How Artificial Intelligence Is Quietly Changing How You Shop Online

How Artificial Intelligence Is Quietly Changing How You Shop Online

The writer William Gibson once said that the future is here, just not evenly distributed. That was the case with the World Wide Web 20 years ago, when some business models – notably e-commerce and new media – took off faster than others. Now a similar trend is happening with artificial intelligence, or AI.

The promise of AI has seemingly been just on the horizon for years, with little evidence of change in the lives of most consumers. A few years ago, buzzwords like “big data” hinted at the potential, but ending up generating little actual impact. That’s now changing, thanks to advancements in AI like deep learning, in which software programs learn to perform sometimes complex tasks without active oversight from humans.

Deep learning algorithms have been powering self-driving cars and making quick progress in tasks like facial recognition. Now these innovations are beginning to find their way into the daily lives of consumers as well.

“For retail companies that want to compete and differentiate their sales from competitors, retail is a hotbed of analytics and machine learning,” says John Bates, a product manager with Adobe Marketing Cloud, which offers machine learning services in e-commerce and other industries. As with the early Web, travel and entertainment are also making early use of machine learning, according to Bates.

A spate of recent experiments and announcements underscore the trend in e-commerce. One notable example is Pinterest Lens, a Shazam-like service that conducts visual searches based on items in the everyday world. Just point your camera at, say, a piece of furniture or item of clothing and Lens can help you find it online.

Lens builds on earlier Pinterest innovations like Related Pins and Guided Search — both based on the idea that you sometimes don’t know what you’re looking for until you see it — as well as a visual search tool that can find similar images inside the billions of pins that Pinterest has collected. Related pins are served up according to a similarity score that Pinterest’s algorithms assign between images.

Lens expands on that earlier search tool beyond images to include things in the real world. Image-detection programs identify an object and visual search digs up similar images, making it easier to buy a coveted item online. The potential for this kind of product-search innovation is interesting: You can search for things that won’t fit in a standard search box, and more tightly connect things found offline with those found online.

“For shopping specifically, improvements to online discovery means new ways to find products you’re interested in but may not have the words for,” says Andrew Zhai, an engineer working on Pinterest’s visual search. “Visual discovery gives people a way to discover new brands and ways of styling that they never knew existed.”

Other e-commerce sites are also adopting deep learning to help shoppers more easily find what they seek. Gilt deploys it to search for similar items of clothing with different features like a longer sleeve or a different cut. Etsy bought Blackbird Technologies last fall to apply the firm’s image-recognition and natural-language processing to its search function.

And notably, Amazon is planning to use the AI technology it offers on its Web Services in its new Amazon Go grocery stores. The company is operating only one store in Seattle, but Chief Financial Officer Brian Olsavsky said during a February earnings call that “it’s using some of the same technologies you would see in self-driving cars: computer vision, sensor, fusion, deep learning.”

Adobe, meanwhile, is taking things a step further by letting people create images of their desired products. Working with researchers at UC Berkeley, Adobe developed an image-editing tool that can turn a crude sketch of a handbag or shoe into a photorealistic and easily tweakable image. The tool also draws on a deep database of related images to turn sketches into pictures.

Adobe’s marketing tools are also incorporating deep learning into offerings for retailers, using AI to predict customer behavior. Shoppers can choose to receive suggestions based on their shopping lists – a belt that matches a pair of pants, painting supplies to help with a can of paint, a wine paired to a dinner recipe. Programs can subtly nudge people along when they are making a big-ticket purchase online but are not ready to hit the buy button.

Subtlety is a key part of these AI-powered marketing tools. People can grow alienated if they feel retailers are snooping on their behaviors or if it comes across as a hard sell. Adobe’s AI learns from past behavior as well as trial and error to learn how to make a gentle nudge without being too pushy.

“That’s a bit of the art and science behind deep learning,” says Bates. “But that’s where a lot of these signals can be built into the algorithms. If it creates an unnatural signal or puts someone off, it can be built into the training itself.”

While deep learning is becoming a part of the retail experience, it’s happening in fits and starts, as Facebook found with chatbots. Touted as a tool that could automate customer-service functions and deepen human engagement, chatbots were added to Facebook Messenger, with more than 11,000 of them available last year. But last week, Facebook scaled back its chatbot ambitions after they clocked a 70% failure rate.

As with the early days of the Web, there remains much work to do before deep learning can be seamlessly integrated into the daily lives of consumers. Compared to expectations of even a few years ago, though, things are a lot farther along than many expected. And that suggests Silicon Valley may again be ready to change how we shop.

How Artificial Intelligence Is Quietly Changing How You Shop Online

4 Models for Using AI to Make Decisions

4 Models for Using AI to Make Decisions

Charismatic CEOs enjoy leading and inspiring people, so they don’t like delegating critical business decisions to smart algorithms. Who wants clever code bossing them around? But that future’s already arrived. At some of the world’s most successful enterprises — Google, Netflix, Amazon, Alibaba, Facebook — autonomous algorithms, not talented managers, increasingly get the last word. Elite MBAs (Management by Algorithm) are the new normal.

