WHITE PAPER
Artificial Intelligence (AI) is transforming every industry it touches from healthcare, to retail and advertising, finance, transport, education, agriculture and so many more. The purpose of AI? To take care of all the mundane tasks employees currently handle, freeing up their time to be more creative and perform the work that machines cannot do. Today, the rapidly advancing technology is used mostly by large enterprises through machine learning and predictive analytics. AI is not a technology of the future, it’s happening now, and companies who fail to adopt it will get left behind.
This paper will explore the application of AI in business, delve in to who should be employing these technologies, and hone in of the
transformative impact AI is having on every industry with research contributions from leading minds in the field. Expert opinions from
academics, industry leaders, researchers, CEOs, founders and many more are included to comment on the impact of artificial intelligence across multiple industries.
1、INTRODUCTION
Artificial Intelligence is disrupting and transforming every industry it touches. From business operations and efficiency to
innovative means of customer service, medical research breakthroughs, smarter transport systems and targeted advertising
campaigns, it’s an inescapable reality of today’s world. Businesses unwilling to adopt AI will fall behind, and it’s predicted that the
revenue generated from both the direct and indirect application of AI software will grow from $1.38 billion in 2016 to $59.75 billion
by 2025. (Tractica, 2017)
“As machines become smarter, consumers will expect flawless customer service around the clock, and by 2025 AI will drive 95% of all customer interactions, with consumers unable to differentiate bots from human workers via online chats as well as over the phone.” (Servion, 2017)
Thanks to the availability of huge amounts of data and increasingly intelligent algorithms, machines can learn, speak,
make informed decisions and carry out complex tasks in an increasingly effective manner. Not only is this driving research
breakthroughs, but implementation in industry is demonstrating the huge potential impact that real-world applications
of AI can have on businesses across all industries from retail and advertising, to healthcare, sales and marketing, transport,
travel and tourism amongst others.
In the 80s when little progress was being made, three pioneers Yoshua Bengio, Yann LeCun and Geoffrey Hinton toiled away
working on neural networks where other scientists had abandoned them due to lack of computational power. ‘In the lean times when
no one believed in neural nets, these [were] the people who really kept the torch burning and really inspired a lot of people.’ (Recode,
2015)
Since then AI has been applied in business, enhanced through research, and has had some astounding breakthroughs: DeepMind
have mastered the Atari console and conquered the game of ‘Go’, and AI is now becoming superior to humans in several areas
including object recognition and face detection, as well as working towards passing the Turing test. (DeepMind, 2013)
It is machine learning, and in particular neural networks, that seem right now to hold huge promise; but the history of artificial intelligence warns us not to assume we can accurately predict what will work, or when.(Ed Newton-Rex, Jukedeck)
WHAT ACTUALLY IS AI?
Depending who you talk to, you usually find two definitions: one where AI aims to embed human intelligence into a machine, and
another where AI aims at discovering possibly super-human levels of intelligence. If interpretability of the AI system is important,
we might prefer an intelligence that's closer to the human's. But if we wish to design the best AI system that detects diseases in
patients, we would be happy if it were better than a human doctor.(Hugo Larochelle, Google Brain)
AI is the simulation of intelligence in computers: behaviour exhibited by nonbiological systems that we would consider intelligent if exhibited by humans. A more recent approach, is ‘machine learning’ where the computer learns how to complete tasks by being exposed to large datasets. (Ed Newton-Rex, Jukedeck)
What’s allowing us to progress so quickly?
AI requires huge data sets and the coupling of ‘really great science with amazing advances in technology has allowed us to collect data
at that we've never had access to before’,enabling models to learn more quickly.(Jasper Snoek, Google Brain) The pace of current advancements wasn’t foreseen, for example Ankur Handa, OpenAI didn’t expect to see ‘super-human performance on ImageNet within only three years of the first paper on CNNs (convolutional neural networks) from Geoffrey Hinton's group in Toronto.’
Whilst these progressions are rapid and impactful, consideration is necessary to identify whether your business in should be applying AI. Factors such as cost, available data, industry relevance and staffing are a consideration for businesses of all sizes as well as the likely ROI. Further chapters will provide solutions to these key points in order to identify the impact of AI in industry and whether you should be employing these
technologies in your business.
Progress has been driven primarily by new ideas and insight rather than bigger datasets and faster computers.(Jörg Bornschien, DeepMind)
Who is driving progressions in AI?
