does deep learning hold the answers?...of facebook’s deepface algorithm (or set of rules): it can...
TRANSCRIPT
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 1
DOES DEEP LEARNING HOLD THE ANSWERS?
The information and opinions in this report were prepared by EyeforTravel Ltd and its partners. EyeforTravel Ltd has no obligation to tell you when opinions or information in this report change. EyeforTravel Ltd makes every effort to use reliable, comprehensive information, but we make no representation that it is accurate or complete. In no event shall EyeforTravel Ltd and its partners be liable for any damages, losses, expenses, loss of data, loss of opportunity or profit caused by the use of the material or contents of this report. No part of this document may be distributed, resold, copied or adapted without EyeforTravel’s prior written permission. © EyeforTravel Ltd ® 2017
Authors: Senay Boztas, Yekko Creations
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 2
ABOUT
AB
OU
T
We bring together everyone in the travel industry, from small tech start-ups to international hotel brands, to form a community working towards a smarter and more connected travel industry.
Our mission is to be the place our industry goes to share knowledge and data so that travel and tech brands can work collaboratively to create the perfect experience for the modern traveler.
We do this through our network of global events, our digital content, and our knowledge hub - EyeforTravel On Demand.
Our Values
We believe the industry must focus on a business and distribution model that always puts the customer at the center and produces great products. However, to deliver an outstanding travel experience, the strength, skills, and resources of all partners in the value chain must be respected and understood.
At EyeforTravel we believe the industry can achieve this goal by focusing on a business model that combines customer insight with great product and, most importantly, places the traveler experience at its core.
At our core we aim to enable the above by valuing impartiality, independent thought, openness and cooperation. We hope that these qualities allow us to foster dialogue, guide business decisions, build partnerships and conduct thorough research directly with the industry.
These principles have guided us since 1997 and will continue to keep us at the forefront of the industry as a vibrant travel community for many more years to come.
Our Services
Our events are the heart of EyeforTravel. These draw in experts from every part of the travel industry to give thought-provoking presentations and engage in discussions. It is our aim that every attendee takes back something new that can help their business to improve. This might be in the fields of consumer research, data insights, technological trends, or marketing and revenue management techniques.
Alongside this we provide our community with commentary, reports, white papers, webinars and other valuable expert-driven content. All of this can be accessed through one place the On Demand subscription service.
We are always expanding the content we create, so please get in touch if you want to write an article for us, create a white paper or webinar, or feature in our podcast.
EyeforTravel by the Numbers
70,000+ database contacts
2,500+ annual event attendees
100,000+ monthly online reach
1,000+ online conference presentations
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 3
ACKNOWLEDGEMENTS
AC
KN
OW
LE
DG
EM
EN
TS
The development of this report was made possible due to the input and support of industry leaders in the tourism, tours and activities fields. EyeforTravel acknowledges and thanks the following for their time and support to this project:
Marion Mesnage | Head of Innovation and Research | Amadeus IT
Janis Dzenis | Group PR Director | Aviasales
Andrei Grintchenko | Head of Business Intelligence Projects | International Air Transport Association (IATA)
Amer Mohammed | Head of Digital Innovation | Stena Line
Dan Christian | Chief Digital Officer | The Travel Corporation
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 4
INTRODUCTION
INT
RO
DU
CT
ION
If data is the new oil in the economy, then deep learning could be said to be the refinery turning the precious information into insights we can all use. Scientists and researchers are beginning to mimic the learning patterns of the human brain in its early stages of development. What were once uniquely human talents of reasoning can now be undertaken by machines. Now, algorithms can diagnose skin conditions with a smaller error percentage than the average doctor.
However, these are only the first shoots of potential. Deep learning is generally only good at specific tasks currently and struggles to replicate the associative and learning power of the human brain. This remarkable organ is an incredibly complex set of specialized centers working together to increase their collective power, creating a system that can constantly learn, categorize incredibly effectively, and infer behaviors and characteristics of almost anything from a vast repository of diverse memories. Deep learning programs on the other hand largely excel at a single task, can be easy to trick and, until very recently, had a habit of overwriting previous calculations when tackling new assignments (also sometimes known as ‘catastrophic forgetting’).
