Michael a Osborne.Indd

An interview with Michael A Osborne Associate professor at the University of Oxford

The New Innovation Wave: Machine Learning and Al

Transform to the power of digital Michael A Osborne

whereas at the end of the 20th century, of some of the most quintessential it was less than 2%. There was quite human characteristics. The boundaries a seismic change in the makeup between all these different fields are of employment due to technology. not very clear. However, you might However, none of these changes associate computer vision, language actually affected unemployment much; processing and speech recognition as the unemployment rate was about 5% other subsets of artificial intelligence. at the beginning of the 20th century and was still 5% at the end of the 20th century. Could you give us some examples of current applications of artificial intelligence and machine learning in business or non-business Roman Emperor environments? Michael A Osborne A great example of this is the retail Associate professor Tiberius, had business, where ‘recommendation at the University of Oxford [an inventor] sentenced engines’ are being used extensively. Take Amazon as example. It recommends to death, fearing the products to millions of customers on the loss of jobs and creative basis of historic data. Algorithms can process data, characterize the spending destruction. patterns of millions of customers, identify latent trends and patterns, and identify clusters and communities for recommendation. Your latest research with Professor Is the current digital innovation Carl Benedikt Frey and Citi on wave different from the previous innovation shows that throughout innovation waves? history powerful interests have Yes, in the past technology has largely hindered innovation. Is history In the past technology just substituted the human muscle, repeating itself? whereas now technologies can has largely just History might be repeating itself. increasingly substitute human brain labor Technological progress has always or human cognitive work. If machines substituted the human created significant disruptions that can substitute for cognitive labor, as muscle, whereas now powerful interests tried to combat to they have done in the past for physical maintain the status quo. In Ancient labor, it’s not very clear to what extent technologies can Rome, the great author, Pliny the Elder, there might be any demand remaining records the story of an inventor who for human labor. Another thing that is increasingly substitute discovered a way to manufacture glass different this time is the accelerating rate human cognitive work. that was unbreakable. Hoping to receive of technological change. It is undeniable a reward, the inventor presented his that we really are making quite dramatic invention to Roman Emperor Tiberius. leaps forwards in designing intelligent Self-driving cars are another application Unfortunately for the inventor, Tiberius algorithms that might be able to where machine learning is really playing had him sentenced to death, fearing the substitute for human workers. quite a pivotal role. Cars today have loss of jobs and creative destruction due sensors, onboard computing and safety to the new technology. Can you give us your definition of systems, and use intelligent algorithms The various innovation waves in history machine learning and artificial to render themselves autonomous. have led to dramatic upheavals in intelligence? These algorithms observe how you society but we have never seen large- drive using the onboard sensors and scale technological unemployment. For Machine learning is a subset of artificial quietly learn to travel from the current example, at the beginning of the 20th intelligence (AI) and can be defined as to desired positions on the basis of this century in the U.S., about 40% of our the study of algorithms that can learn data. employment was engaged in agriculture, and act. It’s about the reproduction Michael A Osborne

