Deep Science: Using machine learning to study anatomy, weather and earthquakes

Research papers come out far too rapidly for anyone to read them all, especially in the field of machine learning, which now affects (and produces papers in) practically every industry and company. This column aims to collect the most relevant recent discoveries and papers — particularly in but not limited to artificial intelligence — and explain why they matter.

A number of recently published research projects have used machine learning to attempt to better understand or predict these phenomena.

This week has a bit more “basic research” than consumer applications. Machine learning can be applied to advantage in many ways users benefit from, but it’s also transformative in areas like seismology and biology, where enormous backlogs of data can be leveraged to train AI models or as raw material to be mined for insights.

Inside earthshakers

We’re surrounded by natural phenomena that we don’t really understand — obviously we know where earthquakes and storms come from, but how exactly do they propagate? What secondary effects are there if you cross-reference different measurements? How far ahead can these things be predicted?

A number of recently published research projects have used machine learning to attempt to better understand or predict these phenomena. With decades of data available to draw from, there are insights to be gained across the board this way — if the seismologists, meteorologists and geologists interested in doing so can obtain the funding and expertise to do so.

The most recent discovery, made by researchers at Los Alamos National Labs, uses a new source of data as well as ML to document previously unobserved behavior along faults during “slow quakes.” Using synthetic aperture radar captured from orbit, which can see through cloud cover and at night to give accurate, regular imaging of the shape of the ground, the team was able to directly observe “rupture propagation” for the first time, along the North Anatolian Fault in Turkey.

“The deep-learning approach we developed makes it possible to automatically detect the small and transient deformation that occurs on faults with unprecedented resolution, paving the way for a systematic study of the interplay between slow and regular earthquakes, at a global scale,” said Los Alamos geophysicist Bertrand Rouet-Leduc.

Another effort, which has been ongoing for a few years now at Stanford, helps Earth science researcher Mostafa Mousavi deal with the signal-to-noise problem with seismic data. Poring over data being analyzed by old software for the billionth time one day, he felt there had to be better way and has spent years working on various methods. The most recent is a way of teasing out evidence of tiny earthquakes that went unnoticed but still left a record in the data.

The “Earthquake Transformer” (named after a machine-learning technique, not the robots) was trained on years of hand-labeled seismographic data. When tested on readings collected during Japan’s magnitude 6.6 Tottori earthquake, it isolated 21,092 separate events, more than twice what people had found in their original inspection — and using data from less than half of the stations that recorded the quake.

Map of minor seismic events detected by the Earthquake Transformer.

Image Credits: Stanford University

The tool won’t predict earthquakes on its own, but better understanding the true and full nature of the phenomena means we might be able to by other means. “By improving our ability to detect and locate these very small earthquakes, we can get a clearer view of how earthquakes interact or spread out along the fault, how they get started, even how they stop,” said co-author Gregory Beroza.

#artificial-general-intelligence, #artificial-intelligence, #cybernetics, #deep-learning, #emerging-technologies, #machine-learning, #science, #simulation, #tc

0

AI’s next act: Genius chips, programmable silicon and the future of computing

If only 10% of the world had enough power to run a cell phone, would mobile have changed the world in the way that it did?

It’s often said the future is already here — just not evenly distributed. That’s especially true in the world of artificial intelligence (AI) and machine learning (ML). Many powerful AI/ML applications already exist in the wild, but many also require enormous computational power — often at scales only available to the largest companies in existence or entire nation-states. Compute-heavy technologies are also hitting another roadblock: Moore’s law is plateauing and the processing capacity of legacy chip architectures are running up against the limits of physics.

If major breakthroughs in silicon architecture efficiency don’t happen, AI will suffer an unevenly distributed future and huge swaths of the population miss out on the improvements AI could make to their lives.

The next evolutionary stage of technology depends on completing the transformation that will make silicon architecture as flexible, efficient and ultimately programmable as the software we know today. If we cannot take major steps to provide easy access to ML we’ll lose unmeasurable innovation by having only a few companies in control of all the technology that matters. So what needs to change, how fast is it changing and what will that mean for the future of technology?