Deep learning therefore currently truly excels at categorizing and finding relationships between sets of big data. This is of enormous use to the travel industry, where there are huge, diverse databases that can be used to unlock insights into customers. Deep learning will turn the web’s vast unstructured data into meaningful analysis for travel brands and will make large proprietary data sets treasure troves for personalization.
New capabilities in AI deep learning are transformational not just to travel but to our entire human experience. Every day leaps are being made towards replication of key activities of the human brain and this progress is occurring at an astounding rate. Google has created an AI that can make its own AI in the form of programming neural net layers, known as AutoML, and its DeepMind division has built an AI that can tackle the huge challenge of multiple tasks at once with only a small drop in accuracy.
The capabilities and systems are only going to become more sophisticated and capable. This will create social, ethical and philosophical challenges, but ones that may again benefit travel and tourism. AI’s ability to replicate and improve on our more unique capabilities may mean that we reassess the way in which we work, with serious discussions already occurring. One of the more likely consequences is that we look at increasing leisure time, through more flexible working, shorter days, and maybe even a four-day week. Some are even debating a universal basic income. This could benefit our industry as consumers find themselves with increased time to focus on experiences and vacations, although there must be the caveat that incomes levels would need to be maintained, even as AI replaces key tasks and possibly jobs.
This is quite some way out though and, in the immediate term, the benefits for our industry, where many are straining under the weight of data, is potentially vast. Deep learning can create the insights necessary to create better pricing structures, increase click-through and conversion rates, improve search, and build dynamic product offers, to name but a few applications.
So, read on and I hope you enjoy this piece of research. Please also look into the many other pieces of original travel research that form our On Demand service.
Alex Hadwick,
Head of Research, EyeforTravel
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 5
CONTENTS
About EyeforTravel ...............................................................................................................2
Acknowledgements ..............................................................................................................3
Introduction ...........................................................................................................................4
1 Deep Learning: A Quick-Fire Explainer .............................................................................6 1.1 Food for thought .........................................................................................................6 1.2 Deeper and deeper ......................................................................................................6 1.3 Why now? ....................................................................................................................7
2 How Neural Networks Work .............................................................................................8 2.1 A case study: how computers learned to see cats ....................................................8 2.2 Simple machine learning versus neural networks ....................................................9
3 How Will Deep Learning Impact Travel and Tourism? ..................................................10 3.1 Can neural networks find you? ................................................................................10 3.2 Making customer predictions ...................................................................................10 3.3 Real-time response ....................................................................................................11 3.4 The time is now ..........................................................................................................11 3.5 Case Study: TTC and Advertising ..............................................................................11
4 Travel Brands Using Deep Learning to Improve Results Today ....................................12 4.1 Full steam ahead for AI at Stena Line ......................................................................12 4.2 A personal trip at Expedia .........................................................................................12 4.3 Data as fuel for tomorrow’s travel: Amadeus .........................................................13
5 The Limits of Deep Learning ...........................................................................................14 5.1 Idiots savants .............................................................................................................14 5.2 The data problem .......................................................................................................15 5.3 Ethics and regulation .................................................................................................15 5.4 Conclusion: dip a toe into the water ........................................................................15
References ...........................................................................................................................16
CO
NT
EN
TS
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 6
1. DEEP LEARNING: A QUICK-FIRE EXPLAINER
1 | DE
EP
LE
AR
NIN
G: A
QU
ICK
-FIR
E E
XP
LA
INE
R
1.1. Food for ThoughtWe don’t often think about how we think – but for decades computer scientists have been doing just that, to develop artificial intelligence.
Thanks to technological advances, their ideas are now taking serious shape. Harry Shum, executive vice president at Microsoft AI [artificial intelligence] and Research Group, isn’t alone in seeing this as a revolution:
“Thanks to the convergence of three major forces — increased computing power in the cloud, powerful algorithms that run on deep neural networks, and access to massive amounts of data — we’re finally able to realize the dream of AI. AI has the potential to disrupt every single vertical industry… and every single business process, from sales and marketing to recruiting.”( Official Microsoft Blog, 2017)
So, what do we mean by artificial intelligence? Stanford University’s One Hundred Year Study on Artificial Intelligence (AI100) defines it as a set of computer technologies inspired by how we move, perceive and respond to the world – although often operating quite differently (Stone, Peter et al, 2016).