It’s not just technology giants leading the AI race, but Universities, venture capitalists (VCs) and internal researchers. Research from institutions and industry experts opens doors for businesses to apply these models to their work, and AI specific VCs are assisting in breakthroughs through their funding.
2. LANDSCAPE
The current landscape of AI consists of four areas: models, training data, hardware and people. New models (such as Generative Adversarial Networks) are being used with great success; bigger datasets are available to train on; hardware improvements are speeding up training; and every success in the field leads to more people entering the industry.(Ed Newton-Rex, Jukedeck)
Progress is consistent, yet to have a direct impact on society takes time and money from organisations implementing research progressions. There are limitations in the availability of data, the power of the computation, and the training period and intelligence of each model.Advancements in unsupervised learning are revolutionising business applications saving both time and money - between 2010 and 2014, global investments in AI tech grew from $1.7bn to $14.9bn (Merrill Lynch, Bank of America, 2016).
Although AI isn’t a new concept, the journey to mainstream implementation has taken years,with recent decades seeing the most rapid
advancements. This means that new advances become obsolete quickly thanks to our current culture of using arXiv and social media for
dissemination. This ‘hyperactivity’ in AI can be disruptive and is forcing companies to reconsider the products they are designing (Hugo
Larochelle, Google Brain).
Factors in the development of AI that have brought us todays models:Researchers figured out how to use graphics processing units (GPUs) to speed up training of neural nets. These methods allowed us to scale up to larger datasets & to iterate faster when doing research on smaller, academic-minded datasets.
The DL community developed much better open source libraries and tools for deep learning. For example, Theano was a breakthrough in how we can implement DL models, which inspired the now widely popular TensorFlow.
ArXiv and social media have become the prime strategy for publicising research. This has made it possible to iterate and build on other researchers' work at a much higher pace. (Hugo Larochelle, Google Brain)
HOW CAN ADVANCEMENTS BE APPLIED IN BUSINESS?
ML is always held back by limitations in the amount of computation we can use.These advances are able to have very apparent impacts on real world problems, and Ian Goodfellow explained: ‘the May 2017 announcement of the new generation of Google TPUs is huge. The new Google TPU helps bridge the gap between the amount of computation we can leverage in DL experiments and the amount of computation used in a biological nervous system. The previous generation was available only to Google engineers, but the new one will be available to Cloud customers, and researchers can apply to get access for free.’
Whilst we have seen a lot of progress in unsupervised learning, especially in generative modelling, we have yet to see it’s full potential. It’s impossible to attribute progress to a specific technique, but it’s fair to say that progress has been driven by new ideas and insight rather than bigger datasets and faster computers.’(Jörg Bornschien, DeepMind)
As more companies are able to implement AI models in their businesses through the availability of cloud platforms (for instance from
Google, Amazon, Microsoft), AI technology will continue to become more accessible to both industry and society at large. There is no need for
any particular industry to get left behind, and industries across the board are being positively disrupted by AI advancements (Hugo Larochelle,
Google Brain). Business intelligence tools are able to source, analyse, transform and report on data to provide valuable customer insights and
allow businesses to invest their time and money in the right areas. Companies striving to build emotional connections with their users will see
increased customer satisfaction through emotional AI, after all ‘in real life situations, people are actually pretty bad at emotional intelligence’, causing us to end up in pointless fights, dismiss good arguments because they go against our biases, and judge people based on stereotypes. Once AI can be trained without bias it should be able to provide more rational responses than its human counterparts (Mikko Alasaarela, Inbot, Oct 2017)
Across multiple industries we see mind-boggling results due to deep learning in data preparation, speech recognition, text understanding, computer games, cybersecurity, etc. Deep learning has provided the greatest leap in performance in the history of AI (and arguably, in the history of computer science), and has rendered many traditional methods obsolete. As a result, within the next decade any company that will not heavily rely on deep learning will be left behind (Eli David, CTO, Deep Instinct).
DEEP DIVE:WHAT ARE GPUs?
Graphics processing units (GPUs) accelerate deep learning applications by allowing computers to offload intense portions of applications to the GPU whilst the remainder of the code runs on the CPU. Pioneered in 2007 by NVIDIA, GPU accelerators now power energy-efficient data centres in government labs, universities, enterprises, and small-and-medium businesses around the world. (NVIDIA, 2017)
WHAT ARE TPUs?