An inevitable democratization of AI: A boon for startups and smaller businesses

If you work at one of the industrial giants (including those “outside” of tech), congratulations — but many of the problems with current AI/ML computing capabilities I present here may not seem relevant.

For those of you working with lesser caches of resources, whether financially or talent-wise, view the following predictions as the herald of a new era in which organizations of all sizes and balance sheets have access to the same tiers of powerful AI and ML-powered software. Just like cell phones democratized internet access, we see a movement in the industry today to put AI in the hands of more and more people.

Of course, this democratization must be fueled by significant technological advancement that actually makes AI more accessible — good intentions are not enough, regardless of the good work done by companies like Intel and Google. Here are a few technological changes we’ll see that will make that possible.

From dumb chip to smart chip to “genius” chip

For a long time, raw performance was the metric of importance for processors. Their design reflected this. As software rose in ubiquity, processors needed to be smarter: more efficient and more commoditized, so specialized processors like GPUs arose — “smart” chips, if you will.

Those purpose-built graphics processors, by happy coincidence, proved to be more useful than CPUs for deep learning functions, and thus the GPU became one of the key players in modern AI and ML. Knowing this history, the next evolutionary step becomes obvious: If we can purpose-build hardware for graphics applications, why not for specific deep learning, AI and ML?

There’s also a unique confluence of factors that makes the next few years pivotal for chipmaking and tech in general. First and second, we’re seeing a plateauing of Moore’s law (which predicts a doubling of transistors on integrated circuits every two years) and the end of Dennard scaling (which says performance-per-watt doubles at about the same rate). Together, that used to mean that for any new generation of technology, chips doubled in density and increased in processing power while drawing the same amount of power. But we’ve now reached the scale of nanometers, meaning we’re up against the limitations of physics.

Thirdly, compounding the physical challenge, the computing demands of next-gen AI and ML apps are beyond what we could have imagined. Training neural networks to within even a fraction of human image recognition, for example, is surprisingly hard and takes huge amounts of processing power. The most intense applications of machine learning are things like natural language processing (NLP), recommender systems that deal with billions or trillions of possibilities, or super high-resolution computer vision, as is used in the medical and astronomical fields.

Even if we could have predicted we’d have to create and train algorithmic brains to learn how to speak human language or identify objects in deep space, we still could not have guessed just how much training — and therefore processing power — they might need to become truly useful and “intelligent” models.

Of course, many organizations are performing these sorts of complex ML applications. But these sorts of companies are usually business or scientific leaders with access to huge amounts of raw computing power and the talent to understand and deploy them. All but the largest enterprises are locked out of the upper tiers of ML and AI.

That’s why the next generation of smart chips — call them “genius” chips — will be about efficiency and specialization. Chip architecture will be made to optimize for the software running on it and run altogether more efficiently. When using high-powered AI doesn’t take a whole server farm and becomes accessible to a much larger percentage of the industry, the ideal conditions for widespread disruption and innovation become real. Democratizing expensive, resource intensive AI goes hand-in-hand with these soon-to-be-seen advances in chip architecture and software-centered hardware design.

A renewed focus on future-proofing innovation

The nature of AI creates a special challenge for the creators and users of AI hardware. The amount of change itself is huge: We’re living through the leap from humans writing code to software 2.0 — where engineers can train machine learning programs to eventually “run themselves.” The rate of change is also unprecedented: ML models can be obsolete in months or even weeks, and the very methods through which training happens are in constant evolution.

But creating new AI hardware products still requires designing, prototyping, calibrating, troubleshooting, production and distribution. It can take two years from concept to product-in-hand. Software has, of course, always outpaced hardware development, but now the differential in velocity is irreconcilable. We need to be more clever about the hardware we create for a future we increasingly cannot predict.