We’re not talking about the kind of computer brain that could pass for human in any context, defined by the British scientist Alan Turing in his 1950 “Turing test”. It’s better to think of American computer scientist Nils John Nilson’s narrower description of computer intelligence as something that will allow a machine to act appropriately, and with foresight, in its own environment (Nilsson, Nils J., 2010).
1.2. Deeper and DeeperThe big step forward has been in a type of machine learning known as deep learning, which sits behind major advances in image and pattern recognition, autonomous driving and speech recognition (also known as natural language processing).
It was originally inspired by how a cat’s brain reacts to light signals and now mimics the human brain’s ability to learn using – as a building block – neural networks.
A neuron is a living cell and part of your nervous system, and the billions of them in your brain pass on pulses of electrochemical activity. Deep neural network systems are organised in hierarchical layers, as they are in a human.
The computer “thought process” is like a ball falling through an early pinball, or a Japanese Pachinko machine, as data passes down a series of processes. Amazon’s director of machine learning Neil Lawrence points out that these simple transformations build into a very complex picture, quite literally in the case of Facebook’s DeepFace algorithm (or set of rules): it can identify faces by using more than 120 million parameters (inverseprobability.com, 2016; Facebook Research, 2014).
The first layer might look for simple edges, the next for shapes, the next for features like eyes, until the network puts the clues together to recognise objects.
Image recognition technology illustrates one challenge: to work, neural networks need masses of classified data and for their errors to be corrected by human hands. Only with this input can these systems learn. However, with sufficient help and information, they can even outperform the human brain at complex tasks, as demonstrated by their victories in increasingly difficult games, first conquering chess, the game show Jeopardy!, and then the board game Go, where there are more potential moves than atoms in the universe according DeepMind founder Demis Hassabis.
1 | DE
EP
LE
AR
NIN
G: A
QU
ICK
-FIR
E E
XP
LA
INE
RDOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 7
1.3. Why Now?Neural networks were proposed in 1944, and got computer scientists excited again in the 1970s and 1980s, but only now are experts really getting to grips with how these systems can function (MIT News, 2017).
The reasons include a democratization of computing power and access to information via the internet, alongside the development of specialised computer tools, like graphics processing units (GPUs) – partly developed for the video game industry and also handy for Bitcoin miners. Our connectedness, booking via telephone, surfing websites, “liking” things on Facebook, is also amassing mind-blowing quantities of data, stored in the cloud.
Amer Mohammed, head of digital innovation at Stena Line, believes AI will transform the travel industry:
“Seven years ago, to do the things we are doing today, you would have to invest USD12million in a super computer. Now we can pay USD6,000 for our algorithms to run for three minutes on Amazon [cloud services]. What’s happening right now is the next major milestone of humanity: the rise of artificial intelligence. Like the new electricity, AI is the foundation of things to come.”
The International Data Corporation predicts that AI will drive worldwide revenues exceeding USD47 billion by 2020, and smart travel firms – which know where customers are, go and behave – could serve themselves a deep slice of that pie (IDC, 2016). Travel firms are aware of this power and those that aren’t investing risk getting left behind. Three quarters of travel professionals working with data expect budget increases in 2017 according to a survey conducted by EyeforTravel as part of its State of Data and Analytics in Travel Report 2017.
75% of Travel Data Professionals Expect to Increase Their Data Spend in 2017
Source: EyeforTravel, 2017
75%
22%
4%
0%
10%
20%
30%
40%
50%
60%
70%
80%
Iexpectourbudgettoincrease Iexpectourbudgettostaythesame Iexpectourbudgettodecrease
“Seven years ago, to do the things we are doing today, you would have to invest USD12million in a super computer. Now we can pay USD6,000 for our algorithms to run for three minutes on Amazon”
2 | H
OW
NE
UR
AL
NE
TW
OR
KS
WO
RK
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 8
2. HOW NEURAL NETWORKS WORK
2.1. A Case Study: How Computers Learned To See CatsTo understand how neural networks work, it’s useful to look at an early example: more than a decade ago, associate professor Fei-Fei Li, director of the Stanford Artificial Intelligence Lab, started training computers to understand pictures.