Speech recognition is one key application What are the most promising an enormous economic impact for taxi of machine learning. Voice recognition opportunities for both AI and and truck drivers. In the near term, self- software, such as Apple’s Siri, has been machine learning? driving technologies are going to have an driven by advances in machine learning. impact in the automation of professions, In our recent paper, “The Future of Deep learning is becoming a mainstream such as forklift drivers, agricultural Employment”, Frey and I have identified technology for speech recognition and vehicle drivers, or mining vehicle drivers. occupations that we thought were the has successfully replaced traditional most susceptible to automation. These algorithms for speech recognition at are the kind of jobs that would be most an increasingly large scale. Historically, Are there sectors where you see easily automated within the near future people were trying to induce speech a lot of potential for machine using machine learning techniques. The recognition by recognizing speech learning? model predicts that most workers in features like words and phrases. transportation and logistics occupations, The retail sector, for example, offers But recently, algorithm designers together with the bulk of office and tremendous opportunities. We have have turned to massive data that has administrative support workers, and seen restaurants deploying tablets on become available for characterizing labor in production occupations, are tables. These tablets can be used to take human speech. If this trend continues, at risk. I think that in all these kinds of orders from customers. They can also we are likely to continue to move from areas, there will be an ever-increasing recommend customers a set of dessert the 95% accuracy rate that we’re able to degree of automation using machine options tailored to what customers achieve on our mobile devices today to learning algorithms. might want based on what they have nearly 99%. This would probably make ordered previously. A customer might these speech recognition techniques even have a profile that’s built up over completely widespread. multiple desserts in the same restaurant More or less anything or even a family of restaurants using the that does not require same software. So, you might get to the Machine learning is point at which this tablet can instantly one of the three know, as soon as you walk through a subset of artificial a door, what your usual order is, and bottlenecks – i.e. expedite the ordering of it. intelligence (AI) and can creativity, social Even in the traditional mall, the customer be defined as the study movement around the store can be intelligence and of algorithms that can closely monitored, indicating customer the requirement interest in products. The change here learn and act. is that these intelligent algorithms to manipulate can become more personalized and gather a lot more data about us as complex objects in You mentioned ‘deep learning’. individuals. Therefore, they can make Can you shed some light on this an unstructured recommendations to us that are more concept? directly targeted at what we want and environment – will what the store might actually want to “Deep learning” is the new big trend in promote. machine learning. It promises general, be potentially powerful, and fast machine learning, moving us one step closer to artificial intelligence. automatable. Can individual companies invest in Deep learning is a set of techniques, inspired artificial intelligence? by biological neural networks, for teaching More or less anything that does not The wonderful thing is that you don’t machines to find patterns and classify require one of the three bottlenecks necessarily need to be developing novel massive amounts of data. It tackles crisp and — i.e. creativity, social intelligence and artificial intelligence techniques and well-defined problems, such as classifying the requirement to manipulate complex machine learning algorithms to benefit images as to whether they contain a car or objects in an unstructured environment from the state-of-the-art. It’s relatively a book, for example. It’s worth noting that — will be potentially automatable in low cost, an individual corporation can deep learning is still not well-understood, or the near future. When self-driving cars just hire a data scientist or a machine provably robust, and that it is likely unsuitable become a reality, there is going to be learning PhD. There is this wealth of for safety-critical applications. Making Machines Learn

Artiﬁcial Intelligence and Machine Learning

What is machine learning? How big is machine learning? How hot is machine learning?

$300

A subset of artiﬁcial Research on AI constitutes Venture capitalists have

intelligence (AI). approximately 10% of all invested over $300 million Study of algorithms that computer science in AI startups in 2014, a 1 can learn and act. research today . twentyfold increase from 2010 2. It aims to reproduce some of the most quintessential human characteristics.

Working with machines

“We need to equip the At the same time, next generation of organizations need to protect workers with the themselves from deskilling – necessary skills that they loss of human skills due to will need for an economy intrusion of automation in that is increasingly occupations. automated”. - Michael A Osborne

No heavy investments needed – a wealth of open-source, state-of-the-art algorithms already available, it costs very little to reproduce an amazing algorithm and it beneﬁts everyone.

1Machine Intelligence Research Institute, “How Big is the Field of Artiﬁcial Intelligence?” 2Bloomberg QuickTake, “Artiﬁcial Intelligence”, July 2015 Michael A Osborne

open-source state-of-the-art algorithms the case that pilots’ increasing reliance What would you recommend already out that could be immediately on automated flying reduces their own to CEOs of large organizations deployed to applications. So, the ability to actually fly a plane in any direct regarding machine learning and bottleneck is really not investment, it’s way. The case of airplane navigators is artificial intelligence? talent. The challenge is just attracting also very interesting – those who have We really need to develop in-house people with the right skills. not lost their jobs have been deskilled; expertise in these techniques because, more and more of their job is being fundamentally, the increasing availability handed over to an algorithm. As a result, of data across a wide range of different when something unforeseen happens, industries is going to be transformational. There is this wealth when there is a kind of an unexpected If your business is not in the forefront of event, the algorithm becomes useless, of open-source using this data, there certainly will be and humans take over, they won’t another business. The potential value necessarily have the same degree of state-of-the-art that could be realized is going to develop understanding and skills that they might into a real competitive advantage. I think algorithms already have had before the automation. We can companies need to be doing a lot more argue that there have been catastrophes out that could be to recruit and hold on to talent such as because of a human skill being lost owing data scientists and machine learning to the intrusion of automation in that immediately deployed experts. occupation. I think this is a real risk for to applications. many businesses; automation of labor might result in the loss of softer skills What are some of the ethics issues that weren’t necessarily quantifiable but that we will face with machine What are the risks for a company are yet essential. learning or AI? investing in AI or machine learning? The big problem here — and one that the machine learning community is If you rush too fast and hard into Automation of labor spending a lot of time on — is privacy. automation, there are skills that will be irrevocably lost to organizations. In his might result in the loss Our objective is to have systems that book, “The Glass Cage: Automation could be verifiably shown to not intrude and Us”, Nicholas Carr explains how of softer skills that upon someone’s private space. Ideally, automation can separate us from weren’t necessarily you would have an algorithm that would reality and deskill us. One example process, for example, medical health that is particularly evocative is that of quantifiable but are yet records without ever returning to the an airline pilot. Carr plausibly makes users of that algorithm any personal essential. details. These algorithms look at this dataset and return aggregate statistics