In fact, the generational way we think about technology is beginning to break down. When it comes to ML and AI, hardware must be built with the expectation that much of what we know today will be obsolete by the time we have the finished product. Flexibility and customization will be the key attributes of successful hardware in the age of AI, and I believe this will be a further win for entire market.

Instead of sinking resources into the model du jour or a specific algorithm, companies looking to take advantage of these technologies will have more options for processing stacks that can evolve and change as the demands of ML and AI models evolve and change.

This will allow companies of all sizes and levels of AI savvy to stay creative and competitive for longer and prevent the stagnation that can occur when software is limited by hardware — all leading to more interesting and unexpected AI applications for more organizations.

Widespread adoption of real AI and ML technologies

I’ll be the first to admit to tech’s fascination with shiny objects. There was a day when big data was the solution to everything and IoT was to be the world’s savior. AI has been through the hype cycle in the same way (arguably multiple times). Today, you’d be hard pressed to find a tech company that doesn’t purport to use AI in some way, but chances are they are doing something very rudimentary that’s more akin to advanced analytics.

It’s my firm belief that the AI revolution we’ve all been so excited about simply has not happened yet. In the next two to three years however, as the hardware that enables “real” AI power makes its way into more and more hands, it will happen. As far as predicting the change and disruption that will come from widespread access to the upper echelons of powerful ML and AI — there are few ways to make confident predictions, but that is exactly the point!

Much like cellphones put so much power in the hands of regular people everywhere, with no barriers to entry either technical or financial (for the most part), so will the coming wave of software-defined hardware that is flexible, customizable and future-proof. The possibilities are truly endless, and it will mark an important turning point in technology. The ripple effects of AI democratization and commoditization will not stop with just technology companies, and so even more fields stand to be blown open as advanced, high-powered AI becomes accessible and affordable.

Much of the hype around AI — all the disruption it was supposed to bring and the leaps it was supposed to fuel — will begin in earnest in the next few years. The technology that will power it is being built as we speak or soon to be in the hands of the many people in the many industries who will use their newfound access as a springboard to some truly amazing advances. We’re especially excited to be a part of this future, and look forward to all the progress it will bring.

#artificial-general-intelligence, #artificial-intelligence, #column, #hardware, #machine-learning, #ml, #natural-language-processing, #neural-networks, #science

0

Abacus.AI raises another $22M and launches new AI modules

AI startup RealityEngines.AI changed its name to Abacus.AI in July. At the same time, it announced a $13 million Series A round. Today, only a few months later, it is not changing its name again, but it is announcing a $22 million Series B round, led by Coatue, with Decibel Ventures and Index Partners participating as well. With this, the company, which was co-founded by former AWS and Google exec Bindu Reddy, has now raised a total of $40.3 million.

Abacus co-founder Bindu Reddy, Arvind Sundararajan and Siddartha Naidu. Image Credits: Abacus.AI

In addition to the new funding, Abacus.AI is also launching a new product today, which it calls Abacus.AI Deconstructed. Originally, the idea behind RealityEngines/Abacus.AI was to provide its users with a platform that would simplify building AI models by using AI to automatically train and optimize them. That hasn’t changed, but as it turns out, a lot of (potential) customers had already invested into their own workflows for building and training deep learning models but were looking for help in putting them into production and managing them throughout their lifecycle.

“One of the big pain points [businesses] had was, ‘look, I have data scientists and I have my models that I’ve built in-house. My data scientists have built them on laptops, but I don’t know how to push them to production. I don’t know how to maintain and keep models in production.’ I think pretty much every startup now is thinking of that problem,” Reddy said.

Image Credits: Abacus.AI

Since Abacus.AI had already built those tools anyway, the company decided to now also break its service down into three parts that users can adapt without relying on the full platform. That means you can now bring your model to the service and have the company host and monitor the model for you, for example. The service will manage the model in production and, for example, monitor for model drift.

Another area Abacus.AI has long focused on is model explainability and de-biasing, so it’s making that available as a module as well, as well as its real-time machine learning feature store that helps organizations create, store and share their machine learning features and deploy them into production.

