said. “So instead of telling somebody to go download this toolkit and then go buy this toolkit from Amazon and then wire all of these together. […] So you have to do like 20 different things, which typically takes two or three days and then you have to put the entire infrastructure together. It takes too long for somebody who’s excited about learning deep learning and building something fun.”

So if you want to get started with deep learning and build some hands-on projects, DeepLens is now available on Amazon. At $249, it’s not cheap, but if you are already using AWS — and maybe even use Lambda already — it’s probably the easiest way to get started with building these kind of machine learning-powered applications.

Audit of NHS Trust’s app project with DeepMind raises more questions than it answers

A third party audit of a controversial patient data-sharing arrangement between a London NHS Trust and Google DeepMind appears to have skirted over the core issues that generated the controversy in the first place.

The audit (full report here) — conducted by law firm Linklaters — of the Royal Free NHS Foundation Trust’s acute kidney injury detection app system, Streams, which was co-developed with Google-DeepMind (using an existing NHS algorithm for early detection of the condition), does not examine the problematic 2015 information-sharing agreement inked between the pair which allowed data to start flowing.

“This Report contains an assessment of the data protection and confidentiality issues associated with the data protection arrangements between the Royal Free and DeepMind . It is limited to the current use of Streams, and any further development, functional testing or clinical testing, that is either planned or in progress. It is not a historical review,” writes Linklaters, adding that: “It includes consideration as to whether the transparency, fair processing, proportionality and information sharing concerns outlined in the Undertakings are being met.”

Yet it was the original 2015 contract that triggered the controversy, after it was obtained and published by New Scientist, with the wide-ranging document raising questions over the broad scope of the data transfer; the legal bases for patients information to be shared; and leading to questions over whether regulatory processes intended to safeguard patients and patient data had been sidelined by the two main parties involved in the project.

In November 2016 the pair scrapped and replaced the initial five-year contract with a different one — which put in place additional information governance steps.

They also went on to roll out the Streams app for use on patients in multiple NHS hospitals — despite the UK’s data protection regulator, the ICO, having instigated an investigation into the original data-sharing arrangement.

And just over a year ago the ICO concluded that the Royal Free NHS Foundation Trust had failed to comply with Data Protection Law in its dealings with Google’s DeepMind.

The audit of the Streams project was a requirement of the ICO.

Though, notably, the regulator has not endorsed Linklaters report. On the contrary, it warns that it’s seeking legal advice and could take further action.

In a statement on its website, the ICO’s deputy commissioner for policy, Steve Wood, writes: “We cannot endorse a report from a third party audit but we have provided feedback to the Royal Free. We also reserve our position in relation to their position on medical confidentiality and the equitable duty of confidence. We are seeking legal advice on this issue and may require further action.”

In a section of the report listing exclusions, Linklaters confirms the audit does not consider: “The data protection and confidentiality issues associated with the processing of personal data about the clinicians at the Royal Free using the Streams App.”

So essentially the core controversy, related to the legal basis for the Royal Free to pass personally identifiable information on 1.6M patients to DeepMind when the app was being developed, and without people’s knowledge or consent, is going unaddressed here.

And Wood’s statement pointedly reiterates that the ICO’s investigation “found a number of shortcomings in the way patient records were shared for this trial”.

“[P]art of the undertaking committed Royal Free to commission a third party audit. They have now done this and shared the results with the ICO. What’s important now is that they use the findings to address the compliance issues addressed in the audit swiftly and robustly. We’ll be continuing to liaise with them in the coming months to ensure this is happening,” he adds.

“It’s important that other NHS Trusts considering using similar new technologies pay regard to the recommendations we gave to Royal Free, and ensure data protection risks are fully addressed using a Data Protection Impact Assessment before deployment.”

While the report is something of a frustration, given the glaring historical omissions, it does raise some points of interest — including suggesting that the Royal Free should probably scrap a Memorandum of Understanding it also inked with DeepMind, in which the pair set out their ambition to apply AI to NHS data.

This is recommended because the pair have apparently abandoned their AI research plans.

On this Linklaters writes: “DeepMind has informed us that they have abandoned their potential research project into the use of AI to develop better algorithms, and their processing is limited to execution of the NHS AKI algorithm… In addition, the majority of the provisions in the Memorandum of Understanding are non-binding. The limited provisions that are binding are superseded by the Services Agreement and the Information Processing Agreement discussed above, hence we think the Memorandum of Understanding has very limited relevance to Streams. We recommend that the Royal Free considers if the Memorandum of Understanding continues to be relevant to its relationship with DeepMind and, if it is not relevant, terminates that agreement.”

In another section, discussing the NHS algorithm that underpins the Streams app, the law firm also points out that DeepMind’s role in the project is little more than helping provide a glorified app wrapper (on the app design front the project also utilized UK app studio, ustwo, so DeepMind can’t claim app design credit either).