Tensor processing units (TPUs) accelerate the neural network computations behind the scenes delivering 15–30X higher performance and 30–80X higher performance-per-watt than contemporary CPUs and GPUs. These advantages help many of Google’s services run state-of-theart
neural networks at scale and at an affordable cost. (Google, 2017)
3. PROGRESSIONS
“There's been such rapid progress in recent years. Since 2012 on the image net, all of that progress is essentially because of the advances we’ve made to the learning algorithms - it’s not only because of the computational resources, but also the research progress"(Aaron Courville, Assistant Professor, University of Montreal, Oct 2017)
There’s a lot of discussion about AI ‘stealing jobs’,but this isn’t the first time. During the industrial revolution factory workers were put out of jobs by machines; fast forward to the internet and industries were disrupted. However, this wasn’t the end for industry: factory workers learned to operate machines, journalists used the internet as a resource rather than a hindrance. For every job that’s destroyed, there will be countless new
opportunities created. (Forbes, 2017)
Hugo Larochelle explained, whilst we can’t know for sure the future of the AI landscape and the economy, historically there are examples of tech
developments that aimed at "replacing people" yet growth in employment for that industry was observed (the advent of ATMs is an often cited one).
For some industries which rely heavily on labor, AI might be better than human beings. Since AI will never be tired and hardly make mistakes, this will create job openings in the development and automation of the robots and machines. Industries which require more human intelligence and emotion however, are unlikely to be completely transformed by AI, but will see the introduction of AI assistants to both save time and helps us do more intelligent work. (Miao Lu, Research Scientist, Yahoo Labs).
AI IS HELPING IMPROVE EFFICIENCY, CUT COSTS & OPTIMISE BUSINESS
AI allows people to spend time on the more thoughtful and creative aspects of their jobs rather than wasting time on menial tasks. For instance, smart customer relationship managers (CRMs) can be used to automate data input, HR employees can use smart recruiting software to sort applicants more efficiently, and marketing and sales teams can send personalised and targeted campaigns. Businesses of all sizes can be reinvigorated by these software processes allowing employees and management to invest their time in growing the business. (Salesforce, 2016)
AI computing has become very fast thanks to GPUs (and now TPUs), and we know speed means productivity. Productivity ultimately leads to performance. Beyond speed, we must make sure it works well and is widely accessible.(Kimberly Powell, NVIDIA)
AI is now far more accessible, even to those outside the world of computer science and AI in general, and as Hugo Larochelle highlighted, the increase in AI cloud platforms is improving the accessibility. It’s now important for Universities and Colleges to embrace the progress of AI and move away from treating it as a narrow topic in computer science meant for graduate study.
Historically, AI was developed by large companies since a lot of training data was required to reap the rewards of AI and ML.For the company to acquire this data meant that the business needed to either have a preexisting and mature product with a lot of user traction, or for the company to have high financial backing to buy the data at an exchange. Today, the cost of acquiring training data is significantly reducing with platforms like
scale.api and Amazon Mechanical Turk matching human annotators to startups at cost efficient prices. Moreover, a lot of Platform as a Service systems that directly serve machine learning models to input datasets like Amazon ML and Azure ML can help bootstrap startups in the AI/ML space. (Vijay Ramakrishnan,Cisco)
AI allows researchers to focus on the problems at hand instead of spending time creating new tools to approach new problems.(Ian Goodfellow, Google Brain)
As consumers demand more personalisation and customisation in their lives, personal assistants that cater for a particular brand and
products catered specific to their likings will increasingly be expected from the consumer.(Kimberly Powell, NVIDIA). Businesses have to
reach such expectations and manually doing this will stop being scalable after a certain point.
Consider the service your company provides and the challenges you’re currently experiencing,whether they lie in ROI, organisation, efficiency,
accuracy, customer service or any other aspect of the business. Consider the impact an AI could have on not only your day to day activity, but on
your company’s success as a whole.
RESEARCH PROGRESSION TESTIMONIAL (Eli David, Deep Instinct)
In 2004 I led a project that created a critical real-world system with immense complexity, entirely based on neural networks and evolutionary computation. Back then, merely mentioning neural networks resulted in raised eyebrows and bewildered looks. And there were good reasons for that: we could train rather shallow neural networks, the training was excruciatingly time consuming, and for most applications the results were inferior to other traditional machine learning methods which were much faster and more convenient to deploy.