“Cameras can take pictures by converting light into a two-dimensional array of numbers known as pixels, but they are just lifeless numbers,” she explains in a TEDTalk. “Ultimately, we want to teach the machine to see as we do, naming objects, identifying people, inferring the 3D geometry of things, understanding relations, emotions, actions and intentions.”
She started with pictures of cats, describing their features in mathematical language for a computer algorithm. But even something this simple has infinite variations: curled up, half-visible, or doing those crazy things that make cats so popular online.
So, instead of improving the algorithm, she started feeding it the kind of “training data” a child accumulates via experiences. In 2007, from Princeton University, she and her team launched the ImageNet.
They downloaded almost a billion images from the internet and used crowdsourcing site Amazon Mechanical Turk, a crowdsourcing service that packages out small tasks to freelance workers, to employ 48,940 people in 167 countries to clean, sort and classify them. Using big data to train computer algorithms, which was in its infancy at that time, allowed a database of 15 million images to be labelled into 22,000 categories by 2009. While they opened the database to the worldwide research community, Li started using neural networks to train their object recognition model:
Even if this architecture helped computer vision leap forward, the computer still might think a child with a toothbrush is holding a baseball bat or misread emotion. If it is as sophisticated as a three-year-old in recognising objects, says Li, the challenge is to get AI into its teens (Li Fei-Fei, 2015).
“Just like the brain consists of billions of highly-connected neurons, a basic operating unit in a neural network is a neural-like node. It takes input from
other nodes and sends output to others. These millions of nodes are organised in hierarchical layers, also similar to the brain. A typical neural network used
to train our object recognition model, has 24 million nodes, 140 million parameters and 15 billion connections. Powered by the mass of data from
ImageNet and modern CPUs and GPUs [computer processors], it blossomed in a way no one expected.”
2 | H
OW
NE
UR
AL
NE
TW
OR
KS
WO
RK
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 9
2.2. Simple Machine Learning Versus Neural NetworksData scientist Daniel Gutierrez suggests that just because we can use neural network-based systems, that doesn’t always mean they are always the best data analysis tools. Instead, he writes in a report for insideHPC, simple techniques – like the classic regression graph – can sometimes better predict business need (insideHPC, 2017).
Neural networks particularly suit large datasets where there are unknown relationships between things (for example, recognising patterns, classifying images and trying to understand language). Here, they can help a business make decisions for new data, but only after being trained with masses of input data. “At first glance,” he notes, “some companies might be considered as ‘under-investing’ in computer systems for AI. But…you have to do a lot of work to make AI solutions more productive.”
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 10
3. HOW WILL DEEP LEARNING IMPACT TRAVEL AND TOURISM?
3 | H
OW
WIL
L D
EE
P L
EA
RN
ING
IMPA
CT
TR
AV
EL
AN
D T
OU
RIS
M?
3.1. Can Neural Networks Find You?Deep learning already sits behind some services: when your customers ask the iPhone personal assistant Siri to search for a holiday, neural networks help it to recognize words, imply meaning and find an answer (Business Insider, 2016).
Even if you aren’t using this kind of technology, you need to appear in voice-based searches, says Dan Christian, chief digital officer at The Travel Corporation (TTC).
His group serves two million customers a year through brands including Contiki, Trafalgar and Lion World Travel, and is investigating how to respond. “My 70-year-old father recently tried to book a cottage, asked Siri, and it predicted a result on the first page, when his efforts typing it in didn’t,” he says. “Voice search makes up 20% of all mobile searches (Forbes, 2017). How do we make sure our brands are there?”
3.2. Making Customer PredictionsNeural networks are increasingly used to forecast tourist demand, when relationships between things are non-linear or hard to understand, as well as to analyze how service improvements affect customer satisfaction.