“Without intending any disrespect to DeepMind, we do not think the concepts underpinning Streams are particularly ground-breaking. It does not, by any measure, involve artificial intelligence or machine learning or other advanced technology. The benefits of the Streams App instead come from a very well-designed and user-friendly interface, backed up by solid infrastructure and data management that provides AKI alerts and contextual clinical information in a reliable, timely and secure manner,” Linklaters writes.

What DeepMind did bring to the project, and to its other NHS collaborations, is money and resources — providing its development resources free for the NHS at the point of use, and stating (when asked about its business model) that it would determine how much to charge the NHS for these app ‘innovations’ later.

Yet the commercial services the tech giant is providing to what are public sector organizations do not appear to have been put out to open tender.

Also notably excluded in the Linklaters’ audit: Any scrutiny of the project vis-a-vis competition law, public procurement law compliance with procurement rules, and any concerns relating to possible anticompetitive behavior.

The report does highlight one potentially problematic data retention issue for the current deployment of Streams, saying there is “currently no retention period for patient information on Streams” — meaning there is no process for deleting a patient’s medical history once it reaches a certain age.

“This means the information on Streams currently dates back eight years,” it notes, suggesting the Royal Free should probably set an upper age limit on the age of information contained in the system.

While Linklaters largely glosses over the chequered origins of the Streams project, the law firm does make a point of agreeing with the ICO that the original privacy impact assessment for the project “should have been completed in a more timely manner”.

It also describes it as “relatively thin given the scale of the project”.

Giving its response to the audit, health data privacy advocacy group MedConfidential — an early critic of the DeepMind data-sharing arrangement — is roundly unimpressed, writing: “The biggest question raised by the Information Commissioner and the National Data Guardian appears to be missing — instead, the report excludes a “historical review of issues arising prior to the date of our appointment”.

“The report claims the ‘vital interests’ (i.e. remaining alive) of patients is justification to protect against an “event [that] might only occur in the future or not occur at all”… The only ‘vital interest’ protected here is Google’s, and its desire to hoard medical records it was told were unlawfully collected. The vital interests of a hypothetical patient are not vital interests of an actual data subject (and the GDPR tests are demonstrably unmet).

“The ICO and NDG asked the Royal Free to justify the collection of 1.6 million patient records, and this legal opinion explicitly provides no answer to that question.”

Zebra Medical Vision gets $30M Series C to create AI-based tools for radiologists

Zebra Medical Vision, an Israeli medical imaging startup that uses machine and deep learning to build tools for radiologists, has raised a $30 million Series C led by health technology fund aMoon Ventures, with participation from Aurum, Johnson & Johnson Innovation—JJDC Inc. (the conglomerate’s venture capital arm), Intermountain Health and artificial intelligence experts Fei-Fei Li and Richard Socher. Existing investors Khosla Ventures, Nvidia, Marc Benioff, OurCrowd and Dolby Ventures also returned for the round.

Zebra also announced its Textray research today, which it claims is the “most comprehensive AI research conducted on chest X-rays to date.” Textray is being used to develop a new product that has already been trained using almost two million images to identify 40 clinical findings. Scheduled to launch next year, it will help automate the analysis of chest X-rays for radiologists.

Founder and CEO Elad Benjamin, who launched Zebra with Eyal Toledano and Eyal Gura in 2014, told TechCrunch in an email that its Series C capital will be spent on new hires to speed up the development of its analytics engine and its go-to-market strategy.

Chest X-rays are one of the most common radiographs ordered, but also among the most difficult for radiologists to interpret. There are several groups of researchers focused on using artificial intelligence algorithms to make the process more accurate and efficient, including teams from Google, Stanford and the Dubai Health Authority.

Benjamin said Zebra, which has now raised a total of $50 million, differentiates its approach by looking at its algorithms from a “holistic product perspective. Developing an algorithm is just one piece,” he added. “Integrating it into the workflow, adapting it to the needs of multiple countries and healthcare environments, supporting and updating it, and regulating it properly globally takes a tremendous focus – and that’s what delivers value, over and above an algorithm.”

He added that “it will take a few more years” before AI becomes mainstream in medical imaging and diagnosis, but he believes that it will eventually be a critical component of radiology. Zebra Medical Vision already markets a bundle of algorithms for lung, breast, liver, cardiovascular and bone disease diagnoses called AI1 that costs hospitals $1 per scan.

In a press statement, aMoon managing partner Dr. Yair Schindel said “We are excited to partner with the Zebra team, which is harnessing the power of data and machine learning to provide physicians and healthcare systems with tools to dramatically increase capacity, while improving patient care. This investment aligns with our vision of backing scalable and sustainable innovations that will have a valuable impact on fundamental facets of global healthcare.”