As for the funding, Reddy tells me the company didn’t really have to raise a new round at this point. After the company announced its first round earlier this year, there was quite a lot of interest from others to also invest. “So we decided that we may as well raise the next round because we were seeing adoption, we felt we were ready product-wise. But we didn’t have a large enough sales team. And raising a little early made sense to build up the sales team,” she said.

Reddy also stressed that unlike some of the company’s competitors, Abacus.AI is trying to build a full-stack self-service solution that can essentially compete with the offerings of the big cloud vendors. That — and the engineering talent to build it — doesn’t come cheap.

Image Credits: Abacus.AI

It’s no surprise then that Abacus.AI plans to use the new funding to increase its R&D team, but it will also increase its go-to-market team from two to ten in the coming months. While the company is betting on a self-service model — and is seeing good traction with small- and medium-sized companies — you still need a sales team to work with large enterprises.

Come January, the company also plans to launch support for more languages and more machine vision use cases.

“We are proud to be leading the Series B investment in Abacus.AI, because we think that Abacus.AI’s unique cloud service now makes state-of-the-art AI easily accessible for organizations of all sizes, including start-ups. Abacus.AI’s end-to-end autonomous AI service powered by their Neural Architecture Search invention helps organizations with no ML expertise easily deploy deep learning systems in production.”

 

#artificial-general-intelligence, #artificial-intelligence, #bindu-reddy, #cloud, #cloud-computing, #co-founder, #coatue, #enterprise, #entrepreneurship, #funding, #fundings-exits, #learning, #machine-learning, #ml, #recent-funding, #science-and-technology, #start-ups, #startups

0

4 things to remember when adapting AI/ML learning models during a pandemic

The machine learning and AI-powered tools being deployed in response to COVID-19 arguably improve certain human activities and provide essential insights needed to make certain personal or professional decisions; however, they also highlight a few pervasive challenges faced by both machines and the humans that create them.

Nevertheless, the progress seen in AI/machine learning leading up to and during the COVID-19 pandemic cannot be ignored. This global economic and public health crisis brings with it a unique opportunity for updates and innovation in modeling, so long as certain underlying principles are followed.

Here are four industry truths (note: this is not an exhaustive list) my colleagues and I have found that matter in any design climate, but especially during a global pandemic climate.

Some success can be attributed to chance, rather than reasoning

When a big group of people is collectively working on a problem, success may become more likely. Looking at historic examples like the 2008 Global Financial Crisis, there were several analysts credited with predicting the crisis. This may seem miraculous to some until you consider that more than 200,000 people were working in Wall Street, each of them making their own predictions. It then becomes less of a miracle and more of a statistically probable outcome. With this many individuals simultaneously working on modeling and predictions, it was highly likely someone would get it right by chance.

Similarly, with COVID-19 there are a lot of people involved, from statistical modelers and data scientists to vaccine specialists, and there is also an overwhelming eagerness to find solutions and concrete data-based answers. Following appropriate statistical rigor, coupled with machine learning and AI, can improve these models and decrease the chances of false predictions that arrive from too many predictions being made.

Automation can help in maintaining productivity if used wisely

During a crisis, time-management is essential. Automation technology can be used not only as part of the crisis solution, but also as a tool for monitoring productivity and contributions of team members working on the solution. For modeling, automation can also greatly improve the speed of results. Every second a piece of software can perform automation for a model, it allows a data scientist (or even a medical scientist) to conduct other more important tasks. User-friendly platforms in the market now give more people, like business analysts, access to predictions from custom machine learning models.

#artificial-general-intelligence, #artificial-intelligence, #automation-technology, #column, #coronavirus, #covid-19, #cybernetics, #data-scientist, #developer, #hiring, #machine-learning, #tc

0

WhyLabs brings more transparancy to ML ops

WhyLabs, a new machine learning startup that was spun out of the Allen Institute, is coming out of stealth today. Founded by a group of former Amazon machine learning engineers, Alessya Visnjic, Sam Gracie and Andy Dang, together with Madrona Venture Group principal Maria Karaivanova, WhyLabs’ focus is on ML operations after models have been trained — not on building those models from the ground up.