More than 3,000 institutions worldwide are studying how countries can predict tourist numbers using data science, and in a recent academic paper, Ying Yu and team write: “Although there are still a few doubts about neural network based tourism demand forecasting, it is generally believed that nonlinear methods outperform linear methods in modelling economic behaviour and efficiently helping wise decision-making.… Neural network approaches have several unique characteristics, such as (1) being both nonlinear and data driven, (2) having no requirement for an explicit underlying model, and (3) being more flexible and universal.” (Yu, Y et al, 2017)
For instance, Damir Kreši, of the Institute for Tourism in Croatia, and his team demonstrated that a neural network-based system can better analyse how tourists rate different elements of their holiday – with suitability for short breaks, friendliness of locals, and personal safety coming out top (Kreši et al, 2013).
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 11
3 | H
OW
WIL
L D
EE
P L
EA
RN
ING
IMPA
CT
TR
AV
EL
AN
D T
OU
RIS
M?
For Aviasales, a Russian-headquartered metasearch company focused on flights, they are aiming to use their deep learning program to understand the dynamics of air fares and then apply this back to their customers. “We created our deep learning library that forecasts air fares with a 5% accuracy,” says J nis Dzenis, group PR director. “In other words, we could predict the exact price for the round trip ticket from Moscow to Phuket for the Christmas holidays that [a customer would] probably buy in November. And it’s not a fortune-telling! It’s a reliable tech based math model.” They are looking to take this into a recommendation system that can display the best-suited fare and personalized alerts, so conversion rates increase and revenues are maximized.
3.3. Real-Time ResponseUnderstanding more about each customer – in real time – can improve customer service, says Andrei Grintchenko, head of business intelligence projects at the International Air Transport Association (IATA).
“Combining data from different sources to draw conclusions related to air travel can make much more sophisticated predictions on customer behaviour,” he says.
“If you’re an airline, and know your passenger has returned a car 30 minutes to flight closing, this passenger won’t make it on a flight. You can start making a proposal to rebook. But if it’s two hours before, maybe you can offer this customer lounge access. That can improve the journey experience.”
3.4. The Time Is NowMarion Mesnage, head of innovation and research at travel technology firm, Amadeus IT Group, says deep learning has “great potential”. “Travel has gone through a digital revolution, and generates a lot of data,” she explains.
There’s transactional data, typically on e-commerce sites, on what people search, buy and spend, plus operational data on flight or train schedules, traveller numbers and delay information. Add to that tourism information on the internet, and pictures, videos, blogs, comments and ratings by travellers. “The potential for learning from and combining these data to develop new offers or improve offers is mind-blowing,” says Mesnage.
3.5. Case Study: TTC and AdvertisingAt The Travel Corporation, Christian is using deep learning to guide corporate advertising with the help of IBM Watson and digital marketing firm Rocket Fuel.
“When there are world events, we don’t want our ads appearing next to tragedies, particularly terrorism in Paris or London,” he says. “IBM Watson allows us to track sentiment [online], and automatically adjust our advertising to ensure we’re not appearing by an adverse situation. Thankfully, travellers are resilient and change their destination: it’s up to us to become more responsive.”
Originally, after a disaster, all TCC adverts would “go dark”. But now if someone is researching their destination on CNN or the BBC, for instance, adverts would not be pulled after regional events, but suggest different destinations. Several weeks later, potential customers might again see the original banner ads.
4 | T
RA
VE
L B
RA
ND
S U
SIN
G D
EE
P L
EA
RN
ING
TO
IMP
RO
VE
RE
SU
LTS
TO
DA
YDOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 12
4. TRAVEL BRANDS USING DEEP LEARNING TO IMPROVE RESULTS TODAY
4.1. Full Steam Ahead for AI at Stena LineIn his third day, a new developer at Stena Line realised they were spending a lot of time manually correcting misspelled names on ferry bookings, mostly in the Netherlands.
So, head of digital innovation Amer Mohammed, explains: “He downloaded a database of 6,000 common names, trained a neural network to mutate those names – basically, misspell them – and trained another neural network to guess the original name. We now have this running every night, checking and correcting names with 99.7% accuracy. It’s working excellently and looks like it’s going to be used in all markets.”
Deep learning is also increasing profits at the ferry operator, which makes money and attracts customers by selling cut-price products including spirits, perfumes and cosmetics.
