What is machine learning and what is it for? It is a question that many people ask themselves but in Signaturit we have resolved thanks to our partner Guiem, responsible for the services we are developing based on this technology.
In this post we wanted to interview him to tell you what machine learning is, what it is for and what his work in Signaturit consists of.
This post is also available in Spanish.
What is machine learning and what is useful for? Meet our Machine Learning Engineer, Guiem Bosch
Guiem was born and raised on the beautiful island of Menorca. His childhood reminds him of Mowgli, spending time with animals on the farm where his parents worked, who he would also help to make artisanal cheese.
Later a question would arise: How could he understand his mind if the tool he used to explore it was his own mind? At that stage he was an idealistic child and decided that the best way to overcome that limitation was to create an artificial mind that would help him understand his own.
When he was 14 he started doing yoga to get a better understanding of his existence and the human mind. 20 years later and he admits that he’s still not that clear on it, but he is sure that yoga has made him a much more loving person!
I studied Computer Science because the curriculum featured a subject called Artificial Intelligence and I also got a degree in Psychology to get a further perspective.
Although he’s been working on Artificial Intelligence projects for 10 years, he’s surprised that it’s become one of the “hottest” jobs in the tech world.
Let’s find out why everyone is talking about this technology from someone at the forefront of it.
«Some people worry that artificial intelligence will make us feel inferior,
― Alan Kay, computer and writer
Guillem Bosch' s favorite quote
1. What are we talking about when we talk about Machine Learning. Can you give me the simple definition :)
Machine Learning encompasses the computing systems that are able to improve on their results in a set task automatically. This progressive improvement is what we call learning and, in general, it is a result of giving the system large quantities of data/experiences.
To the untrained eye, the algorithms that make up the mathematical model of the data may seem complex, and even intimidating, but the underlying idea is simple.
Just like in humans and other animals, skill and knowledge are the product of a series of repeated experiences. Except for some instances where learning stems from one simple event (e.g. don’t touch fire because it burns), humans learn and improve through exposure to new cases and repeated experiences.
Machine Learning operates on the same principle, an image classifier that distinguishes between cats and dogs improves as it sees more photos, just like a system that learns to play chess can win more games as it gains more experience by playing against itself.
2. How would you explain the benefits of Machine Learning to a basic user that hates computers? How do you think Machine Learning can help society in general?
First of all, in the short term the uptake of machine learning solutions has ushered in a paradigm shift in programming. There are certain problems that are impossible if you have to codify them in the classic way. They require the explicit detection of thousands upon thousands of rules. Machine Learning doesn’t need the rules; it just needs a supply of data to learn them by itself.
In the long term, machine learning can be helpful in almost any area you can think of. It’s actually already helping. However, let me digress a bit and stress the responsibility we all have as the creators and users of technology.
In the field of machine learning it is especially important to be fully aware that machines learn from the data supplied and obtained from interactions with humans. The potential of machine learning to help wider society is enormous, but just as astonishing advances appear so do racist chatbots, sexist recruiting algorithms and so on.
As S.Polyak said: “Before we work on artificial intelligence, why don’t we do something about human stupidity?” I don’t think we need to be insulting about ourselves, but I do agree on the need for a solid base as a species before we start blindly delegating various responsibilities to machines.
Guiem concentrated while we took the picture.
3. It seems like technology from the future but how is it being felt in the present? What uses cases can we find today?
Well, there are unlimited applications in the market today. Every time someone asks me about what it can do nowadays I recommend they visit kaggle.com. This is a platform where the world’s leading Machine Learning/Data Scientists come together to find solutions to challenges with real data.
- Quora: data analysis to predict if a user question on a Q&A social network is sincere or just trolling.
- TwoSigma: use of news to predict stock market movements.
- Earthquake prediction.
- Petfinder: predict how quickly a pet will be adopted.
- Google: predict how much their consumers will spend.
- Detect power outages in pylons.
- Detect whales in photos for population studies.
- Predict house prices.
- Detect cancer.
- Reduce accidents on commercial flights.
- And countless more.
4. Tell us about the start of year career and how you discovered Signaturit.
Whenever I’m asked this, I feel like an old man telling war stories. When I started out in the world of machine-learning, just over 10 years ago, the number of applications for machine learning was quite small. Completely different to the amount of data you can dive into today!
Applied artificial intelligence was usually linked to military projects, though I was lucky enough to work as a researcher at university on an European assistive technology project. I took part in the development of a wheelchair control system for a person with cognitive difficulties that was able to autonomously steer the wheelchair.
The trick was to get the wheelchair to help with steering only when the user needed help (e.g. someone suffering from hemispatial neglect as a result of a stroke who has problems in turning left).
It was a really rewarding project because at that time I was also just starting to study psychology and that allowed me to better understand user needs.
Coming back to the present, I think I first discovered Signaturit in the same way as lots of other people do: without even realising. I was looking for new professional challenges when I came across a job offer for Machine Learning Engineer. At that point I realised I recognised the logo and then a little bit later I remembered that someone had sent me a document to sign using Signaturit’s solution two years earlier.
I have it on good authority that I’m not the first worker at Signaturit to use the platform before working here. I love that, the contrast: on the one hand the process with digital signatures was so natural and sudden that it almost went unnoticed. But on the other hand, you realise just how disruptive it is (a legal solution available everywhere).
That’s why my memory of having used Signaturit before was almost subconscious in a way, I was really drawn to that kind of subtle ubiquitousness in Signaturit’s technology.
5. What is your specialisation in Machine Learning?
Although I’ve worked in a wide range of disciplines, ranging from those closely linked to robotics and computer vision to jobs on knowledge representation, in recent years I’ve been focused on Natural Language Processing.