Executives dedicated to data-driven excellence accept the reality that smart algorithms need greater autonomy to succeed. Empowering algorithms is now as organizationally important as empowering people. But without clear lines of authority and accountability, dual empowerment guarantees perpetual conflict between human and artificial intelligence.

Computational autonomy requires that C-suites revisit the hows and whys of delegation. CEOs need to clarify when talented humans must defer to algorithmic judgment. That’s hard. The most painful board conversations that I hear about machine learning revolve around how much power and authority super-smart software should have. Executives who wouldn’t hesitate to automate a factory now flinch at the prospect of deep-learning algorithms dictating their sales strategies and capex. The implications of success scare them more than the risk of failure.

“Does this mean that all our procurement bids will be determined by machine?” asked one incredulous CEO of a multibillion euro business unit. Yes, that’s exactly what it meant. His group’s data science, procurement, and supply chain teams crafted algorithmic ensembles that, by all measures and simulations, would save hundreds of millions. Even better, they would respond 10 times faster to market moves than existing processes while requiring minimal human intervention. Top management would have to trust its computationally brilliant bidding software. That was the challenge. But the CEO wouldn’t — or couldn’t — pull the autonomy trigger.

“You need a Chief AI Officer,” Baidu chief scientist Andrew Ng told Fortune at January’s Consumer Electronics Show. (He explained why he thinks so in a recent HBR article.) Perhaps. But CEOs serious about confronting autonomy opportunity and risk should consider four proven organizational options. These distinct approaches enjoy demonstrable real-world success. The bad news: Petabytes of new data and algorithmic innovation assure that “autonomy creep” will relentlessly challenge human oversight from within.

The Autonomous/Autonomy Advisor

McKinsey, Bain, and BCG are the management models here. Autonomous algorithms are seen and treated as the best strategic advisors you’ll ever have, but they’re ones that’ll never go away. They’re constantly driving data-driven reviews and making recommendations. They both take initiative on what to analyze and brief top management with what they find. But only the human oversight committee approves what gets “autonomized” and how it is implemented.

In theory, the organizational challenges of algorithmic autonomy map perfectly to which processes or systems are being made autonomous. In reality, “handoffs” and transitions prove to be significant operational problems. The top-down approach invariably creates interpersonal and inter-process frictions. At one American retailer, an autonomous ensemble of algorithms replaced the entire merchandising department. Top management told store managers and staff to honor requests and obey directives from their new “colleagues”; the resentment and resistance were palpable. Audit software and human monitors were soon installed to assure compliance.

In this model, data scientists are interlocutors and ambassadors between the autonomy oversight committee and the targets of implementation. They frequently find the technologies are less of a hassle than the people. They typically become the punching bags and shock absorbers for both sides. They’re the ones tasked with blocking efforts to game the algorithms. Their loyalty and accountability belongs to top management.

The Autonomous Outsourcer

“Accenturazon” — part Accenture, part Amazon Web Services — is the managerial model here. Business process outsourcing becomes business process algorithmization. The same sensibilities and economic opportunities that make outsourcing appealing become managerial principles for computational autonomy.

That means you need crystal-clear descriptions and partitioning of both tasks to be performed and desired deliverables. Ambiguity is the enemy; crisply defined service level agreements and explicit KPI accountability are essential. Process and decision owners determine the resource allocations and whether autonomy should lead to greater innovation, optimization, or both. Predictability and reliability matter most, and autonomy is a means to that end.

As with traditional outsourcing, flexibility, responsiveness, and interoperability invariably prove problematic. The emphasis on defined deliverables subverts initiatives that might lead to autonomous-driven new value creation or opportunity exploration. The enterprise builds up a superior portfolio of effective autonomous ensembles but little synergy between them. Smarter C-suites architect their autonomous Accenturazonic initiatives with interoperability in mind.

Data scientists in business process algorithmization scenarios are project managers. They bring technical coherence and consistency to SLAs while defining quality standards for data and algorithms alike. They support the decision and process owners responsible for autonomy-enabled outcomes.

World-Class Challenging/Challenged Autonomous Employee

Even the most beautiful of minds can come with intrinsic limitations, and in that way algorithms resemble eccentric geniuses. Can typical managers and employees effectively collaborate with undeniably brilliant but constrained autonomous entities? In this enterprise environment, smart software is seeded wherever computational autonomy can measurably supplement, or supplant, desired outcomes. The firm effectively trains its people to hire and work with the world’s best and brightest algorithms.

The software is treated as a valued and valuable colleague that, more often than not, comes up with a right answer, if not the best one. Versions of this are ongoing at companies such as Netflix and Alibaba. But I cannot speak too highly of Steve Levy’s superb Backchannel discussion of how Google has committed to becoming a “machine learning first” enterprise.

“The machine learning model is not a static piece of code  —  you’re constantly feeding it data,” says one Google engineer. “We are constantly updating the models and learning, adding more data, and tweaking how we’re going to make predictions. It feels like a living, breathing thing. It’s a different kind of engineering.”

Comingling person/machine autonomy necessarily blurs organizational accountability. In such fast-changing learning environments, project and program managers can’t always know whether they will get better results from retraining people or retraining algorithms. That said, a culture of cocreation and collaboration becomes the only way to succeed.

