Photo: autonomous AI/IDx diagnostics
How can AI augment a doctor’s intelligence to improve medical diagnoses, patient care and decision making? Artificial intelligence is revolutionising the healthcare sector in many ways. In this podcast, Esade Associate Professor Esteve Almirall talks with leading AI expert Xavier Amatriain, the mastermind behind the Netflix and Quora recommendation algorithms and the co-founder and chief technology officer of Curai, a company dedicated to using AI in primary care.
In this episode, both AI experts delve into the promising AI revolution in the healthcare sector, how Curai is helping to solve the bottlenecks of primary care, how recommender systems like Netflix’s work to predict user wishes – and what the future of AI may hold.
- 2:33 | What is the augmented intelligence paradigm and why does it matter in areas such as medicine?
- 5:41 | Will sensors be enough to help us solve the last mile of healthcare, that is, primary care? What is the approach of Curai?
- 9:11 | Where does the data for medical AI come from?
- 13:00 | Healthcare is a major spend for all governments, yet we still experience long delays and waiting lists. Will AI ever solve these problems?
- 17:05 | Natural language is opening new frontiers. What are the opportunities and challenges for this in healthcare?
- 20:50 | How far are we from a widespread use of natural language and having it included in home assistants such as Alexa or Google Assistant?
- 23:30 | Let’s dream. If tomorrow the Spanish National Health System asks for your help, what would be your contribution?
- 26:32 | How important is machine learning and recommendation algorithms for a company like Netflix or Quora? Can the business value be quantified?
- 29:42 | Right now, recommenders like Netflix’s tell us what we know we like, any chance that one day they could find what we don’t know we like?
- 34:00 | At Esade we work a lot on machine learning with tabular data. The state of the art is still gradient boosting machines (GBM). It is now more sophisticated in pre-processing, as well as in automatically choosing the best hyperparameters, and handling categorical data. But there is no advance that is really a game-changer. In your opinion, will this change in the near future?
- 42:23 | What businesses do you think will be most disrupted by AI?
- 47:05 | Which is more important, algorithms or data?
- 50:32 | How different is the current research on machine learning compared to the actual production of machine learning?
- 54:35 | At Esade, we have included AI, particularly machine learning, in most of our programmes. What is your advice in this area for business schools?
Esteve Almirall: Welcome everyone. This is Esade's podcast and today we have with us Xavier Amatriain, one of the leading experts in artificial intelligence. Today is a special podcast devoted to AI in one of the hottest topics these days: AI and healthcare. Welcome to this podcast, Xavier.
Xavier Amatriain: Thanks for having me, Esteve.
Esteve Almirall: Xavier is the co-founder and chief technology officer of Curai, one of these special companies that work on healthcare and AI. Xavier graduated as a telecommunications engineer and did his PhD in artificial intelligence in Spain. He was a lecturer in both Spain and the US, and later moved to the private sector, first working in Telefónica, and then joining Netflix, where he became engineering director. He was also the VP of engineering at Quora. I’m sure many of you use Netflix and Quora.
Xavier is a leading expert in AI, particularly in machine learning, in areas such as recommender systems. He contributed to developing the Netflix algorithms that suggest all the content that you are glued to all day long during this crisis. Xavier, thank you very much for taking our call, we are glad to have you here. Our first question: AI and healthcare have a long history together, particularly in areas such as test analysis, X-rays and so on. You sometimes describe AI more as augmented intelligence rather than artificial intelligence itself. What is the augmented intelligence paradigm and why does it matter in areas such as medicine?
Xavier Amatriain: First of all, thank you for the nice introduction. That is a great question. You were saying that AI and healthcare have a long tradition. I would even go further: some of the first applications of AI were in healthcare. Some of the famous expert systems from 50-60 years ago were applied to things like psychology and even healthcare. They were applied to the early diagnosis systems in the University of Pittsburgh, where this started around 50 years ago.
AI is just a way to augment the intelligence we already have
There is a long history of connection between healthcare and AI because when you look at healthcare it seems clear that there are many things that could be automated and there are many algorithms that are driving many decisions.