“We present a price on board and a comparison price, based on what we know the cheapest land price is,” says Mohammed. “If you have 30,000 products, it takes six weeks to do that manually and standard web crawling matching the product names only gives a 10% match.”
Now, though, it has the AI Crawler, a smart system using machine learning, neural networks and image recognition software that can guess if products match based on name formatting, descriptions, product numbers, bottle size and alcohol level. It has 91.3% accuracy, cost EUR15,000 to build and will run once a month.
“The reason you buy on board is that it’s cheaper, but we don’t want to be too cheap!” Mohammed adds. “Sometimes they raise prices on land and we lose more than we need to lose.”
4.2. A Personal Trip at ExpediaOnline travel agent Expedia has artificial intelligence at its core: just to propose the prices in its Best Fare Search service is an “unbounded computer science problem,” according to vice president of global product David Fleishman (Computerworld UK, 2016).
Its 700 data scientists are constantly tweaking the basic algorithm and have incorporated neural networks to process customer question-based searches.
Fleishman says Expedia’s smart hotel finder, its bot on Facebook Messenger, and boxes allowing a client to “search anything” are products of this: “If someone says in casual conversation, ‘I want to go to Paris,’ we make an assumption that they mean Paris, France instead of Paris, Illinois,” he explains. “At the heart of this understanding is Natural Language Processing (NLP) or the task of turning travellers’ common questions…into searchable content. We’ve been experimenting with NLP ever since we realized that the standard travel search framework doesn’t work as well on mobile devices.” (Expedia.com, 2016)
Jan Krasnodebski, director of lodging revenue optimization at Expedia, told the EyeforTravel Smart Travel Data Summit 2016 that by learning from data, the business can personalise offers. “If you are not willing to put the customer first, all the terabytes of data and machine learning algorithms aren’t going to help you,” he adds.
4 | T
RA
VE
L B
RA
ND
S U
SIN
G D
EE
P L
EA
RN
ING
TO
IMP
RO
VE
RE
SU
LTS
TO
DA
YDOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 13
4.3 Data as Fuel for Tomorrow’s Travel: Amadeus Amadeus IT, which runs one of the world’s major Global Distribution Systems, helping agents book airline tickets and travel services, believes small tweaks using AI will have extraordinary results (Amadeus and ATKearney, 2017).
Marion Mesnage, head of innovation and research, says it is already exploring how to use deep learning to maximise prices and revenues from airline tickets.
“Instead of building very complex models to understand the parameters that influence revenue, you just feed the data into a system and let the system – thanks to a deep learning algorithm – learn what works,” she says. “There is no assumption on the model whatsoever. That’s very disruptive…but we believe it could equal or outperform what humans can achieve.”
The “reinforcement learning for revenue management” project started in 2017 and so far has successfully been used to make predictions for one airline on one segment of a route. Amadeus is working to expand it to far more complex situations.
It is also using deep learning to analyse customer bookings data, and personalise offers for different classes of passenger – again, without applying any a priori assumptions. “We need to learn as a company to treat data as a real asset,” adds Mesnage. “It’s the fuel that will be used for applications, many of which we don’t even know [about] yet.”
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 14
5. THE LIMITS OF DEEP LEARNING
5 | T
HE
LIM
ITS
OF
DE
EP
LE
AR
NIN
G5.1. Idiots SavantsSingularity – the idea of human civilization utterly changed by artificial superintelligence – isn’t happening just now. As Google DeepMind research scientist Raia Hadsell says, today’s systems might be incredibly good at recognising cats but they are a long way from human multitasking intelligence (The Verge, 2016).
“There is no neural network in the world, and no method right now that can be trained to identify objects and images, play Space Invaders, and listen to music,” she admits. Still, she and others are working on a system of progressive neural networks aimed at transferring skills and avoiding “catastrophic forgetting”, with some success (Hadsell R et al, 2016).
“What we are doing now is artificial narrow intelligence, AI that’s specific to a certain task,” says Stena Line’s Amer Mohammed. “We need to come up with mathematical models that can actually understand the world, not just fake understand it.”