Data scientists here facilitate. They’re analogous to autonomous resources, as opposed to human resources, departments. They do things like write chatbots and adopt Alexa-like interfaces to make collaboration and collegiality simpler and easier. They look to minimize discrimination, favoritism, and tension in person/machine relationships. C-suites depend on them to understand the massive cultural transformation pervasive autonomy means.

All-In Autonomy

Renaissance Technologies and other, even more secretive investment funds are the management models here. These organizations are fully committed to letting algorithmic autonomy take the enterprise to new frontiers of innovation, profitability, and risk. Their results should humble those who privilege human agency. Human leadership defers to demonstrable algorithmic power.

One quant designer at a New York hedge fund (that trades more in a week than a Fortune 250 company makes in a year) confided: “It took years for us to trust the algorithms enough to resist the temptation to override them….There are still [occasional] trades we won’t make and [not doing them] almost always costs us money.”

Firms look to leverage, amplify, and network autonomy into self-sustaining competitive advantage. They use machine learning software to better train machine learning software. Machine learning algorithms stress-test and risk-manage other machine learning algorithms.

Autonomy is both the organizational and the operational center of gravity for innovation and growth. People are hired and fired based on their abilities to push the algorithmic boundaries of successful autonomy.

Leadership in these organization demands humility and a willingness to convert trust in numbers into acts of faith. Academic computational finance researchers and fund managers alike tell me their machines frequently make trades and investments that the humans literally and cognitively do not understand. One of the hottest research areas in deep learning is crafting meta-intelligence software that generate rationales and narratives for explaining data-driven machine decision to humans.

Risk management and the imperative to acquire accessible human understanding of complex autonomy dominates data science for all-in enterprises.

Admittedly, these four managerial models deliberately anthropomorphize autonomous algorithms. That is, the software is treated not as inanimate lines of code but as beings with some sort of measurable and accountable agency. In each model C-suites rightly push for greater transparency and accessibility into what makes them tick. Greater oversight will lead to greater insight as algorithmic autonomy capabilities advance.

CEOs and their boards need to monitor that closely. They also need to promote use cases, simulations, and scenarios to stress-test the boundary conditions for their autonomous ensembles.

CEOs and executive leaderships should be wary of mashing up or hybridizing these separate approaches. The key to making them work is to build in accountability, responsibility, and outcomes from the beginning. There must be clarity around direction, delegation, and deference.

While that maxim is based on anecdotal observation and participation, not statistical analysis, never underestimate how radical shifts in organizational power and influence can threaten self-esteem and subvert otherwise professional behavior. That’s why CEOs should worry less about bringing autonomy to heel than making it a powerful source and force for competitive advantage.

Without question, their smartest competitors will be data-driven autonomous algorithms.

4 Models for Using AI to Make Decisions

Why nature is our best guide for understanding artificial intelligence

Why nature is our best guide for understanding artificial intelligence

In living organisms, evolution is a multi-generational process where mutations in genes are dropped and added. Well-adapted organisms survive and those less fortunate go extinct. This is Natural Selection. Resilience is great, but if you don’t grow gills in time for the flood, then tough luck.

Engineering, on the other hand, is a deliberate process with reliable steps designed to reach a stated objective. With the emergence of artificial intelligence, we are beginning to see the convergence of evolution and engineering as machine learning algorithms begin to evolve.

For the sake of our comparison (natural evolution to machine evolution), let’s consider data and how it is normalized as “the environment” and the training process as “Natural Selection.” The training process can be supervised or unsupervised learning, reinforcement learning, clustering, decision trees or different methods of “deep learning.”

Much like natural evolution, different organisms solve for the same problem differently depending on their environment, but ultimately reach the same outcome. Sharks and dolphins wound up with similar mechanisms to survive despite starting from completely different beginnings. In technology, we see similar patterns. The K-means clustering algorithm, a technique often used for image segmentation, for example, ingests essentially unlabeled inputs (usually images) and coherently grouped clusters are produced until a desired grouping is reached. If you gave 10 people the same data set and asked them to solve the same problem using different algorithms, it’s possible that they could each take a different approach and get the same outcome. Problem solving in nature and machines are, in a sense, quite similar.pasted-image-0

Why does this matter for companies?

As machine learning techniques find their way into commercial applications, businesses are faced with the challenge of developing strategies to implement this technology safely and efficiently.

Historically speaking, technologists have often looked to nature for inspiration. Here are a few ways businesses can use evolution to understand the potential implications of artificial intelligence:

  • Divergent evolution: It’s harder to move into adjacency even in seemingly related data sets than at first glance. Just because you have ImageNet to train on object recognition doesn’t mean you master video recognition or facial recognition.
  • Convergent evolution: Always be on the lookout for what are fundamentally the same problems even if it’s a different data set. Think about how Google uses search query data to build a better spell checker. They keep track of what users are querying, and when they notice that millions of others have spelled something differently, they’ll suggest that you do the same. A happy accident.
  • Predator and prey or parasite and hosts co-evolving: Interesting things can happen if two AIs co-evolve. Cybersecurity companies (like Cylance and Bromium) are developing machine learning solutions that are constantly training their systems to detect new threats.