Now to answer your question. Yes, I often talk about augmented intelligence. When we talk or read about AI we sometimes hear a definition of AI that seems like science fiction; robots are going to come and kill us. I try to demystify that and say: “Hey, AI is just a way to augment the intelligence we already have.”
And this is particularly important in healthcare, because we are not trying to replace doctors. What we're trying to do is augment their intelligence by giving them tools that help them make better and faster decisions. If you think about a decision that a doctor on average needs to make, what do they need to do? They need to capture information from a patient in about 10 to 15 minutes, remember everything they studied in medical school and be informed about all the research that has come up in the last few years. They need to combine all of that and very quickly make a decision.
AI can augment doctors' intelligence by giving them tools that help them make better and faster decisions
That is really hard to do. It’s like asking a mathematician not to use a computer or even a calculator. Clearly, there are tools that are able to deal with that information, process it, present it better, and help doctors, and people in general, make better decisions. That's why I call it augmented decisions: it's not a replacement for our intelligence, but an augmentation of our intelligence.
Esteve Almirall: Many times most of the costs come from the last mile and healthcare is no exception. Most of the bottlenecks in healthcare are taking place in this last mile, which is primary care. Now we have many sensors, such as bands and the Apple Watch, and many of us believe that these sensors can help us. Will this be enough? What is the approach of Curai?
Xavier Amatriain: I totally agree that one of the biggest problems with healthcare is scaling the last mile. It's about scalability and availability of healthcare for everyone. And it's becoming harder and harder to do that because, first of all, the population is growing. Secondly, we also live longer, and therefore there are more diseases and we want more care, but we don't have more doctors. And as I said before, being a doctor is extremely hard, as dealing with all this information.
There are many sensors, and this is going to become especially important. But the reality, and this is something we need to acknowledge, is that the kind of sensor that you mention – for instance, the Apple Watch and so on – only reaches a minority of the population. And as a matter of fact, it's the minority of the population that can usually pay for expensive healthcare, because these are the rich people.
Our approach at Curai is: “How can we use AI and the latest technologies, for things that can make access easier for the average person, and easier for those who have difficulties accessing healthcare? Unfortunately, it’s not going to be the Apple Watch – we're not going to be able to send one to every citizen.
One way to scale healthcare is by using the communication channels that most people have
But language – and everything that deals with language – is something that everyone has. Natural language processing and image recognition (almost everyone can send a picture on a phone) scale much better for the time being. Our hope is that over time, we are going to be able to scale different kinds of sensors and have the ability to read everyone's temperature and heartbeat from their homes. And I think we're not very far from that.
But for the time being, I think the way to scale healthcare is by using the communication channels that most people have, which are language, text, voice and images. This is something that is cheap, available, and there's a lot that you can do with that kind of technology. If you think about it, most of what doctors do in real life is talk to patients, ask questions, get answers from patients, take measurements and that's it. Of course, this is something that is also useful because a lot of medical practice, particularly in primary care, is about asking questions and getting answers from patients.
Esteve Almirall: Natural language is fantastic, it brings a universal approach. You focus on medical diagnosis. We had expert systems in medical diagnosis back in the early eighties and now we are in the decade of data – in many areas everything is about data and deep learning. In your case, where does the data for medical AI come from?
Xavier Amatriain: That was one of my concerns when I got into medical AI because I was coming from Netflix and Quora where we had tons of data. My first question was: where are we going to get the data from? It's not easy to get that data. Now, it turns out that in medical AI you can combine quite different kinds of sources of data. None of these sources are perfect, but if you combine them in the right way, you can leverage the benefits of each kind of data.
For example, one particular kind of data that's interesting is electronic health records – the digital transcript of everything that happens during a medical visit.
In medical AI you can combine quite different kinds of sources of data and leverage the benefits of each kind of information
When you go to a medical visit and see the doctor typing and entering things into a computer, that goes into an electronic medical record with notes from the doctors, codes about diagnosis, symptoms and results from the lab. That is remarkably interesting data. Traditionally, this data has only been available in medical systems. That's one of the reasons why at Curai we partner with, for example, Stanford Hospital and the Massachusetts General Hospital in Boston. We work with their medical data because it is a way to bootstrap the systems.