During the past few years we have witnessed two revolutions: The first, much improved neural networks methods that allow us to train "deep" neural networks (and hence the name "deep learning"), and the second, the use of GPUs which allow us to train neural networks which are by orders of magnitude larger. And these two breakthroughs are responsible for the current ongoing deep learning revolution.
4.USE CASE & BENEFITS
Five years ago I was talking to a friend who was leading one of the leading computer vision groups in the world. When discussing deep learning he referred to it as "another transient hype". Today, his entire research group is focused only on deep learning, obtaining amazing breakthroughs, and barely using any of the traditional image processing methods which have been used in previous decades. I expect every leading business in
every industry to heavily rely on deep learning in the upcoming years (or otherwise be completely left behind).(Eli David, Deep Instinct)
Whilst the obvious contender for implementing AI is ‘tech companies’, cross-industry enterprises are seeing benefits incomparable to previously
used models when implemented successfully.Smaller companies can be deterred from investing in an long term AI strategy due to the
vast amounts of labelled data that’s required.Many businesses don’t have access to this, but that isn’t to say that it isn’t possible or advisable
to build a successful strategy.
Companies currently benefitting from AI are the ones which already have an existing store of archived data, like hospitals or businesses where
the cost of acquiring incremental data is low, like internet companies. However, the time investment required to collect training data or develop ML models is decreasing. Additionally, cloud infrastructure services like Amazon AWS and Google Cloud ML reduces the need to acquire expensive capital goods upfront, thereby reducing the friction of small/medium size companies entering into this space. (Vijay
Ramakrishnan, Cisco)
Companies currently benefitting from AI are the ones which already have an existing store of archived data, like hospitals or businesses where
the cost of acquiring incremental data is low, like internet companies. However, the time investment required to collect training data or
develop ML models is decreasing. Additionally, cloud infrastructure services like Amazon AWS and Google Cloud ML reduces the need to
acquire expensive capital goods upfront, thereby reducing the friction of small/medium size companies entering into this space. (Vijay
Ramakrishnan, Cisco)
4.1 BUSINESS OPTIMISATION
Tasks that have previously relied on manual labour from employees such as customer support, data management, supply chain management, and marketing strategy decisions and automation are areas predicted to be the most affected by AI in the next five years (Forbes, 2017). There are very few companies who don’t rely on employees in any of these areas, and paying large teams is costly, so maximising their efforts by alleviating time consuming tasks that could be managed by an AI model is an obvious direction for businesses to take. Introducing a technology that can provide more insights into making an informed decision like a human can is sure to improve not only ROI, but optimise efficiency and allow staff to excel in areas that require human skill (Ankur Handa, OpenAI) .
DeepMind is working and researching with the NHS to build a ML system to recognise impending blindness. (The Guardian, 2016)
Researchers have created ‘dermatologist level’ classification of skin cancer using neural nets. (Springer Nature,2017)
IBM Watson carried out cancer research where 99% of the treatment recommendations were coherent with the physician decisions.(SVN, 2017)
4.2 HEALTHCARE
The adoption of AI in healthcare is on the rise,helping not only with diagnoses but also in the production of new drugs and research
advancements. Deep learning methods are already enabling analysis that outperforms experts in applications such as the detection
of diabetic eye disease and cancerous tumours. These progressions will not replace doctors and medical experts, but will allow them
to achieve higher accuracies with more time to focus on the hardest cases and with a strong second opinion to help minimise human error
(Jasper Snoek, Google Brain). For example, in the U.S, there is one radiologist for every 10,000 people, and in India there is one for every
100,000 people. With AI, radiologists will be more effective and will therefore have more time to focus on those cases that are more difficult.
(Kimberly Powell, NVIDIA). Jasper Snoek expects this to spread across medical domains to, for example, cardiology where algorithms
could automatically analyse EKGs taken at home.
4.3 TRANSPORT
Soon, autonomous vehicles won’t just be delivering pizza. Their development will see reduced congestion, improved mobility for those
who have difficulty getting around today, greener cities, and increased access to public transit.Raquel Urtasun, Head of Uber Advanced
Technology Group (ATG) is using ML techniques to make their self-drive cars more intelligent and safer than human drivers. Raquel explains that
when she thinks about cities today she ‘can’t help but see all of the parking spaces and garages dedicated to our personal vehicles.’ A
future where we can begin to transform those parking structures into green space and new businesses is not too far away. Additionally, the
huge popularity of Amazon has created a massive opportunity for autonomous vehicles, as the already prevalent shortage of last-mile drivers will continue to increase based on currentconsumer trends (Kimberly Powell, NVIDIA).