There are a handful of brilliant AI companies helping us work more efficiently, (we have X.aihelping us manage our hectic lives, Diffbot helping us intelligently organize the web, etc.), but these applications are still in their infancy and there needs to be a fundamental shift in how we anticipate their arrival. Perhaps it’s best to place them into the context of a phenomena that we already understand — evolution.

There’s great opportunity in AI, and natural evolution provides a framework for us to study and prepare for the future of machine evolution. In the meantime, it’s important that company leadership seriously consider their strategy for AI and invest in the requisite talent and infrastructure to turn their data into transformative solutions.

Why nature is our best guide for understanding artificial intelligence

Setting the AI record straight (for now)

Setting the AI record straight (for now)

Artificial intelligence is changing at a breathtaking pace the way we interact with our devices, friends and data. It promises to revolutionize how we work, shop, socialize, date, bank, heal, navigate… how we live.

Setting the AI record straight (for now)

Artificial intelligence is changing at a breathtaking pace the way we interact with our devices, friends and data. It promises to revolutionize how we work, shop, socialize, date, bank, heal, navigate… how we live.

Yet here I am in the engine room of the revolution, getting increasingly frustrated with the prevalent AI conversations and memes. We’re painting a partial, bipolar, even misleading picture. “We” sometimes includes me, and it includes journalists, pundits, analysts and even practitioners.

What do I mean? Specifically, we highlight the highs (AlphaGo beats world champion!) and lambast the lows (Chatbot Tay turns racist and hateful!), while brushing off smaller (but meaningful) advances. We keep giving air time to robot-takeover fears, while failing to evangelize and thus guide the real value of AI today.

As humans, we’re drawn to the dramatic, but as an industry, we have a responsibility to get everyone on the same, grounded-in-reality page. So forgive me for sharpening my proverbial axe. Here are three specific AI topics that need a reboot.

Pervasive if mundane

We are all rightly in awe of the splashy, sexy and sometimes scary AI concepts we read about in the headlines. The end of apps, the end of screens, autonomous vehicles,autonomous weapon systems… Though not difficult to imagine, many of these scenarios are several years out. But less sensational AI models are already here, now, improving our everyday experiences. They’re great. And we should talk about them more.

A broad array of companies are quietly innovating on top of AI technologies — and integrating the resulting innovations into services today. Their models are showing us the right products (e.g. Expedia, Getty Images), serving smarter ads (e.g. GumGum, Pinterest), organizing our content (e.g. GoPro, Microsoft), guiding investments (e.g. Sentient), talking to us in the words we understand (e.g. DoCoMo, IBM Watson) and more. Each of these is one of our customers, and each of them is filled with smart people working hard to improve their products using AI. Many of them work on agile software teams. Over time, their incremental progress will revolutionize technology. But don’t expect to wake up tomorrow as Bruce Willis in “Surrogates.”

Not just for the big platforms anymore

In this AI Spring, we’re seeing big news almost daily, much of it brought to us by the tech behemoths we all know and love. Google, Amazon, Microsoft, IBM, Facebook, Apple and others are constantly making waves with new AI projects and developments. And these companies are paying big money for AI startups (recent examples include Turi andNervana). It’s war.

It’s not just who has the models, it’s about who applies the right data.

But AIs are not just for the big guys. Why? Three reasons. First, the affordability of massive computing power. Indeed, the commoditization of compute power is a large part of why the big cloud platforms want you developing models on top of Azure, AWS, Watson and GCP, and why Intel and NVIDIA want you doing it on their silicon. Second, there are more ML- and AI-trained engineers than ever before, taking advantage of decades of science advancing the underlying algorithms.

Third, largely behind the scenes, product teams from startups to retailers and systems integrators are applying others’ AI platforms using their own proprietary training data. The most sophisticated algorithms are only able to see, listen, talk and “think” like humans if the right humans are training and re-training them as our fickle human opinions evolve. Nobody understands shoes like Shoes.com, surfing videos like GoPro, top-shelf stock imagery like Getty — and their customers.

Sometimes the ability to leverage cognitive computing is less about the underlying algorithms and more about how they are tuned and productized. Mark my words: It’s not just who has the models, it’s about who applies the right data.

A means, not an end

Lastly, can we please stop evaluating AIs based on whether they are saving us from cancer or threatening humanity? These questions bait valuable conversations, but miss the bigger, more important “middle.” I love telling my mom that we train AIs. But AIs are not the goal. When we all go to that tech-loving part of our brain, we forget that AIs are the meansto improved experiences.

We all accept that a picture is “worth a thousand words.” We also know that most of our communication is non-verbal (just contrast your last date, or in-person interview, with your last telecon call). So how much are 100 million images worth? How about terabytes of CT scans of your loved one’s brain tumor? And what about 24 hours of video, at 30 fps? Well, of course, that depends on what’s in the data — and, more importantly, what we can do with it.

AIs hold the key to unlocking the value buried in the massive, exponentially growing sets of unstructured content with which we each interact daily.

As such, I humbly propose that we measure the value of an AI based on its ability to contribute positive utility to people. We all know that at the end of the day, it’s the people who matter in our lives. Just the same, we need to remember that AIs should be measured by their ability to understand, interact with and improve people’s lives. This utilitarian framework offers hundreds of years of insights from philosophers, economists and anthropologists.