Medical health records, however, contain many errors and they haven't been designed to learn machine learning algorithms – they are not a silver bullet. So, as I said before, you need to combine them with different kinds of data.
We use expert systems. Believe it or not, the expert systems from years ago, which have been evolving for many years, are really interesting because they condense centuries of medical knowledge into rules and probabilistic networks. You can use those systems to generate data.
Expert systems from years ago condense centuries of medical knowledge into rules and probabilistic networks
As a matter of fact, one of our unique approaches (we've published a couple of papers about this in Pure AI) is that we combine synthetic data generated from expert systems with natural data coming from electronic health records and other places. The combination of synthetic data from expert systems, data from electronic health records and data extracted from medical publications offers unique and interesting information.
Esteve Almirall: Synthetic data is on the rise. People are using it more and more to complement and augment existing data, particularly in areas where synthetic data performs very well. This is a fantastic approach. On the other hand, healthcare is a major spend for all governments: the US spends around 14.4% of its GDP, Germany and France around 9%, the UK around 7.5% and Spain 6%. Even with this level of expense we have long delays and long waiting lists. Will AI ever solve these problems?
Xavier Amatriain: That’s exactly what we’re working on at Curai. Our hope is that we can implement an online, always on, always available, very easy-to-access medical service that does precisely what you’re saying: scaling the last mile, the primary care access through automation and AI.
AI can provide suggestions to the doctors as they are making their diagnosis and treatment plans
And as I said at the beginning, it's doing this not by entirely replacing doctors, but by augmenting them. The approach that we have at Curai is a combination of three levels: AI, health coaches (humans with medical knowledge) and medical doctors. We combine these three levels of knowledge to make the system scalable. There are many ways you can do that. One of them is leveraging the ability of AI to capture and extract the right kind of information from the patient in the initial phases of the interview, and then inject what the AI is capturing from that conversation to provide suggestions to the doctors as they are making their diagnosis and treatment plans.
Another interesting thing that we do is related to telemedicine. Telehealth or telemedicine has traditionally used video, which has a number of issues. One in particular is that it is hard to parallelise and automate. If you are connected with a doctor over video for 30 minutes, those are 30 minutes that the doctor must spend only talking to you in the video. There's no way that you can scale that, it is what it is, 30 minutes of doctor time.
Text, on the other hand, is interesting because if you have a text conversation with a doctor, you can inject AI, and you can parallelise it. You can have a much more agile and scalable approach to having medical conversations without the need for synchronous face-to-face communication. You can leverage a lot of the benefits of a synchronous messaging and text communication – and that goes a long way. This is the approach we're using at Curai.
Esteve Almirall: One thing that you mentioned before is natural language. When we work on user interfaces we are still using the same metaphors from the seventies of Alan Kay in Palo Alto – we still have folders, desktops, use a mouse, drag and drop things... Natural language is opening new frontiers and in contrast with everything else, there is no standard user interface. It’s a blank slate full of opportunities. What are the opportunities and challenges for this in healthcare and with applications like yours?
Natural language is opening new frontiers in artificial intelligence
Xavier Amatriain: I agree with you. I think we designed some interface paradigms to interact with computers, because that was the best thing we had. The mouse now seems an obvious way to interact with a computer, but it's not that obvious – it’s something that somebody invented and it worked. Clicking, dropping and dragging is not a natural way of interacting with a complex interface.
We humans have much richer modes of communication and language is the obvious one. Language is the way that we should aspire to communicate with many of these rich interfaces. It is challenging and I'm sure many of you have experienced frustration when interacting with Amazon's Alexa, Google Home, Google Assistant, Siri, or any of these devices.
But although we sometimes experience frustration, when they work there is also the delight of thinking “whoa,” this is really what it should be. I think the idea that natural language is the right mode of interaction with computers is, at this point, pretty obvious, and it's more a matter of whether the technology will be able to catch up.