And with less effort too: Amadeus IT’s Marion Mesnage points out that the AlphaGo AI computer program beat the top Go player, but using 50,000 times more energy than his 20-watt human brain.
These processing power requirements for deep learning are the main limiter for Aviasales: “A system resource is the only limit. Even our test library consumes a lot. Hence, we could be more productive by achieving [a] new level of computer performance,” says Dzenis.
x50,000 more electrical power is needed for the AlphaGo computer than the human brain
Deep learning is currently a ‘narrow intelligence’ that struggles to come close to the complex multitasking, social capabilities and creative thought of the human brain
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 15
5 | T
HE
LIM
ITS
OF
DE
EP
LE
AR
NIN
G
5.2. The Data ProblemTo train neural networks, you don’t just need data: You need a mass of cleaned, standardized data. Neil Lawrence, head of Amazon Research Cambridge, points out that computer systems need dramatically more input than humans to understand concepts and features and there’s currently a “data efficiency deficit.”
There could also be resistance to gathering information. Consumers might search anonymously or delete cookies, while a trust in Britain’s National Health Service was reprimanded after allowing Google research access to patient data as it failed to comply with data protection rules and gather consent. This was even though records were anonymised (New Scientist, 2016).
Human bias can be a problem too: when Microsoft released chatbot Tay on Twitter in 2016, it mirrored other commentators, apparently becoming racist and generally objectionable within hours. The deep learning network, clearly, is only as intelligent as its diet.
5.3. Ethics and RegulationMicrosoft CEO Satya Nadella believes “ethics and design go hand in hand” and artificial intelligence must be infused with “protections for privacy, transparency and security.” (Slate, 2016)
In their 2017 report, What If? Imagining the Future of the Travel Industry, Amadeus and ATKearney suggest the level of regulation and consumer data protection could lead to very different scenarios. In 2018, Europe’s new data protection regulations become law – including giving consumers the “right to an explanation” about algorithmic decision making (Amadeus and ATKearney, 2017).
This, though, isn’t straightforward: neural networks are so complex that it’s difficult to know why they have gone right, let alone wrong. After a Florida driver died when his Tesla car collided with a tractor trailer, in autopilot mode, a federal investigation eventually concluded there were no safety defects in the system and that human drivers are responsible all of the time (National Transportation Safety Board, 2017). As Nadella says, “we want not just intelligent machines but intelligible machines.”
Stanford AI100 points out many limitations to deep learning: “the difficulty of creating safe and reliable hardware…smoothly interacting with human experts … gaining public trust … overcoming fears of marginalizing humans … and diminishing interpersonal interactions.”
5.4. Conclusion: Dip a Toe into the WaterThese limitations aren’t the end of the story: in fact, they could ensure that computer scientists adopt a more humanistic approach.
Developers point out that – if you have enough data – you don’t need to spend millions to start a small project exploring whether deep learning could overturn your assumptions and better serve your customers.
“We have passed a threshold of feasibility in open source [algorithms], hardware and data,” says Mesnage. “It’s exciting.”
DOES DEEP LEARNING HOLD THE ANSWERS?
www.eyefortravel.com 16
REFERENCES
RE
FE
RE
NC
ES
Amadeus and ATKearney, 2017. What If? Imagining the Future of the Travel Industry report [Online]. Available at: http://www.amadeus.com/documents/reports/what-if-imagining-the-future-of-the-travel-industry.pdf. Accessed 26 July 2017.
Business Insider, 2016. Apple completely changed how Siri works and almost nobody noticed [Online]. Available at: http://www.businessinsider.com/apples-siri-using-neural-networks-2016-8?international=true&r=US&IR=T. Accessed 26 July 2017.
Computerworld UK, 2016. How Expedia.com was built on machine learning [Online]. Available at: http://www.computerworlduk.com/it-business/how-expediacom-was-built-on-machine-learning-3644963/. Accessed 26 July 2017.
Expedia.com, 2016. Exploring Artificial Intelligence via Natural Language Processing [Online]. Available at: http://blog.expedia.com/tag/machine-learning/. Accessed 26 July 2017.
EyeforTravel, 2017. The State of Data and Analytics in Travel Report 2017 [Online]. Available at: http://www.eyefortravel.com/revenue-and-data-management/state-data-and-analytics-travel-report-2017. Accessed 26 July 2017.