We use ML to identify and distinguish between high and low value customers to optimise market spend and minimise exposure to unprofitable customers. A large percentage of customers churn and have a 0% CTLV, whilst some will spend millions each year on ASOS.It’s important to identify which users are which to drive sales and provide the best UX.(Ben Chamberlain, ASOS)
4.4 RETAIL
Retail sales are dropping as the cost of living increases. This is problematic for retailers who’ve had to turn to tech to encourage
consumers to jump back on the spending wagon. Using machine learning, retailers can capitalise on market data to understand and
anticipate consumer behaviours and trends (RTInsights, 2017). ASOS are using machine learning to predict customer lifetime value. In
an ideal world, ASOS would ‘know every action a customer will make for the rest of time, but [they] can’t do that’. Distinguishing between
high and low value customers allows companies to optimise market spend and minimise advertising to unprofitable customers.Retail has massive amounts of consumer data so is able to use AI to show real-time behaviours in the customer base. This further benefits the industry by allowing companies to implement better inventory management systems (Ben Chamberlain, ASOS).
The examples above demonstrate the positive transformations of applying AI in a variety of sectors, but are there any negative ramifications?Data scientists or an AI officer can be expensive hires, but there are an abundance of online training courses available such as Andrew Ng’s Coursea as well as platforms like Kaggle. Hugo Larochelle explained that they have seen examples of people without formal training in AI who are able to produce high quality solutions to problems using AI. This suggests that the barrier of entry to using AI is lowering, and new AI-enabled systems in new domains should continue to be observed. There are some industries that will be required to tread more carefully than others in their implementation of AI, for example law enforcements agencies that have an enormous stake on public life that employ AI need to tread carefully since the outputs of the such systems could over-fit on past data, so racial and gender bias could be learned.
In this case interpretable models are preferred to black box models. Additionally, deep learning systems can be easily tricked using carefully
crafted input data. MindMeld’s Vijay Ramarkishnan explains that in these circumstances, systems will need ample supervision and common-sense before being the first responders to control important decisions like physical gate-keeping of sensitive areas. It’s important to remember however, that fictional representations of AI tend to paint a dystopian picture. We need to be very careful with AI development, as it could certainly bring about negative scenarios; but we shouldn’t assume that things will go badly, in any industry. AI has the potential to do enormous good, probably more than any technology we’ve ever created before (Ed Newton-Rex, Jukedeck).
TAKING AI FURTHER: APPLYING DEEP LEARNING ALGORITHMS
Whilst AI is now readily available for implementation cross-industry, there are still challenges in the commercial applications of deep learning. Despite the availability of multiple research oriented opensource deep learning frameworks, they are not efficient enough for application in many domains (e.g.,cybersecurity); their incorporation in the final product is complicated; and even where applicable,highly skilled and scarce deep learning experts are required to employ them. As a result, today mostof the real-world applications of deep learning in industry are in computer vision, because they mainlyrely on convolutional neural networks only, for which the barrier of entry is lower.Eli David from Deep Instinct shared his expertise on overcoming this obstacle: I expect that in the next few years we will see deep learning being employed effectively in additional industries in realworld solutions. We will have more deep learning frameworks which focus on efficiency in real-world rather than being geared towards research, as is the case with currently available deep learning frameworks, and they will lower the barrier of entry for adoption in the industry.
5.CASE STUDY:DANIEL GOLDEN, DIRECTOR OF MACHINE LEARNING,ARTERYS
Artificial Intelligence has the capacity to transform all aspects of healthcare from the research and production of new drugs and treatments, to the way in which patients are cared for. Cloud based company, Arterys, has created an intelligent medical imaging platform that supports advanced AI tools and ultra-fast image processing. Back in 2011 when Arterys was founded, most medical imaging happened on hospital premises with limited computing power, with measurements taken manually and automated tools minimal. Out of this, the vision to advance medical imaging via
cloud computation and advanced analytics was born. Arterys was the first clinical cloud-based deep learning in healthcare model to receive FDA approval to be used in a clinical setting, a huge step for AI in healthcare. Their aim of creating ‘A world where clinical care is data-driven, intelligent and patient focused’ comes to life through their ‘deep learning engine with unlimited computation to translate deep learning models into clinical products.’