There have been many advantages recently. Just five years ago, natural language processing was barely using deep learning. There were a lot of structural approaches for dialogue systems and frame-based dialogue systems and so on. There have been many improvements in the past five to ten years that are slowly making a difference. We've seen some early impressive demos from some of the big companies working in this space such as Google and so on. I'm convinced that in the next five years, we will see a huge improvement in the natural language interaction between humans and computers, and it's going to become the default paradigm for artificial general intelligence.
In the next five years, we will see a huge improvement in the natural language interaction between humans and computers
Esteve Almirall: I think that everyone agrees that natural language will be a normal way of interaction between very complex and very simple things. We will use natural language in many situations, from switching channels to a doctor's diagnosis. How far are we from a widespread use of natural language and having it included in home assistants such as Alexa or Google Assistant?
Xavier Amatriain: I don't think we are very far away. Like everything in technology, progress is nonlinear. Breakthroughs move very quickly when they are applied, commercialised or integrated into products. I think we're already at the point where a lot of the research approaches are really good in natural language processing. There are question answering systems that are already performing better than humans.
It's hard to agree that they are as smart as a human, but for some metrics, some question-answering systems, and even automated translations and so on, are really good at this point. Now, regarding a purely natural language system that is able to do things that are convincingly feel-good for the user, we are currently at the demo stage. But my prediction is that in the next five years, it is going to become the de facto interface for many systems, and it’s going to be changing the way that we interact with computers.
Breakthroughs move very quickly when they are applied, commercialised or integrated into products
I have to say that there are different modes for natural language processing, such as voice or language. I think it's going to be applied to both. So, I'm not saying that it's only going to be voice. It could be text. But I think, in any case, it will be natural language interacting via text or via voices with our computers.
Esteve Almirall: Absolutely – nobody uses fonts anymore, we all text. Let’s dream for a second. If tomorrow the Spanish National Health System asks for your help, and God knows we need help, what would be your contribution?
Xavier Amatriain: I think one of the things that I have to be realistic about is the scale at which we are operating at Curai. We're a small start-up with over 30 people and we are starting on this. We have collaborations with some hospitals, but they are early stage collaborations. If a big institution like the Spanish Health Ministry asked for our help in delivering healthcare to millions of patients in Spain, the honest answer would be that we're not yet ready. We're at the early stage of growing and scaling this.
For example, Curai is not even available in all of the US, it's only available in California. We are scaling little by little and making sure that we're able to scale. It's going to take a bit of time and a lot of work to get to a point that we feel we can scale to many more millions of people. That being said, the overall volume of people in California is roughly comparable to Spain. So, of course, it could work but it would be hard for us to take on another country.
Esteve Almirall: I'm convinced that primary care will be full of AI in a few years. I would like to take advantage of having you here to talk about recommender systems. You are one of the top experts in recommender systems worldwide. Believe it or not, recommender systems are in many of our classrooms – many of our professors teach about recommender systems, and particularly the ones you built at Netflix. How important are these recommender systems for a company like Netflix or Quora? Can the business value be quantified somehow?
At Netflix everything you see is a recommendation
Xavier Amatriain: For sure. When I was at Netflix, I would start my talks saying: “At Netflix everything is a recommendation.” So that's how valuable recommendations were and are at Netflix. Everything you see on the service, including when you're searching, is a recommendation. All of the interface, how it reacts, what it responds to and what you're being offered is personalised and geared towards recommending things that you might be interested in.
It has a huge business value, which was actually quantified at one point in many millions of dollars. It's not hard to quantify, at least to some extent. As you may know, in industry whenever we run an experiment with any kind of improvement to a recommender system, we do what is called an A/B test. We present version A and version B of the recommender system – and then we measure the difference for the users and how they respond to each choice.
A relatively easy experiment is to say: “Okay, I'm going to remove all the recommendations and I'm going to see what's the effect of recommendations on the actual business.” That's something that every now and then we used to do at Netflix just for a very small subset of people and for a small period of time, because it meant losing a lot of money. But this way you can measure the impact and it is really huge. It's incredible how, as much as people might think that some of the recommendations could be better, if they didn't have those recommendations, they probably wouldn't be even using a service like Netflix, or YouTube, or Quora.