The team also today announced that it has raised a $4 million seed funding round from Madrona Venture Group, Bezos Expeditions, Defy Partners and Ascend VC.

Visnjic, the company’s CEO, used to work on Amazon’s demand forecasting model.

“The team was all research scientists, and I was the only engineer who had kind of tier-one operating experience,” she told me. “So it was like, ”Okay, how bad could it be?’ I carried the pager for the retail website before it can be bad. But it was one of the first AI deployments that we’d done at Amazon at scale. The pager duty was extra fun because there were no real tools. So when things would go wrong — like we’d order way too many black socks out of the blue — it was a lot of manual effort to figure out why was this happening.”

Image Credits: WhyLabs

But while large companies like Amazon have built their own internal tools to help their data scientists and AI practitioners operate their AI systems, most enterprises continue to struggle with this — and a lot of AI projects simply fail and never make it into production. “We believe that one of the big reasons that happens is because of the operating process that remains super manual,” Visnjic said. “So at WhyLabs, we’re building the tools to address that — specifically to monitor and track data quality and alert — you can think of it as Datadog for AI applications.”

The team has brought ambitions, but to get started, it is focusing on observability. The team is building — and open-sourcing — a new tool for continuously logging what’s happening in the AI system, using a low-overhead agent. That platform-agnostic system, dubbed WhyLogs, is meant to help practitioners understand the data that moves through the AI/ML pipeline.

For a lot of businesses, Visnjic noted, the amount of data that flows through these systems is so large that it doesn’t make sense for them to keep “lots of big haystacks with possibly some needles in there for some investigation to come in the future.” So what they do instead is just discard all of this. With its data logging solution, WhyLabs aims to give these companies the tools to investigate their data and find issues right at the start of the pipeline.

Image Credits: WhyLabs

According to Karaivanova, the company doesn’t have paying customers yet, but it is working on a number of proofs of concepts. Among those users is Zulily, which is also a design partner for the company. The company is going after mid-size enterprises for the time being, but as Karaivanova noted, to hit the sweet spot for the company, a customer needs to have an established data science team with 10 to 15 ML practitioners. While the team is still figuring out its pricing model, it’ll likely be a volume-based approach, Karaivanova said.

“We love to invest in great founding teams who have built solutions at scale inside cutting-edge companies, who can then bring products to the broader market at the right time. The WhyLabs team are practitioners building for practitioners. They have intimate, first-hand knowledge of the challenges facing AI builders from their years at Amazon and are putting that experience and insight to work for their customers,” said Tim Porter, managing director at Madrona. “We couldn’t be more excited to invest in WhyLabs and partner with them to bring cross-platform model reliability and observability to this exploding category of MLOps.”

#amazon, #artificial-general-intelligence, #artificial-intelligence, #bezos-expeditions, #cybernetics, #datadog, #defy-partners, #engineer, #enterprise, #jeff-bezos, #machine-learning, #madrona-venture-group, #ml, #mlops, #performance-management, #recent-funding, #science-and-technology, #startups, #tc, #whylabs

0

What does GPT-3 mean for the future of the legal profession?

One doesn’t have to dig too deep into legal organizations to find people who are skeptical about artificial intelligence.

AI is getting tremendous attention and significant venture capital, but AI tools frequently underwhelm in the trenches. Here are a few reasons why that is and why I believe GPT-3, a beta version of which was recently released by the OpenAI Foundation, might be a game changer in legal and other knowledge-focused organizations.

GPT-3 is getting a lot of oxygen lately because of its size, scope and capabilities. However, it should be recognized that a significant amount of that attention is due to its association with Elon Musk. The OpenAI Foundation that created GPT-3 was founded by heavy hitters Musk and Sam Altman and is supported by Mark Benioff, Peter Thiel and Microsoft, among others. Arthur C. Clarke once observed that great innovations happen after everyone stops laughing.