Just before I started at Signaturit I was analysing millions of social media comments to monitor trends that would enable predictions on tourism dynamics.
Trying to get machines to understand the language of human communication is a discipline that I absolutely love. Apart from my personal interest in language, I think that it will be one of the pillars of artificial intelligence in the future.
6. How would you describe your job? What is your Machine Learning project at Signaturit?
Thousands of documents needing to be signed come through the Signaturit platform every day. Although at first glance it may seem impossible, there are many variables (time of day, day of the week, etc.) that allow us to predict whether a document will be signed or not. The algorithms of machine learning are especially effective when dealing with the huge amounts of data that occur in the signature process.
So, my first objective is to obtain a robust system that makes reliable predictions about whether a document will be signed or not. In a second phase, if it ´s predicted that there is a risk that the document is not signed, we work on a system that can offer suggestions to the user to increase the probability of signing.
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement Nº 778550
7. How do you think Machine Learning will have advanced applications in 5 years time?
Well, I think there are two possible scenarios. The first, more conservative one sees a continuation of the steady rate of change. Which isn’t slow because there is a constant stream of new solutions, strategies, and implementations that improve on previous results in fields such as computer vision, natural language processing, and so on.
That leads to the increasing penetration of machine learning into our daily lives, better virtual assistants, better and fairer facial recognition systems, better automated driving systems, better algorithms for medical diagnosis, and so on.
The other scenario involves the possibility of there being an advance that is a clear break from everything that’s come before. Although the functionalities of the latest Machine Learning advances have revolutionised the technology sector, it is clear that we are still far from achieving a general learning ability that can be compared to humans.
In fact, Geoffrey Hinton himself, considered the godfather of Deep Learning, has suggested we set aside everything done so far and start again from scratch, “the future depends on some graduate student who is deeply suspicious of everything I have said”.
If there is some groundbreaking discovery that leads to new forms of machine learning, we could move into a very different evolutionary framework where changes and improvements happen in a completely different timescale. So in 5 years time things could be happening that are just unimaginable at present.
Guiem in one of our team meetings.
8. What is your goal/dream/target in Machine Learning?
I would love to live in a time in which machines can really connect words with meaning. To date, natural language processing systems are mostly based on co-occurrence. This is known as distributional semantics and is based on a hypothesis that holds that words which are used and appear in the same contexts share the same meaning.
Obviously words like “dog” and “cat” appear in more similar contexts than “tiger” and “lion”, which means they can be grouped together by meaning. You can download Wikipedia, feed it into a neural network, and create representations of the words according to their co-occurrence.
That has allowed us to “play” with words in a way that was just unimaginable previously, to the point that you can do arithmetic operations with language (a famous example: “king - man + woman = queen”).
However, we are still far from creating a system that, as well as being able to use the word “red” in the right context, can understand the subtleties of meaning that we as humans are able to envisage (e.g. “warmth”, “passion”, and so on in the case of the colour red).
Seeing a solution to the problem of how words (symbols) are linked to their meanings would be a childhood dream come true! It would surely be one of the most revolutionary changes in the history of humanity because a machine being able to understand the subtleties of human language would mean we would have solved some of the mysteries of consciousness.
9. What do you most appreciate about working at Signaturit?
Without a doubt, being able to create something that is technologically sound and innovative in such a fresh, clever, and dynamic environment. Working in an environment like that is a real pleasure.
Just like with the finished product, mutual respect between colleagues and the feeling of belonging come so naturally that you tend to take it for granted. However, behind it all there is a lot of hard and painstaking work that goes into making it like that. From the very first minute, you can sense the great collaborative spirit of the whole team.
Bringing it back to my area of work, I really appreciate the drive and ambition that has been put into innovation. Independent thinking, having an eye on what more we can do to improve what we already have.
Learning in Machine Learning is based on trial and error, on testing lots of models until we find the one that works best. All of this comes together perfectly at Signaturit, where there is always room to experiment.
10. What are your biggest challenges at Signaturit?
There is a saying in applied machine learning circles that states “machine learning in a company is 10% Data Science and 90% other challenges”.
I’m not going to get into how accurate those figures are but I do agree that a large part of the job isn’t strictly related to using/creating learning algorithms. After years of experience, I think the biggest challenge outside of academic settings revolves around finding a balance between exploring and exploiting knowledge.
I’ve seen entire teams researching and improving models that are completely useless when finished because the market has moved on altogether.
That’s why it’s important to always be aware of the client’s needs, be hands-on with product development, and not switch off in a world full of big questions, which can undoubtedly be intellectually rewarding but not very useful in the context of a company project.
11. What are the characteristics of a Signabuddy for you?
Let me be a bit mysterious... I’d invite anyone who wants to get to know us to come visit and find out for themselves what makes a Signabuddy ;)
Signaturit kickoff last January
12. What recommendations would you give to future Signabuddies and future professionals in your sector?
I would advise them not to let this opportunity pass them by. We are in a period of expansion but we are always careful to keep our energy and dynamism intact. “Come and see”.
One specific piece of advice for future professionals in the sector would be: always try to work on projects that you find inspiring. Don’t be frightened by how difficult it might seem, working on something you love always pays off.
13. How does the mind of an engineer that wants computers to think work?
I have no idea, that’s what we’re working on... For me, I find being a bit eclectic useful, I combine an almost computer-like analytical mind with a bit of creativity and thinking outside the box.
The good thing about machine learning is that all of us involved in it have very different backgrounds and motivations. My take on it, as you can see, verges on the metaphysical but that doesn’t stop me having a great time every day on the project we’re working on here.