Esteve Almirall: That's great. We all use recommender systems, we love them. But we also get frustrated with recommender systems because they always recommend us what we know we like. For example, I love science fiction and so the system keeps recommending science fiction films. The problem is that the recommendations keep getting worse because I have seen all the good movies. Is there any chance that one day these recommender systems could find what we don’t know we like?
In recommender systems, the less you know about a user, the more you need to encourage that user to explore
Xavier Amatriain: Hopefully, the answer is yes. The concept you're talking about is very well known and studied in recommender systems, it’s serendipity. Serendipity is when you find something that you didn't know you were looking for. There's a lot of work on how to include serendipity into the algorithms. Another way to think about it is the well-known tradeoff between exploration and exploitation. Exploitation means presenting more of the things that I know you like, while exploration is presenting something that I'm not sure you're going to like but it's going to give me more information about you and will enable you to explore a different area of the catalogue.
Any well-designed recommender system has to include a way to tune the exploit tradeoff. It's a very long topic, I could go on for hours about this. But the bottom line is the less you know about a user, the more you need to encourage that user to explore, and so the more you need to tune the algorithm for exploration.
However, if I have shown you all different kinds of things in the catalogue, and it turns out, you've always said no to everything except for science fiction, then I need to tune down the exploration and say: “Well, maybe all that he likes is science fiction.” I do think there's still a lot of room for improvement and for growth.
One of the tricky things with recommender systems is that it's unclear what's the right objective function that they need to optimise for. There's a lot of discussion about that, particularly, for example, in the context of YouTube and how it might be promoting some extreme point of views, or it might just be optimising for engagement.
When we at Netflix asked people what they liked and what they wanted to be recommended everyone had a very high-brow concept of themselves
I'll give you one example. When we at Netflix asked people what they liked and what they wanted to be recommended, everyone had a very high-brow concept of themselves. So everyone would say: “Oh, I want documentaries and I want the best movies that won all the Oscars.” But when we measured what they were watching, it turned out when they came home tired after a full workday, many would tune into a dumb comedy show just because they wanted to laugh. They didn't want to acknowledge that they liked that, but that is what they were actually watching.
Then there was the tricky decision of: “Should we optimise for what users say they want to watch, or for what users actually watch in the end?” Questions like this are very interesting and very tricky. When I have conversations with people they tell me: “Hey, you're recommending things that are different from the ones I want,” This concept of what people want is different from what they do – it's an interesting psychological effect that we could talk about for some time.
Esteve Almirall: At Esade we work a lot on machine learning with tabular data. The state of the art is still gradient boosting machines (GBM). It is now more sophisticated in pre-processing, as well as in automatically choosing the best hyperparameters, and handling categorical data. But there is no advance that is really a game-changer. In your opinion, will this change in the near future?
Xavier Amatriain: We could go on for a long conversation because it's not an easy answer. I agree with you. I think GBM are a really powerful approach – that is almost my default approach to anything. In fact, I even usually start simpler and go for logistic regression first and then GBM, and then complicate things only if necessary.
There's an ongoing debate about whether deep neural nets are better or not for categorical variables
There's an ongoing debate right now about whether deep neural nets are better or not for categorical variables. I'm still honestly not convinced one way or the other. If you look at some of the Kaggle competitions, it is true that most of them are won by a combination of both. There is the combination of gradient boosted decision trees, and then categorical embeddings put into some kind of neural network. I think that is the state-of-the-art winning approach.
Now, would I recommend that in practice to a company? Well, it really depends. I can go back to the recommender system community where we have exactly the same discussion because recommender systems have a lot of similar data – that's based on matrices and tabular data too. In recommender systems, many companies have switched to neural nets over the past three to five years. Google started doing that and about five years ago, they published a few papers on it. They claimed improved results.
Now the reality is, and this is fascinating, that when you talk to the people in those teams, they didn't switch to embeddings and neural nets for the accuracy gains, but more because of the engineering improvements that they got out of reducing the amount of feature engineering that they needed to do, and this enabled faster innovation and faster iteration.