MEDICAL IMAGING PROBLEMS:
Medical images contain a wealth of information that isn’t being leveraged to optimise patient care.Through images, it’s difficult to precisely track disease and even more so to predict the course a disease will take or how it will respond to specific treatments. Several challenges in medical imaging are currently preventing data-driven diagnoses:
Increasing Productivity Demands
Complex scans with more images at a higher resolution are becoming the norm.Scanning volume has increased significantly in the past 10 years,overburdening radiologists and causing burnout.
Challenges in Quantifying Diagnostic Imaging
Quantification of anatomy is tedious, and clinicians don’t have time to measure precisely. While volumetric measurement is ideal, disease is often measured with few linear dimensions which can be inconsistent between readers.
High Error Rates
The retrospective error rate among radiologic examinations is around 30% with real time errors in daily radiology practice averaging 3-5%. These errors lead to delays in treatment, unnecessary biopsies and mis-informed treatment decisions.
These problems deliver real-world ramifications to healthcare professionals and patients alike. The lack of efficiency means that delays and errors in screening, diagnosis and monitoring are common and lead to poor outcomes at higher costs. Delays can come from either the time it takes for an image to be interpreted, or the time it takes for physicians to detect disease progression or assess the effectiveness of a treatment. Arterys identified this issue and are leveraging their platform to accelerate these insights to provide more timely treatments as well as stopping and altering treatments that do not work.
How can AI Help?
Whilst the system augments the radiologist in several aspects, it does not aim to replace the human, but enhance it in the following ways:
Expediting the tedious work surrounding setup, interpretation, and reporting on image findings.
Improving accuracy and consistency when measuring anatomy so the radiologist can better track changes in patients.
Helping the community standardise how disease is measured and categorised, so that there is more consistency across practices.
This will allow for real-world study of outcomes of similar patients and optimisation of treatment for individual patients.Arterys have faced a good amount of skepticism surrounding AI in the clinical community, however having allowed clinicians to try the system for themselves and compare the results has proven a convincing demonstration of the success and positive impact of AI. The clinicians are then able to experience the benefits of the automation and assess its accuracy. The AIenabled platform simultaneously enables the clinician to edit the measurements and
assessments it automates, allowing them to remain in full control of the interpretation.
What’s Stopping Mainstream Implementation, and What’s Next?
AI requires very large datasets, and due to the physical recording of data that were popular until recently, these high-quality datasets can be hard to come by. Data found in a single institution can be biased to how clinicians practice there, so real leaps will occur when clinicians from different institutions collaborate in creating datasets that represent the population in a given disease. As more and more clinicians and researchers embrace AI and see the emergence of new patterns and biomarkers thanks to their research, more people are going to be keen to employ the technologies.
What does Arterys implementation of AI teach us about cross-industry applications?
• Costs can be cut whilst improving efficiency of data collection
• Image analysis does not need to be time consuming, and can become more accurate
• Jobs do not need to be lost, the employee simply becomes more effective working alongside the intelligent system
6.CASE STUDY:MAGGIE MHANNA, DATA SCIENTIST, RENAULT DIGITAL
Renault Digital is 100% Renault-owned subsidiary that started operations on January 1st, 2017 and aims at digitalising Renault’s core business for its employees, partners and clients worldwide. Renault Digital's mission is to deliver digital projects for all of the company's business lines, whether for the upstream (in factories, for the design of vehicles) or for the downstream (points of sale, e-commerce, after-sales).
“We have based our approach on an inspiring start-up ecosystem. One of the most important core competence at Renault Digital is Data Science
and AI. Driverless Cars are the main goal. In this project (AD/ADAS), the focus is on offering one the widest range of advanced driver assistance
systems (ADAS) on the market, enhancing vehicle safety and in some cases operating without human intervention. These ADAS form the starting point for autonomous vehicles, capable of instigating decisions on behalf of the driver.”
AI FOR A POSITIVE IMPACT
•ML and data science for predictive maintenance:
Predicting what will fail and when promises cost savings over routine or time-based preventive maintenance because tasks are performed only when warranted.
•Accurately incorporating analysis of customer feedback:
Helps in building vehicle and subsystems performance for guiding future product design.This also helps in detecting failure patterns for
establishing a relationship between the failure and causes of failure. Take an example of a cause of failure in several operations in the car
which happens to be associated with regionspecific issues such as inferior fuel quality, climatic conditions, road infrastructure, etc.