Facebook Research, 2014. DeepFace: Closing the Gap to Human-Level Performance in Face Verification [Online]. Available at: https://research.fb.com/publications/deepface-closing-the-gap-to-human-level-performance-in-face-verification/. Accessed 26 July 2017.
Forbes, 2017. OK Google, Let’s Talk About Voice Search [Online]. Available at: https://www.forbes.com/sites/forbesagencycouncil/2017/07/07/ok-google-lets-talk-about-voice-search/#1a4333e67c50. Accessed 26 July 2017.
Hadsell R et al, 2016. Progressive Neural Networks paper, Cornell University Library [Online]. Available at: https://arxiv.org/abs/1606.04671. Accessed 26 July 2017.
IDC, 2016. Worldwide Semiannual Cognitive/Artificial Intelligence Systems Spending Guide [Online]. Available at: https://www.idc.com/getdoc.jsp?containerId=prUS41878616. Accessed 26 July 2017.
insideHPC, 2017. Drilling Down into Machine Learning and Deep Learning. Available at: https://insidehpc.com/2017/05/drilling-down-into-machine-learning-and-deep-learning/. Accessed 26 July 2017.
inverseprobability.com, 2016. Deep Learning, Pachinko, and James Watt: Efficiency is the Driver of Uncertainty [Online]. Available at: http://inverseprobability.com/2016/03/04/deep-learning-and-uncertainty. Accessed 26 July 2017.
Kreši et al, 2013. Artificial Neural Network-Based Applications in Travel and Tourism Research: A Review and Case Study, presented at the International Critical Tourism Studies Conversnce V, 2013 [Online]. Available at: http://cts.som.surrey.ac.uk/publication/view/artificial-neural-network-based-applications-in-travel-and-tourism-research-a-review-and-case-study/. Accessed 26 July 2017.
Li Fei-Fei, 2015. TEDTalk, March 23, 2015, How we teach computers to understand pictures [Online]. Available at: https://www.youtube.com/watch?v=40riCqvRoMs. Accessed 26 July 2017.
MIT News, 2017. Explained: Neural networks [Online]. Available at: https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414. Accessed 26 July 2017.
National Transportation Safety Board, 2017. Crash Summary [Online]. Available at: https://dms.ntsb.gov/pubdms/search/hitlist.cfm?docketID=59989. Accessed 26 July 2017.
New Scientist, 2016. Revealed: Google AI has access to huge haul of NHS patient data [Online]. Available at: https://www.newscientist.com/article/2086454-revealed-google-ai-has-access-to-huge-haul-of-nhs-patient-data/. Accessed 26 July 2017.
Nilsson, Nils J., 2010. The Quest for Artificial Intelligence, Stanford University [Online]. Available at: https://ai.stanford.edu/~nilsson/QAI/qai.pdf. Accessed 26 July 2017.
Official Microsoft Blog, 2017. Microsoft Build 2017: Microsoft AI – Amplify human ingenuity [Online]. Available at: https://blogs.microsoft.com/blog/2017/05/10/microsoft-build-2017-microsoft-ai-amplify-human-ingenuity. Accessed 26 July 2017.
Slate, 2016. The Partnership of the Future [Online]. Available at: http://www.slate.com/articles/technology/future_tense/2016/06/microsoft_ceo_satya_nadella_humans_and_a_i_can_work_together_to_solve_society.html. Accessed 26 July 2017.
Stone, Peter et al, 2016. Stanford University, One Hundred Year Study on Artificial Intelligence (AI100) [Online]. Available at: https://ai100.stanford.edu. Accessed 26 July 2017.
The Verge, 2016. These are three of the biggest problems facing today’s AI [Online]. Available at: https://www.theverge.com/2016/10/10/13224930/ai-deep-learning-limitations-drawbacks. Accessed 26 July 2017.
Yu, Y et al, 2017. Statistical Modeling and Prediction for Tourism Economy Using Dendritic Neural Network, Computational Intelligence and Neuroscience Journal [Online]. Available at: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5299217/. Accessed 26 July 2017.