Musk has made the world stop laughing in so many ambitious areas that the world is inclined to give a project in which he’s had a hand a second look. GPT-3 is getting the benefit of that spotlight. I suggest, however, that the attention might be warranted on its merits.

Why have some AI-based tools struggled in the legal profession, and how might GPT-3 be different?

1. Not every problem is a nail

It is said that when you’re a hammer, every problem is a nail. The networks and algorithms that power AI are quite good at drawing correlations across enormous datasets that would not be obvious to humans. One of my favorite examples of this is a loan-underwriting AI that determined that the charge level of the battery on your phone at the time of application is correlated to your underwriting risk. Who knows why that is? A human would not have surmised that connection. Those things are not rationally related, just statistically related.

#artificial-general-intelligence, #artificial-intelligence, #column, #deep-learning, #developer, #gpt-3, #lawyers, #machine-learning, #openai, #science, #startups, #tc, #verified-experts

0

Hour One raises $5M Seed to generate AI-driven synthetic characters from real humans

All of the people pictured above are real, but what you are seeing are synthetically-generated versions of their real selves. And they can be programmed to say anything. Tech futurists have long warned about humans being replaced by life-like AI-driven figures, where it would be almost impossible to tell between machine and human. Indeed, there’s even a new book on this subject of ‘deep fakes’.

But that future comes a step closer today with the news that Hour One, which creates AI-driven synthetic characters based on real humans, closes a $5 million seed funding led by Galaxy Interactive (via its Galaxy EOS VC Fund), Remagine Ventures and Kindred Ventures (with participation of Amaranthine).

Hour One will use the funds to scale its AI-driven cloud platform, onboard ‘thousands’ of new characters, and expand its commercial activities.
 
Founded in 2019, Hour One develops technologies for creating high-quality digital characters based on real people. The idea is to generate production-grade video-based characters in a highly scalable and cost-effective way. The upshot of this is that what appears to be a real human could talk about any product or subject at all, to the point of infinite scale.

This was showcased at its “real or synthetic” likeness test at CES 2020, challenging people to distinguish between real and synthetic characters generated by its AI.

Oren Aharon, Hour One’s Founder and CEO said in a statement: “We believe that synthetic characters of real people will become a part of our everyday life. Our vision is that Hour One will drive the use of synthetic characters to improve the quality of communication between businesses and people across markets and use cases. By enabling each person to create their own character together with our scalable cloud platform, we will provide a variety of solutions for next-gen remote business-to-human interactions.”
 
Hour One is currently working with companies in the e-commerce, education, automotive, communication, and enterprise sectors, with expanded industry applications expected throughout 2020.

The company also showcased its “real or synthetic” likeness test at CES 2020, challenging people to distinguish between real and synthetic characters generated by its AI.

The real issue, however, is how will this technology be deployed without it being abused.

Lior Hakim, cofounder and CTO says this potential problem is dealt with via encryption technologies to secure the use and rights of the characters enabling “anyone to identify our videos as well as mark them as altered to notify the viewers”. The company also says it has an ethical policy code for how its technology is used.

Sam Englebardt, Co-Founder and Managing Director of Galaxy Interactive says the startup’s “ethics-driven approach to the creation of synthetic video” is key and that “given how challenging production with live actors has become as a result of COVID-19, now is the perfect time for businesses of all sizes to produce their content with Hour One’s synthetic characters.”

Clearly this will reduce the cost of synthetic character creation meaning any textual content could be “automatically translated into a live-action video of a person that engages an audience by speaking the text” said Eze Vidra, Co-Founder and Managing Partner at Remagine Ventures .

Speaking to TechCrunch, Business strategy lead for Hour One Natalie Monbiot, said the company has a unique ability to onboard “basically any human being and turn them into a synthetic character that’s a lifelike replica of that person. So it’s not an avatar or a version of that person. It really does look and behave like that person. You can then basically generate new content by uploading new texts. So, for example, in e-commerce, you can pick your characters and get them to present your product or do a product presentation. This means every single product SKU can have its own video presentation.”