BRINGING CONNECTED CARS TO THE ROADS
The product lifecycle of a vehicle from inception through engineering design and manufacture to service and disposal is 5 years. New features, such as system upgrades and applications are provided regularly for smartphones; this can't be said for connected vehicles. AI models must be optimised to speed up this process and allow connected cars to run sophisticated applications.
As well as consuming power and battery, connected cars also consume bandwidth. Here, we need to talk about mobile operators: passing data from the vehicle to the network and vice versa will represent a non-negligible amount of data. Operators need to work with car manufacturers to make the transition seamless.
Obstacles in the Roll-Out of Connected Cars:
Data & Privacy
• The privacy of the individuals whose data is being collected and analyzed is at risk daily.Data is produced via navigation, multimedia
and other applications.
• It provides information on the driver's habits,which can be transmitted to the car manufacturer, as well as to third parties. The EU’s General Data Protection Regulation (GDPR) will come into force next year with tougher punishments covering a larger scope.
Obstacles in the Roll-Out of Autonomous Cars:
Reliability
• The vehicle must be able to perceive its environment perfectly, to make the right decision and carry out this action safely.
• In the real world, roads are a highly dynamic and varied environment, therefore there are millions of use cases the vehicle must learn.
Artificial Intuition
• Behaviours of other users can sometimes escape AI, humans are able to behave intuitively where computers are not.
• Humans take subtle signs for granted: the body language of a traffic controller, or a cyclist making eye contact. Machines will be able to learn these interactions with AI, but this may take infinite hours of driving and testing.
Legal Framework
• The Vienna Convention on road traffic, (the basis of European road codes), must evolve - it stipulates that the driver must keep control of
the vehicle at all times.
• Most manufacturers are working with authorities on these changes, which can be assumed to be completed before 2020.
HOW RENAULT DIGITAL ARE USING AI IN OTHER PROJECTS
Demand forecasting
Optimising manufacturing
Distribution and the supply chain processes and networks
Finding correlations between incidents occurring at different phases of a project by identifying existing questions having the same context and linking them using NLP.
Optimising On-board diagnostics (OBD) by grouping diagnostic trouble codes and proposing them as labeled patterns to technicians with the corresponding solutions, saving them both time and effort.
How is Renault Digital’s implementation of AI relevant?
The structure of their AI teams isn’t that of a giant corporation, and can be replicated by businesses of varying sizes: The projects development methodology used isn't waterfall but rather an agile approach relying on small teams. The teams selforganize and run on short cycles, with a strong link with the experts of the field and systematic feedback to users to validate the concept and quickly derive value.
7. WHERE ARE WE HEADING?
Contrary to what people think, scientific progress is slow and continuous, but social and economic impact can be disruptive - there’s no doubt that we will reach human level AI but we don’t know how long it’s going to take. We need to make sure that AI and society will bloom for the benefit of all.(Yoshua Bengio, Université De Montreal, RE•WORK Deep Learning Summit 2017)
Companies will dedicate up to 20% of their workforce to neural nets, and 85% of customer interactions will be managed by AI.
57% of consumers expect voice activated smart assistants to have a major or moderate impact on their daily lives.(Salesforce, 2017)
Revenue generated from the direct and indirect application of AI software will grow from $1.38 billion in 2016 to $59.75 billion by 2025. This represents a significant growth curve for the forecast period with a compound annual growth rate (CAGR) of 52%. (Tractica, 2017)
38% of U.S. jobs and 30% of U.K. jobs could be automated with the industries leading being transportation & storage (56$), manufacturing (46%) wholesale and retail (44%) and healthcare (17%).(Tractica, 2017)
IN THE NEXT 5 YEARS: WHAT THEY SAY
Eli David, Deep Instinct:
Whilst we may face an initial wave of job loss, many new jobs will be created, which instead of competing with AI, will work together to achieve
complex tasks not feasible today.
Raquel Urtasun, Uber ATG:
Thanks to AI, I see a future with reduced congestion, improved mobility for those who have difficulty getting around today, greener cities, and increased access to public transit.
Ankur Handa, OpenAI:
I see healthcare, public sector and government being affected positively as these are the places where most often important decisions are made.
Jasper Snoek, Google Brain:
New techniques for merging disparate data sources and NLP will allow us to analyze medical records, discover symptoms and predict medical
outcomes .
Jörg Bornschien, DeepMind:
We will see progress related to rapid acquisition of knowledge: Few shot learning for example, where a generative or discriminative model has
to generalize from only a few positive examples.