The new nukes

AIs are pervasive, and are creating the power to spread good and bad. Time and time again, people have managed to guide innovations for the good of humanity. In some ways, AIs remind me of nuclear technologies. Handled poorly, they could destroy the world. But they haven’t. In fact, we have avoided another world war thanks in part to mutual assured destruction, and atomic know-how has quietly revolutionized healthcare, energy, robotics, physics and aerospace.

Yes, the extremes are scary, but they are only part of the story. If we are to harness AIs for universal good, we need to be more realistic about their daily implications and the net benefits thereof. Less “duck and cover” and more comprehensive conversation.

TC

Setting the AI record straight (for now)

Google wants to improve artificial intelligence to prevent robot screw-ups

Google wants to improve artificial intelligence to prevent robot screw-ups

More and more artificial intelligence will soon enter our lives. And Google would very much like its AI systems to be front and center.

That’s why the company is putting resources into making sure AI systems don’t go off the rails.

Last month, Google researchers teamed with scientists from two universities and OpenAI, the consortium backed by Elon Musk, to investigate potential pitfalls for AI that may arrive in the not-too-distant future.

The paper walks through five different ways a household cleaning robot could encounter safety problems, then details some math to avoid these problems. (Google hasn’t released plans for this kind of robot, but one of its robotics projects — no, not the frightening humanoid one — fits the bill.)

Here are the five issues that Google flags:

  • Whoops: If you train a robot to clean your house as quickly as possible, how do you ensure that it doesn’t knock over the lamp?
  • “Reward hacking”: What if the robots cheat? Google wonders aloud if a robot could, say, sweep a mess under a rug or disable its vision (nothing to see here!).
  • Reasonable human feedback: Can robots figure out how to improvise on the job — like toss a candy wrapper, but not a cellphone — without constant human oversight?
  • Physical safety: We probably don’t want our robots shoving wet mops in sockets.
  • Different environments: Once the robot learns to clean a room in a house, can the same robot clean an office?

It’s important research — before these types of robots arrive en masse, they will surely need to pass certain safety requirements. And Google needs to prove that AI is a force for good. The Google paper also nods to other important research topics like ensuring that AI doesn’t fall into “malicious” hands or “attack or harm people.”

Google wants to improve artificial intelligence to prevent robot screw-ups

Can artificial intelligence wipe out cyber terror?

Can artificial intelligence wipe out cyber terror?

Slowly but surely, cyber security is evolving from the days of castles and moats into the modern era of software driven business. In the 1990s, after several failed attempts to build secure operating systems, the predominant security model became the network-perimeter security model enforced by firewalls. The way it works is clear: Machines on the inside of the firewall were trusted, and anything on the outside was untrusted. This castle-and-moat approach failed almost as quickly as it began, because holes in the wall had to be created to allow emerging internet services like mNews, email and web traffic through.

Artificial intelligence will replace large teams of tier-1 SOC analysts who today stare at endless streams of threat alerts.

With a security wall that quickly became like Swiss cheese, machines on both sides were still vulnerable to infection and the antivirus industry emerged to protect them. The model for antivirus then and now is to capture an infection, create a signature, and then distribute it widely to “immunize” other machines from getting infected by the same malware. This worked for vaccines, so why not try for cyber security?

Fast-forward to 2016, and the security industry hasn’t changed much. The large security companies still pitch the castle-and-moat model of security — firewalls and signature-based detection — even though employees work outside the perimeter as much as inside. And in spite of the fact that most attacks today use one-and-done exploit kits, never reusing the same malware again. In other words, the modern work force coupled with modern threats has rendered traditional security techniques obsolete.

Software is eating security

While most enterprises today still employ these dated security techniques, a new model of security based on artificial intelligence (AI) is beginning to take root in organizations with advanced security programs. Necessity is the mother of invention, and the necessity for AI in security became obvious when three phenomena emerged: (1) The failure of signature-based techniques to stop current threats; (2) the voluminous amounts of security threat data; and (3) the scalability challenges in addressing security threat data with people.

“Software is eating the world,” the noted venture capitalist Marc Andreessen famously said in 2011 about such obvious examples as Amazon, Uber and Airbnb disrupting traditional retail and consumer businesses. The security industry is ripe for the same kind of disruption in the enterprise space, and ultimately in the consumer product space. Artificial intelligence will replace large teams of tier-1 SOC analysts who today stare at endless streams of threat alerts. Machines are far better than humans at processing vast amounts of data and finding the proverbial needle in the haystack.

Deep learning

Artificial Intelligence is experiencing a resurgence in commercial interest because of breakthroughs with deep learning neural networks solving practical problems. We’ve all heard about IBM’s Watson winning at “Jeopardy,” or making difficult medical diagnoses by leveraging artificial intelligence. What is less well known is that Watson has recently undergone a major deep learning upgrade, as well, allowing it to translate to and from many languages, as well as perform text to speech and speech to text operations flawlessly.