#articles, #artificial-general-intelligence, #artificial-intelligence, #ceo, #e-commerce, #emerging-technologies, #europe, #eze-vidra, #kindred-ventures, #remagine-ventures, #science-and-technology, #tc, #techcrunch

0

AI is struggling to adjust to 2020

2020 has made every industry reimagine how to move forward in light of COVID-19: civil rights movements, an election year and countless other big news moments. On a human level, we’ve had to adjust to a new way of living. We’ve started to accept these changes and figure out how to live our lives under these new pandemic rules. While humans settle in, AI is struggling to keep up.

The issue with AI training in 2020 is that, all of a sudden, we’ve changed our social and cultural norms. The truths that we have taught these algorithms are often no longer actually true. With visual AI specifically, we’re asking it to immediately interpret the new way we live with updated context that it doesn’t have yet.

Algorithms are still adjusting to new visual queues and trying to understand how to accurately identify them. As visual AI catches up, we also need a renewed importance on routine updates in the AI training process so inaccurate training datasets and preexisting open-source models can be corrected.

Computer vision models are struggling to appropriately tag depictions of the new scenes or situations we find ourselves in during the COVID-19 era. Categories have shifted. For example, say there’s an image of a father working at home while his son is playing. AI is still categorizing it as “leisure” or “relaxation.” It is not identifying this as ‘”work” or “office,” despite the fact that working with your kids next to you is the very common reality for many families during this time.

Image Credits: Westend61/Getty Images

On a more technical level, we physically have different pixel depictions of our world. At Getty Images, we’ve been training AI to “see.” This means algorithms can identify images and categorize them based on the pixel makeup of that image and decide what it includes. Rapidly changing how we go about our daily lives means that we’re also shifting what a category or tag (such as “cleaning”) entails.

Think of it this way — cleaning may now include wiping down surfaces that already visually appear clean. Algorithms have been previously taught that to depict cleaning, there needs to be a mess. Now, this looks very different. Our systems have to be retrained to account for these redefined category parameters.

This relates on a smaller scale as well. Someone could be grabbing a door knob with a small wipe or cleaning their steering wheel while sitting in their car. What was once a trivial detail now holds importance as people try to stay safe. We need to catch these small nuances so it’s tagged appropriately. Then AI can start to understand our world in 2020 and produce accurate outputs.

Image Credits: Chee Gin Tan/Getty Images

Another issue for AI right now is that machine learning algorithms are still trying to understand how to identify and categorize faces with masks. Faces are being detected as solely the top half of the face, or as two faces — one with the mask and a second of only the eyes. This creates inconsistencies and inhibits accurate usage of face detection models.

One path forward is to retrain algorithms to perform better when given solely the top portion of the face (above the mask). The mask problem is similar to classic face detection challenges such as someone wearing sunglasses or detecting the face of someone in profile. Now masks are commonplace as well.

Image Credits: Rodger Shija/EyeEm/Getty Images

What this shows us is that computer vision models still have a long way to go before truly being able to “see” in our ever-evolving social landscape. The way to counter this is to build robust datasets. Then, we can train computer vision models to account for the myriad different ways a face may be obstructed or covered.

At this point, we’re expanding the parameters of what the algorithm sees as a face — be it a person wearing a mask at a grocery store, a nurse wearing a mask as part of their day-to-day job or a person covering their face for religious reasons.

As we create the content needed to build these robust datasets, we should be aware of potentially increased unintentional bias. While some bias will always exist within AI, we now see imbalanced datasets depicting our new normal. For example, we are seeing more images of white people wearing masks than other ethnicities.

This may be the result of strict stay-at-home orders where photographers have limited access to communities other than their own and are unable to diversify their subjects. It may be due to the ethnicity of the photographers choosing to shoot this subject matter. Or, due to the level of impact COVID-19 has had on different regions. Regardless of the reason, having this imbalance will lead to algorithms being able to more accurately detect a white person wearing a mask than any other race or ethnicity.