Kimberly Powell, NVIDIA:
More enterprises will implement AI in their own industry by crafting ideas to service each unique customer with AI. With an increased accessibility,everyone in every industry will be allowed to innovate with AI for themselves.
BEYOND THE NEXT FIVE YEARS…
“Research in real intelligence (humans and other animals) strongly suggests that there is a correlation between the number of neurons in the brain and the overall intelligence. This seems to hold true for artificial neural networks as well, as one could argue that our current state-of-the-art deep learning models are very similar to the neural networks from the 1990s, with the main difference that they contain about a million times more synapses (connections). Extrapolating from this trend, I predict that within the next several decades as the size of our DL models grows exponentially, and we would find better and better neural network architectures and parameters, we will get close to or surpass human-level. My guess is that this will happen in the lifetime of most of us. “(Eli David, Deep Instinct)
8. GLOSSARY
ArXiv- An open-source repository of electronic pre-prints of scientific papers
AD - Advanced driver
ADAS - AdvancedAdvanced driver-assistance systems
CPU - The central processing unit is the electronic circuitry within a computer
Deep Learning - Area of machine learning research which has been introduced with the objective of moving machine learning closer to one of its original goals of AI.
GAN - Generative adversarial networks are a class of AI algorithms used in unsupervised introduced by Ian Goodfellow in 2014 learning
GPU - Graphics processing units accelerate deep learning applications by allowing computers to offload portions of the application to the GPU whilst the remainder runs on the CPU
Machine Learning - An application if AI that provides systems the ability to automatically learn and improve from experience without being explicitly programmed.
ROI - Return on investment
TPU - Tensor processing units are an application specific integrated circuit developed by
Google especially for machine learning
Theano - Numerical computation library for Python
Tensorflow - Open-source software library for data flow programming across a range of tasks.
9. REFERENCES & ADDITIONAL READING
• ‘The RIC Report’, Bank of America, Merrill Lynch, (12 Jan 2016View our calendar of events here. )
• ‘Is Your Business Ready for Artificial Intelligence?', BGCG Henderson Institute, Philipp Gerbert, Martin Reeves, Sebastian Steinhäuser and Patrick Ruwolt, (6 Sept 2017)
• ‘Reshaping Business With Artificial Intelligence’, MITSloan, SamRansbotham, David Kiron, Phillipp Gerbert and Martin Reeves, (6 Sept 2017)
• ‘Playing Atari with Deep Reinforcement Learning’, DeepMind Technologies, Volodymyr Mnih, Koray Kavukcuoglu, David Silver & others, (19 Dec 2013)
• ‘Voice and Facial Recognition Could Help AI Surpass Humans in Emotional Intelligence’, Venture Beat, Mikko Alasaaela, Inbot, (Oct 15 2017)
• ‘How AI is Already Changing Your Life Backed By 6 Incredible Stats’, Salesforce, Kim Honjo, (1 Mar2017)
• ‘IDC FutureScape: Worldwide IT Industry 2017 Predictions’ IDC FutureScape, Frank Gens & others,(Nov 2016)
• ‘What is GPU-Accelerated Computing?’, NVIDIA, (2017)
• ‘An in-depth look at Google’s first Tensor Processing Unit (TPU)’, Google Cloud Platform, Kaz Sato,(12 May 2017)
• Video: ‘Visual Reasoning via Feature-wise Linear Modulation’, RE•WORK Deep Learning Summit Montreal, Aaron Courville, University of Montreal, (11 Oct 2017)
• ‘The Paradox of Robots Taking All Our Jobs’, Forbes, Adam Oximek, (28 Oct 2017)
• ‘Artificial Intelligence - Where and How to Invest’, Capgemini, (2017)
• ‘Artificial Intelligence not to kill jobs yet: Survey’, Economic Times, (15 Nov 2017)
• ‘Website Personalization Powered by Artificial Intelligence | Stats and Facts’, Boomtrain, Nitin Deshdeep, (14 Feb 2017)
• ‘How Artificial Intelligence is Revolutionizing Business in 2017’, Forbes, Louis Columbus, (10 Sept 2017)
• ‘Artificial Intelligence in Healthcare: Past, Present and Future’, SVN, Fei Jiang, Hui Zhi, Yi Dong & others, (21 June 2017)
• ‘Google DeepMind pairs with NHS to use machine learning to fight blindness’. The Guardian, Alex Hen, (5 July, 2016)
• ‘Artificial Intelligence Market Forecasts’, Tractica, (Q2, 2017)