Now, that for me is not intuitive and it's a bit mind-blowing. I didn't believe it at first. When I first heard it, I couldn't believe that a deep neural net with many layers and different approaches to introducing embeddings and projections could be simpler to maintain and faster to innovate than a gradient boosted decision tree.
Many companies have switched to neural nets over the past three to five years
But the reality, if you've seen some of those approaches, is that every time you have a GBM or logistic regression, you are going to be spending a lot of time on feature engineering. And that is almost art. You need to really understand the domain, you need to encode the variable in a particular way, you need to transform the features and normalise them. It's a lot of work with a lot of data science behind it.
While many of these approaches to injecting the same information into a neural net are more automatic, if you start including things like auto-ML and architecture search in neural nets, you need a lot of machines like Google has, but these things almost happen magically.
I wouldn't say it's a clear winner, but there are situations where many people rightfully claim that deep neural nets are superior to GBM for categorical data. For example, if you follow the online Fast-AI course from Jeremy Hayward, he's a strong proponent and believer that deep neural nets for tabular data are better than GBM. There are a few examples which are pretty convincing, but there are also examples I can find where it's the other way around. So, it's an interesting debate, and I don't think there's a clear answer.
I think teaching GBMs for now as a default is a good option, but I wouldn't discard going into deep neural nets as a way to extend the toolkit.
The other thing with deep neural nets – and Netflix has published on this subject – is that for example, if you have tabular data but you want to introduce things like time sequences, it's very hard to do it with a GBM. It is much easier to do it with a neural net because you can combine. You can even combine convolution neural nets with recurrent neural nets and embeddings, and then you start having some architectures that are really powerful. Netflix did publish that they had started to consider sequence and timing, then neural nets beat the GBMs because they didn't have a good way to inject time and sequence information into a GBM.
Esteve Almirall: That's a wonderful answer. This Fast-AI course is very popular among our students. We also work a lot with platforms. In platforms, everything is ready for deep learning and it takes you less time to implement things, not because they consume less CPU but because you have the tools ready.
Because of these tools and many developments, we have seen how deep learning is eating machine learning in many ways. But there are other things happening as well. We have been talking about the use of synthetic data and its use for semi-supervised learning, causality in recommender systems and many more. Do you see anything that could disrupt this machine learning panorama or we are set with deep learning and that's it?
One of the real big debates in AI is how much of the knowledge should be innate and how much needs to be learned
Xavier Amatriain: Deep learning has already happened. The question now is what’s next. There are a few areas that are interesting, some of which we have already discussed. I’m really interested in the notion of synthetic data, but I’m also interested from the perspective of how to inject priors, knowledge and invariance into machine learning systems.
One of the real big debates in AI is how much of the knowledge should be innate and therefore injected into the structure and architecture of the model that is being learned, and how much needs to be learned. There are extremes, people like Yann LeCun say everything needs to be learned, while others say: “No, there needs to be a lot of structure and priors in the system, otherwise, it is impossible to build smart AI”.
I really believe that it’s a combination of both: injecting this combination of structure, prior knowledge and expertise into systems, and then combining with learning. This is something that we haven’t solved – although there’s a lot of research around it. I think this is an important area where we will see huge improvements and advances. Synthetic data is one approach, you can generate synthetic data with structured expert systems or similar, which are injecting structure into the data set itself – and that’s a powerful approach.
There are other hard research areas such as self-supervision or transfer learning, which we all use now. I think this is a key area where we could do more. This notion of being able to learn a particular problem, and then transfer the knowledge that you’ve learned to a relatively similar problem – that’s really powerful and it has been important for many key advances in the past few years.
Auto ML is another area that is exciting and powerful – this notion of being able to have machine learning models or meta models that learn the optimal parameters and architectures by themselves. They are basically doing the work of data scientists.
And of course, there’s the whole reinforcement learning approach and the big advances we’ve seen in games being played, from chess to go to poker, and how reinforcement learning is the key paradigm in all those areas. We will soon see big things happening in these areas.
The best algorithm cannot be trained on useless data and the best data doesn’t work unless you use the right algorithm
Esteve Almirall: We are oftten so focused on new algorithms that we see them as magical. Which is more important, algorithms or data?