Data scientists and those who build products with models have an increased responsibility to check for the accuracy of models in light of shifts in social norms. Routine checks and updates to training data and models are key to ensuring quality and robustness of models — now more than ever. If outputs are inaccurate, data scientists can quickly identify them and course correct.

It’s also worth mentioning that our current way of living is here to stay for the foreseeable future. Because of this, we must be cautious about the open-source datasets we’re leveraging for training purposes. Datasets that can be altered, should. Open-source models that cannot be altered need to have a disclaimer so it’s clear what projects might be negatively impacted from the outdated training data.

Identifying the new context we’re asking the system to understand is the first step toward moving visual AI forward. Then we need more content. More depictions of the world around us — and the diverse perspectives of it. As we’re amassing this new content, take stock of new potential biases and ways to retrain existing open-source datasets. We all have to monitor for inconsistencies and inaccuracies. Persistence and dedication to retraining computer vision models is how we’ll bring AI into 2020.

#articles, #artificial-general-intelligence, #artificial-intelligence, #column, #coronavirus, #covid-19, #cybernetics, #emerging-technologies, #getty-images, #machine-learning, #opinion, #science-and-technology

0

DeepMind’s Agent57 AI agent can best human players across a suite of 57 Atari games

Development of artificial intelligence agents tends to frequently be measured by their performance in games, but there’s a good reason for that: Games tend to offer a wide proficiency curve, in terms of being relatively simple to grasp the basics, but difficult to master, and they almost always have a built-in scoring system to evaluate performance. DeepMind’s agents have tackled board game Go, as well as real-time strategy video game StarCraft – but the Alphabet company’s most recent feat is Agent57, a learning agent that can beat the average human on each of 57 Atari games with a wide range of difficulty, characteristics and gameplay styles.

Being better than humans at 57 Atari games may seem like an odd benchmark against which to measure the performance of a deep learning agent, but it’s actually a standard that goes all the way back to 2012, with a selection of Atari classics including Pitfall, Solaris, Montezuma’s Revenge and many others. Taken together, these games represent a broad range of difficulty levels, as well as requiring a range of different strategies in order to achieve success.

That’s a great type of challenge for creating a deep learning agent because the goal is not to build something that can determine one effective strategy that maximizes your chances of success every time you play a game – instead, the reason researchers build these agents and set them to these tasks at all is to develop something that can learn across multiple and shifting scenarios and conditions, with the long-term aim of building a learning agent that approaches general AI – or AI that is more human in terms of being able to apply its intelligence to any problem put before it, including challenges it’s never encountered before.

DeepMind’s Agent57 is remarkable because it performs better than human players on each of the 57 games in the Atari57 set – previous agents have been able to be better than human players on average – but that’s because they were extremely good at some of the simpler games that basically just worked via a simple action-reward loop, but terrible at games that required more advanced play, including long-term exploration and memory, like Montezuma’s Revenge.

The DeepMind team addressed this by building a distributed agent with different computers tackling different aspects of the problem, with some tuned to focus on novelty rewards (encountering things they haven’t encountered before), with both short- and long-term time horizons for when the novelty value resets. Others sought out more simple exploits, figuring out which repeated pattern provided the biggest reward, and then all the results are combined and managed by an agent equipped with a meta-controller that allows it to weight the costs and benefits of different approaches based on which game it encounters.

In the end, Agent57 is an accomplishment, but the team says it can stand to be improved in a few different ways. First, it’s incredibly computationally expensive to run, so they will seek to streamline that. Second, it’s actually not as good at some of the simpler games as some simpler agents – even though it excels at the the top 5 games in terms of challenge to previous intelligent agents. The team says it has ideas for how to make it even better at the simpler games that other, less sophisticated agents, are even better at.

#articles, #artificial-general-intelligence, #artificial-intelligence, #atari, #deep-learning, #deepmind, #emerging-technologies, #science, #tc, #technology, #video-game

0