Many of us interact with deep learning algorithms unwittingly when we see TV show and movie recommendations on Netflix based on what we’ve viewed previously or when your Mac properly identifies everyone in a picture uploaded from your phone. Or when we ask Alexa a question and Amazon Echo gives an intelligent response — likewise for Cortana and Siri. And one of the most hotly debated topics in machine learning these days is self-driving cars, like Tesla’s amazing Model S.

Deep learning allows a machine to think more like a human. For instance, a child can easily distinguish a dog from a cat. But to a machine, a dog is just a set of pixels and so is a cat, which makes the process of distinguishing them very hard for a machine. Deep learning algorithms can train on millions of pictures of cats and dogs so that when your in-house security camera sees the dog in your house, it will know that it was Rover, not Garfield, who knocked over the vase.

With deep learning, today’s next-generation security products can identify and kill malware as fast as the bad guys can create it.

The power of deep learning becomes clear when you consider the vast speed and processing power of modern computers. For instance, it takes a child a few years to learn the difference between a house cat and a dog. And if that child grew up to be a cat “expert,” it would take Gladwell’s 10,000 hours to become a feline whisperer. The amount of time it takes to expose a human to all of the training data necessary to classify animals with near perfection is long. In contrast, a deep learning algorithm paired with elastic cloud computing resources can consume hundreds of millions of samples of training data in hours, to create a neural network classifier so accurate and so fast that it would outperform even the most highly trained human experts.

What’s more fascinating than this new technology allowing machines to think like a human, is allowing machines to act like a human. Since the 1950s, we’ve been fascinated with the notion that robots might one day be able to think, act and interact with us as our equals. With advances in deep learning, we’re one giant step closer to that reality. Take the Google Brain Team’s DeepDream research, for instance, which shows that machines trained in deep learning can create beautiful pieces of art, in a bizarre form of psychedelic machine “dreaming.” For the first time, we see incredible creativity from machines because of deep learning, as well as the ability to make decisions with incredible accuracy.

Because of this ability to make classification decisions with incredible accuracy, deep learning is leading a renaissance in security technologies by using the technology to identify unknown malware from benign programs. Like the examples above, this is being done by training the deep learning neural networks on tens of millions of variants of malware, as well as on a representative sample of known benign programs.

The results are industry-changing, because unlike legacy security products that provided protection either through prior knowledge of a threat (signature-based) or via segmentation and separation, today’s next-generation security products can identify and kill malware as fast as the bad guys can create it. Imagine a world where security technologies actually enable more sharing rather than less, and allow a more open approach to data access rather than restrictive. This is the direction deep learning is allowing us to go.

Are you ready?

Disruption is clearly coming to the security space. The market has been waiting for better technology that can keep pace with the fast-evolving adversarial threat. Breakthroughs in deep learning artificial neural networks are now stopping attacks previously unseen in real time before they even have a chance to run. It’s time to get on-board with a new generation of technology that is disrupting traditional castle-and-moat security models.


Anup Ghosh is the founder and CEO of Invincea Inc., and is a holder of seven patents in next-generation security. Reach him @AnupGhosh_.

Can artificial intelligence wipe out cyber terror?

We need to talk about AI and access to publicly funded data-sets

We need to talk about AI and access to publicly funded data-sets

For more than a decade the company formerly known as Google, latterly rebranded Alphabet to illustrate the full breadth of its A to Z business ambitions, has engineered an annually increasing revenue generating empire which last year pulled in ~$75 billion. And it’s done this mostly by mining user data for ad targeting intel.

Slice it and dice it how you like but Google’s business engine needs data like the human body needs oxygen. Most of its products are thus designed to remove friction to accessing more user data; whether it’s free search, free email, free cloud storage, free document editing tools, free messaging apps, a fuzzy social network that no one loves but which is somehow still hanging around, free maps, a mobile OS platform that OEMs can load onto smartphone hardware without paying a license fee… Most of what Google builds it opens to all comers to keep the data pouring in. The bits and bytes must flow.

The trade off for consumers handing over data is of course access to a particular Google service without any up front cost. Or getting to buy a cheaper piece of hardware than they might otherwise be able to. Or the convenience of using a dominant digital service. Of course they are ‘paying’ with their data, but few will think of it that way. It’s an abstract idea for starters, and a personal cost that’s far harder to quantify given how unclear it is what Google really does with the data it gathers and processes in its algorithmic black boxes.

Google certainly isn’t spelling that out. Rather it makes noises about the benefits of it knowing more about you (savvier virtual assistants, more powerful photo search and so on). And without explicit knowledge of what the trade-off entails — coupled with noisy PR about the convenience of data-powered services — most consumers will simply shrug and carry on handing over the keys to their lives. This is the momentum that fuels Mountain View’s ad-targeting empire. The more it knows about you, the richer it bets it can get.

You can dislike Google’s business model but you can also argue that consumers do (in general) have a choice about whether to use its services. Albeit in markets where the company has a defacto monopoly there may be doubt about how much choice people really have. Not least if the company is found to have been abusing a dominant position by demoting alternatives to its services in its search results (Google is facing just such antitrust claims in Europe, where it has a hugely dominant marketshare in search, for example).