Want to fool a computer vision system? Just tweak some colors

Research into machine learning and the interesting AI models created as a consequence are popular topics these days. But there’s a sort of shadow world of scientists working to undermine these systems — not to show they’re worthless but to shore up their weaknesses. A new paper demonstrates this by showing how vulnerable image recognition models are to the simplest color manipulations of the pictures they’re meant to identify.

It’s not some deep indictment of computer vision — techniques to “beat” image recognition systems might just as easily be characterized as situations in which they perform particularly poorly. Sometimes this is something surprisingly simple: rotating an image, for example, or adding a crazy sticker. Unless a system has been trained specifically on a given manipulation or has orders to check common variations like that, it’s pretty much just going to fail.

In this case it’s research from the University of Washington led by grad student Hossein Hosseini. Their “adversarial” imagery was similarly simple: switch up the colors.

Probably many of you have tried something similar to this when fiddling around in an image manipulation program: by changing the “hue” and “saturation” values on a picture, you can make someone have green skin, a banana appear blue and so on. That’s exactly what the researchers did: twiddled the knobs so a dog looked a bit yellow, a deer looked purplish, etc.

The original images are at left; color-shifted versions and the systems’ best guesses at right.

Critically, however, the “value” of the pixels, meaning how light or dark it is, wasn’t changed, meaning the images still look like what they are — just in weird colors.

But while a cat looks like a cat no matter if it’s grey or pink to us, one can’t really say the same for a deep neural network. The accuracy of the model they tested was reduced by 90 percent on sets of color-tweaked images that it would normally identify easily. Its best guesses are pretty random, as you can see in the figure at right. Changing the colors totally changes the system’s guess.

The team tested several models and they all broke down on the color-shifted set, so it wasn’t just a consequence of this specific system.

It’s not too hard to fix — in this case, all you really need to do is add some labeled, color-shifted images into the training data so the system is exposed to them beforehand. This addition brought success rates back up to reasonable (if still fairly poor) levels.

But the point isn’t that computer vision systems are fundamentally bad at color or something. It’s that there are lots of ways of subtly or not-so-subtly manipulating an image or video that will devastate its accuracy or subvert it.

“Deep networks are very good at learning (or better memorizing) the distribution of training data,” wrote Hosseini in an email to TechCrunch. “They, however, hardly generalize beyond that. So, even if models are trained with augmented data, it’s likely that we can come up with a new type of adversarial images that can fool the model.”

A model trained to catch color variations might still be vulnerable to attention-based adversarial images and vice versa. The way these systems are created and encoded right now simply isn’t robust enough to prevent such attacks. But by cataloguing them and devising improvements that protect against some but not all, we can advance the state of the art.

“I think we need to find a way for the model to learn the concepts, such as being invariant to color or rotation,” Hosseini suggested. “That can save the algorithm a lot of training data and is more similar to how humans learn.”

You can read the full pre-print paper on Arxiv (PDF).

IBM brings its Power9 servers with Nvidia GPUs to its cloud

IBM is hosing its annual THINK conference to packed halls in Las Vegas this week. Given how important its cloud business has become to its bottom line, it’s no surprise that this event features its fair share of cloud news. Among today’s announcements it the launch of the third generation of Power Systems servers in the IBM Cloud. This comes a day after Google also confirmed that it is using these processors in its data centers, too.

These servers are designed around the recently launched Power9 RISC processor (which are themselves the latest generation of the PowerPC processors Apple once used) and Nvidia Tesla V100 GPUs. Thanks to their use of the high-speed NVLink interface, these machines are especially powerful when it comes to training machine learning models.

In addition, IBM is also bringing its PowerAI distribution to the cloud. PowerAI is essentially IBM’s deep learning platform that supports frameworks like TensorFlow, Torch and Caffe, as well as IBM’s own deep learning frameworks. Given that PowerAI has long been optimized for exactly the kind of Power servers IBM is now bringing to its Cloud (the AC922, to be exact), it’s no surprise that PowerAI will be available in the Cloud, too.

We need to improve the accuracy of AI accuracy discussions

 Reading the tech press, you would be forgiven for believing that AI is going to eat pretty much every industry and job. Not a day goes by without another reporter breathlessly reporting some new machine learning product that is going to trounce human intelligence. That surfeit of enthusiasm doesn’t originate just with journalists though — they are merely channeling the wild optimism… Read More

Atomwise, which uses AI to improve drug discovery, raises $45M Series A

 Atomwise, which uses deep learning to shorten the process of discovering new drugs, has raised a $45 million Series A. The round was led by Monsanto Growth Ventures, Data Collective (DCVC) and B Capital Group. Baidu Ventures, Tencent and Dolby Family Ventures, which are all new investors in Atomwise, also participated, as well as returning investors Y Combinator, Khosla Ventures and DFJ. Read More