Occupations vs. Automation

Carl Benedikt Frey and Michael About 47% of total US A. Osborne studied 702 detailed 47% employment was found to be at occupations to estimate the risk of computerization probability of computerization

21% of US employment Creativity is a hindrance to As a result, 86% of US and 24% of UK employment is automation – creative skills workers in the highly creative highly creative, involved in areas cannot be readily replaced by category are at low or no risk such as development of machines of automation ; the equivalent novel ideas number for the UK is 87%

Source: Nesta, “Creativity vs. Robots: The Creative Economy and the Future of Employment”, April 2015; Carl Benedikt Frey and Michael A. Osborne, “The Future of Employment: How Susceptible are Jobs to Computerisation?”, September 2013 Michael A Osborne

useful to researchers without ever How can governments and be delivering enormous value to society. returning to those people, the details of regulators try to mitigate the We have seen in the last 10 years the any individual patient. This relies upon adverse effects of machine delivery of amazing technologies. the robustness of the algorithm. learning? Even the smartphone in your pocket, for example, has a range of apps that I feel the most imminent concern is the would have been unthinkable even effect of machine learning technologies 10 years ago. My point is that value upon employment. People have rightly is delivered to us as customers or as recommended that these technological Value is delivered to consumers. Let’s make sure that all this events would necessitate the change wealth that is created is distributed in of education. I think we need to equip us as customers or a fair way throughout society at large. the next generation of workers with the Moreover, the wonderful thing about as consumers. Let’s necessary skills that they will need for an the technology is that it has very low economy that is increasingly automated. make sure that all this marginal cost of reproduction. So, at the Governments do have a role in investing point that you develop an amazing new in the technological research. There wealth that is created is speech recognition algorithm, it costs are profound challenges here in getting very little to deploy that algorithm on distributed in a fair way governments to select the right type of everybody’s mobile phone. Everybody technological development. throughout society at benefits from the introduction of those large. techniques as customers and as Is there an argument to slow down consumers. The problem is that some the progress of machine learning? people necessarily will be put out of work by these technologies. I think there I don’t think there is. An important thing is a challenge to make sure that there to note is that all these technologies will are jobs created to replace those that are eliminated by new technologies. Michael A Osborne

Michael A Osborne

Associate professor at the University of Oxford

Michael A Osborne is an associate professor at the University of Oxford. He co-leads the Machine Learning Research Group, a sub-group of the Robotics Research Group in the Department of Engineering Science. Professor Osborne designs intelligent algorithms capable of making sense of complex data through the usage of techniques such as Machine Learning and Computational Statistics. His work in data analytics has been successfully applied in fields as diverse as astrostatistics, labor economics, and sensor networks. Professor Osborne is also the co-author of the widely publicized research “The Future of Employment: How susceptible are jobs to computerisation?” with fellow Oxford academic Carl Benedikt Frey. They concluded in the paper that about 47% of total US employment was at risk of being automatable. Capgemini Consulting spoke with Professor Osborne to understand Machine Learning and its impact on the business world.

About Capgemini

Capgemini Consulting is the global strategy and transformation Now with 180,000 people in over 40 countries, Capgemini is consulting organization of the Capgemini Group, specializing one of the world’s foremost providers of consulting, technology in advising and supporting enterprises in significant and outsourcing services. The Group reported 2014 global transformation, from innovative strategy to execution and with revenues of EUR 10.573 billion. Together with its clients, an unstinting focus on results. With the new digital economy Capgemini creates and delivers business, technology and creating significant disruptions and opportunities, our global digital solutions that fit their needs, enabling them to achieve team of over 3,600 talented individuals work with leading innovation and competitiveness. A deeply multicultural companies and governments to master Digital Transformation, organization, Capgemini has developed its own way of working, drawing on our understanding of the digital economy and the Collaborative Business ExperienceTM, and draws on our leadership in business transformation and organizational Rightshore®, its worldwide delivery model. change.

Find out more at: Learn more about us http://www.capgemini-consulting.com/ at www.capgemini.com.

Rightshore® is a trademark belonging to Capgemini

Contacts: Didier Bonnet, [email protected], Jerome Buvat, [email protected]