Another caveat is that Google has worked to join up more personal data dots, undermining how much control users have over how they share data with the centralizing Alphabet entity — by, for example, consolidating the privacy policies of multiple products to enable it to flesh out its understanding of each user by cross-referencing their usage of different services. That collapsing of prior partitions between products has also caused Google headaches with European data protection regulators. And contributed to a caricature of it as a vampire octopus with masses of tentacles all maneuvering to feed data back into a single, hungry maw.

But if you think Google has a controversial reputation at this point in its business evolution, buckle up because things are really stepping up a gear.

The Google/Alphabet octopus, via its artificially intelligent DeepMind tentacle, is being granted access to public healthcare data. Lots and lots of healthcare data. Now personal data doesn’t really get more sensitive than people’s medical records. And these highly sensitive bits and bytes are now being sucked towards Google’s algorithmic core — albeit indirectly, via the DeepMind division, which so far this year has two publicly announced data-sharing collaborations with the UK’s National Health Service (NHS).

The public data in question is tied to the two specific projects. But the most recent of these collaborations, with Moorfields Eye Hospital NHS Trust in London, entails DeepMind applying machine learning to the data. Which is a key development. Because, as New Scientist noted this week, Google will be keeping any AI models DeepMind is able to build off of this public data-set. The trained models are effectively its payment in this trade — given it’s not charging the NHS for its services.

So yes, this is another Google freebie. And the cash-strapped, publicly (under)funded NHS has obviously leapt at the chance of a free-at-the-point-of-use high tech partner who might, in time, help improve healthcare outcomes for patients. So it’s granting the commercial giant access to patients’ data.

And while we are told the first NHS DeepMind collaboration, announced back in February with the Royal Free Hospital Trust in London, does not currently involve any AI component, the five-year strategic partnership between the pair does include a wide ranging memorandum of understanding in which DeepMind states its hope to also conduct machine learning research on Royal Free data-sets. So advancing AI is the clear objective for DeepMind’s NHS engagement, as you’d expect. It is a machine learning specialist. And its learning algorithms need the lifeblood of data in order to develop and thrive.

Now we’re all, as individuals, used to getting Google freebies in exchange for sharing some of our data. But the thing is, the data trade off here — with the publicly funded NHS — is a rather different beast. Because the people whose personal data is being pumped into Google-owned databanks are not being asked for their individual consent to the exchange.

Patient consent has not been sought in either of the current NHS collaborations. In the Moorfields project, where the data is being anonymized (or pseudonymized), NHS information governance rules allow for data to be shared for medical research purposes without obtaining patient consent (although NHS patients can opt out of supplying their data to all research projects) — so long as the relevant Health Research Authority clears the project. And DeepMind has applied to be cleared access in this case.

In the first collaboration, with the Royal Free, where DeepMind is helping co-design an app to detect acute kidney injury, the patient data being supplied is not anonymized or pseudonymized. In fact full patient medical records are being shared with the company — likely millions of people’s medical records, given it’s getting real-time data across the Trust’s three hospitals, along with five years’ worth of historical inpatient data.

In that case patient consent has not been sought because the Royal Free argues consent can be implied as it claims the app is for “direct patient care”, rather than being a medical research project (or another classification, such as indirect patient care). There has been controversy over that definition — with health data privacy groups disputing the classification of the project and questioning why DeepMind has been handed access to so much identifiable patient data. Regulators have also stepped in after the fact to take a look at the project’s parameters.

Whatever the upshot of those complaints, it’s fair to say NHS rules on information governance are not an exact science, and do involve interpretation by individual NHS Trusts. There is no definitive set of NHS data-sharing commandments to point to to definitely denounce the scope of the arrangement. The best we have is a series of principles developed by the NHS’ national data guardian, Fiona Caldicott. And, perhaps, our public sense of right and wrong.

But what is absolutely crystal clear is that millions of NHS patients’ medical histories are being traded with DeepMind in exchange for some free services. And none of these people have been asked if they agree with the specific trade.

No one has been asked if they think it’s a fair exchange.

The NHS, which launched in 1948, is a free-at-the-point of use public healthcare service for all UK residents — currently that’s around 65 million people. It’s a vast repository of medical data so it’s not at all hard to see why Google is interested. Here lies data of unprecedented value. And not for the relatively crude business of profiling consumers via their digital likes and dislikes; but for far more valuable matters, both in societal and business terms. There could be considerable future revenue-generating opportunities if DeepMind’s AI models end up being able to automate and/or improve complex diagnostic and healthcare challenges, for example. And if the models prove effective they could end up positively impacting healthcare outcomes — although we don’t know exactly who would benefit at this point because we don’t know what pricing structure Google might impose on any commercial application of its AI models.

One thing is clear: large data-sets are the lifeblood of robust machine learning algorithms. In the Moorfields case, DeepMind is getting around a million eye scans to train its machine learning models. And while those eye scans will technically be handed back at the end of the project, any diagnostic intelligence they end up generating will remain in Google’s hands.

The company admits as much in a research outline of the project, though it steers the focus away from these trained algorithms and back to the original data-set (whose value the algorithms will now have absorbed and implicitly contain):

The algorithms developed during the study will not be destroyed. Google DeepMind Health knows of no way to recreate the patient images transferred from the algorithms developed. No patient identifiable data will be included in the algorithms.