Xavier Amatriain: That’s another very hot debate. I think the answer is that you need a combined approach to both. The best algorithm cannot be trained on useless data and the best data doesn’t actually work unless you use the right algorithm. The thing that people sometimes forget is that both things go hand in hand, and you need to adapt your approach to the data, and adapt the data to the approach.
I had an example of this that I presented several times on the courses I’ve given. For example, if your data is linear because you’ve engineered your features in a way that they are linear, and all of the sudden you try a GBM, then you might wrongly decide that GBM doesn’t work because you’re getting the same results or worse. But what you don’t realise is that the reason it’s doing that is because you have prepared the data for a linear model, so the non-linearities that the GBM introduces are not going to matter.
If you redo your feature engineering and start injecting non-linearities in your features and add more complex features into the model, it turns out that GBM beats the logistic system. You basically have to increase the complexity of your data when you increase the complexity of your model, otherwise it doesn’t matter. That’s why both approaches need to go hand in hand, and you can’t think independently about one or the other.
When you increase the complexity of your model you have to increase the complexity of your data
Esteve Almirall: That's interesting. A problem many companies face is how to make products using machine learning. Many companies try to transform functions that are already available into models that are put into the cloud. By doing this companies gain almost zero marginal cost, infinite scalability, immediate deployment... How different is the current research on machine learning compared to the actual production of machine learning? The DevOps movement seems to go in this direction but it seems there is a lot of tension there.
Xavier Amatriain: There has always been a big separation between what was presented in research conferences and what was introduced in products. I think that nowadays the distance is becoming smaller and smaller. The reason for that is the huge availability of open source frameworks, which are democratising access to this kind of ready-to-go platforms in production. The other reason is the availability of those tools in cloud environments. You can go to these platforms and they will have machines and infrastructures with those environments ready to go. It is much easier than when you had to set up your own server.
So yes, there’s still distance between what’s done in research and what goes into production – particularly the scale and cost. Some of the models we see in research take huge resources to train. There’s no way a company is going to spend that much money in, for instance, graphic processing units in the cloud to make it happen. But on the other hand, once those models have been trained, because of transfer learning they could be used in production.
Data science and machine learning are going to be even more revolutionary for business in general
The distance between research and production has reduced. Another reason for that is because most advanced research right now is coming from industry, from companies like Google, Amazon Web Services, Facebook and Microsoft. This makes things ready to use in production sooner than before.
Esteve Almirall: One more thing about Esade. We have included AI, particularly machine learning, in most of our programmes. What is your advice in this area for business schools like ours?
Xavier Amatriain: I taught in a business school for a couple of years and I think I can empathise with the kind of things you are trying to teach your audience. I think data science and machine learning are going to be even more revolutionary for business in general. Not only for technology businesses, but for any business.
One of the interesting things that I’m seeing here in Silicon Valley is that many people are starting in companies as data scientists, and they eventually become business managers and executives. The reason is very obvious: now business is about understanding data more than it has ever been in the past. Many decisions are now based on processing data, understanding trends in data, inferring models from that and making projections for the future.
Understanding data and data science for businesspeople in general is key. Even further than that, I think understanding the capabilities of technology – like AI and machine learning – is also important. The way to approach it, from my perspective, is to demystify what AI and machine learning are. AI is about automating processes that we used to do in a manual way, and now we can do much more efficiently thanks to automation. If you think about it, it’s not much different to the industrial revolution, but applied to knowledge and decision-making, which is huge.
If you think about it this way you come to the conclusion that data, data science and AI/automation should be injected into the whole curriculum of anyone who is going to be a business leader in the next few years.
Esteve Almirall: Indeed we are facing a new revolution of data that is characterised by AI. And also, a new revolution in healthcare. If you think of the new revolution in healthcare, think of Curai – they are probably going to be one of the leaders in this revolution. Thank you very much for this enlightening conversation.
Xavier Amatriain: Thank you for having me Esteve, always great to talk with you.
Join the Do Better community
Become a member and enjoy our free benefits. Get recommendations, receive personalised content in your inbox and save your favourite articles to read later.