DeepMind says it will be publishing “results” of the Moorfields research in academic literature. But it does not say it will be open sourcing any AI models it is able to train off of the publicly funded data.

Which means that data might well end up fueling the future profits of one of the world’s wealthiest technology companies. Instead of that value remaining in the hands of the public, whose data it is.

And not just that — early access to large amounts of valuable taxpayer-funded data could potentially lock in massive commercial advantage for Google in healthcare. Which is perhaps the single most important sector there is, given it affects everyone on the planet. If you don’t think Google has designed on becoming the world’s medic, why do you think it’s doing things like this?

Google will argue that the potential social benefits of algorithmically improved healthcare outcomes are worth this trade off of giving it advantageous access to the locked medicine cabinet where the really powerful data is kept.

But that detracts from the wider point: if valuable public data-sets can create really powerful benefits, shouldn’t that value remain in public hands?

Or shouldn’t we at least be asking if we have a public duty to disseminate the value of publicly funded data as widely as possible?

And are we, as a society, comfortable with the trade off of a few free services — and some feel-good but fuzzy talk of future social good — for prematurely privatizing what could be our core IP?

Shouldn’t we, as the data creators, as the patients, at least be asked if we are comfortable with the terms of the trade?

Fiona Caldicott’s, the UK’s national data guardian, happened to publish her third review of how patient data is handled within the NHS just this week — and she urged a more extensive dialogue with the public about how their data is used. And a proper informed choice to opt in or out.

The old rules about information governance — which still talk in terms of shredding pieces of paper as a viable way to control access to data — have certainly not kept up with big data and machine learning. Stable doors and bolting horses spring to mind when you combine these old school data access rules with the learning and evolving character of advanced AI.

Access to data-sets is undoubtedly the core competitive advantage for AI builders because really good data is hard to come by and/or expensive to create. And that’s why Google is pushing so hard and fast to embed itself into the NHS.

You can’t blame the company for this healthcare data-grab. It’s just doing what successful commercial enterprises do: figuring out what the future looks like and plotting the fastest route to get there.

What’s less clear is why governments and public bodies find it so hard to see the value locked up in the publicly funded data-sets they control.

Or rather why they fail to come up with effective structures to support maintaining public ownership of public assets; to distribute benefits equally, rather than disproportionately rewarding the single, best-resourced, fastest-moving commercial entity that happens to have the slickest sales pitch. It’s almost as if the public sector is being encouraged to privatize yet another public resource… ehem

Inject a little more structured forward-thinking and public healthcare data could, for example, be contributed (with consent) to machine learning research departments in domestic universities so that AI models can be developed and tested ‘in house’, as it were, with public parents.

Instead we have the opposite prospect: public data assets stripped of their value by the commercial sector. And with zero guarantees that the algorithms of the future will be free at the point of use. Of course Google is going to aim to turn a profit on any healthcare AI models DeepMind creates. It’s not in the business of only giving away freebies.

So the really pressing question — roundly ignored by web consumers going about their daily Googling but perhaps moving into clearer focus, here and now, as commercial thirst to accelerate AI advancements is encouraging public sector bodies to over-hastily ink wide-ranging data-sharing arrangements — is what is the true cost of free?

And if we’ve inked the contracts before we even know the answer to that question won’t it be too late for us to haggle over the price?

Even DeepMind talks publicly about the need for new models of information governance and ethics to be put in place to properly oversee the coupling of AI with data…

Screen Shot 2016-07-09 at 12.38.45 PM

So we, the public, really need to get our act together and demand a debate about who should own the value locked up in our data. And preferably do so before we’ve handed over any more sets of keys.

We need to talk about AI and access to publicly funded data-sets

What’s Next for Artificial Intelligence

What’s Next for Artificial Intelligence

How do you teach a machine?

Yann LeCun, director of artificial-intelligence research at Facebook, on a curriculum for software

The traditional definition of artificial intelligence is the ability of machines to execute tasks and solve problems in ways normally attributed to humans. Some tasks that we consider simple—recognizing an object in a photo, driving a car—are incredibly complex for AI. Machines can surpass us when it comes to things like playing chess, but those machines are limited by the manual nature of their programming; a $30 gadget can beat us at a board game, but it can’t do—or learn to do—anything else.

This is where machine learning comes in. Show millions of cat photos to a machine, and it will hone its algorithms to improve at recognizing pictures of cats. Machine learning is the basis on which all large Internet companies are built, enabling them to rank responses to a search query, give suggestions and select the most relevant content for a given user.

Deep learning, modeled on the human brain, is infinitely more complex. Unlike machine learning, deep learning can teach machines to ignore all but the important characteristics of a sound or image—a hierarchical view of the world that accounts for infinite variety. It’s deep learning that opened the door to driverless cars, speech-recognition engines and medical-analysis systems that are sometimes better than expert radiologists at identifying tumors.

Despite these astonishing advances, we are a long way from machines that are as intelligent as humans—or even rats. So far, we’ve seen only 5% of what AI can do.

What’s Next for Artificial Intelligence

%d bloggers like this: