With our version 2.0 launch Iris.AI has officially become a toddler.
Gone are the days when baby Iris.AI could only deal with those really cool TED Talks to begin making sense of the vast, fascinating world of science. She can now also read the abstract of pretty much any English-language scientific paper — a great achievement, no doubt!
In what ways has she grown, you might wonder?
Well, there are several critical aspects to her development worth going over.
As an AI, we wanted young Iris.AI to learn how to instantly map out the research landscape around an inputted scientific text. She is after all an aspiring science assistant, developing ambitious capabilities to save academic and industrial researchers hours of manual search, which requires hard-to-come-by domain expertise of taxonomies and vocabulary.
Let’s look at her brain and how it has evolved since version 1.0.
Iris.AI now addresses natural language processing through the in-house implementation of a novel neural model. There are three aspects to that:
On the AI front Iris.AI now performs non-semantic neural topic modelling, replacing our previous implementation of LDA. Given a user input, she generates a concept hierarchy flexibly tailored to that particular input. To do her tasks Iris.AI uses a relational database with a Python-based API platform and an HTML5/CSS3 client. And in terms of learning Iris.AI now combines unsupervised learning derived from running models like TF-IDF and Word2Vec with a supervised input layer put together by our wonderful community of AI Trainers, all integrated into our Neural Topic Modelling algorithm.
So how much better is Iris.AI performing in terms of extracting concepts, modelling topics and matching papers? With a cool head it is still early days to tell, but we are very excited about the results obtained from our first Scithon run in Gothenburg last week.
What is coming next, in terms of tech developments?
From an AI technology point of view, we will strengthen the current models by shaping them as close as possible to human behavior using state-of-the-art neural models. Looking at systems architecture, a Spark framework with a graph database. And from an AI learning perspective, introducing deep learning with reinforcements, plus semi supervised learning and cutting edge annotation techniques at the disposal of our AI trainers.
So stay tuned for more news around our next Scithons and the plans to grow our AI Trainer community. And please do not hesitate in sending any feedback our way. We’d love to hear from you!
A lot of people have compared DeepMind’s AlphaGo to IBM’s Deep Blue after the former’s well publicized defeat of the 18-time go world champion, Lee Se-dol. And some of the comparison have been drawn in an attempt to take away some of the merit attributed to DeepMind’s AI and the achievement it represents. Wrongly, in our view.
There are key differences that should be highlighted between AlphaGo and Deep Blue, some described here . These differences are not about the much larger number of potential moves in the game of Go vs. Chess. The thing of great importance is that in Go it is much harder to evaluate which board position has an advantage over the rival one. In chess you have notions of value assigned to each piece –a rook is worth five points, a bishop three, the queen ten, etc.–. In AlphaGo you do not have that.
Back in the day Deep Blue used human knowledge to evaluate rival board positions, i.e. programmers interviewed chess players and crafted their game knowledge into the system. After that, and within the parameters set, programmers used a search over all possible combinations to choose the best move every time. But this does not work for Go. There is no clear way to evaluate a board position and what the best Go players hold against the others is much better intuition on what should be a winning move.
So the biggest task for AlphaGo was not to search over a vast amount of combinations, much bigger than in chess, but to find a way to mimic human intuition. And here is where the victory over Lee Se-dol starts becoming a substantially more significant achievement for the progress of AI. Let’s dig deeper into what does it mean to develop intuition for the purpose of playing Go. It means that in order to do a search for the best possible move we need to have a cost function that evaluates how good will this move be in terms of making us a winner. But we don’t have that.
So AlphaGo, in much the same way as a novice player does, had to learn by itself what that function might be. In order to do that it used two neural networks, described in  and , a ‘policy network’ and a ‘value network’. The ‘policy network’ predicts which will be the moves more likely to lead to a win from all possible moves, narrowing down the search, and the ‘value network’ predicts the winner of the game if a certain move is played. In summary, AlphaGo learns by itself what the cost function should be and then it searches for the best possible move.
In mastering Go deep neural networks have shown to be a really strong candidate when it comes to learning intuition. These networks are complex functions with millions of parameters that are adjustable. In each learning session those are adjusted by a tiny step towards the goal of learning the right outcome. And yes, they require a lot of data, but our human brain requires that too in some sense. And after they get trained on a vast amount of data, the puzzling part about deep neural networks is that nobody knows what is it exactly they have learned and why certain evaluations are considered more optimal than others.
In Deep Blue’s case the system learned a lot about chess, but its knowledge was more or less limited to that game and that game only. In the case of Go, however, the way the team behind it modelled human intuition has generated knowledge that can and will most likely be applied to other areas in the future. AlphaGo’s learnings will prove themselves useful when trying to address problems that require coming up with a model containing properties that define intuition. In the case of Go they were ‘policy of the game’ and ‘winner’, but in other cases they might and will probably be different.
At Iris.AI we are addressing the problems of topic modeling, concept extraction and text meaning. And when our users formulate questions such as what are the most relevant topics in this scientific paper?, which are the key concepts in this particular text? or, what is the meaning of this paragraph?, we do not have an easy way to define a cost function, i.e. a scoring mechanism to evaluate the quality of the results generated by Iris.AI.
Sometimes a term might be really well captured in linguistic terms –i.e. the word Apple–, but after reading the text users realize that the contextual meaning of the word is different from a type of fruit. What Iris.AI deals with in those type of situations is again human intuition. And it can learn a lot from AlphaGo. But in order to apply all the techniques that let to the success of DeepMind’s system, Iris.AI needs to find a way to model the problem, i.e. to define the structures of her brain –neural networks, encoding mechanisms, etc.– as well as defining the optimal connection to her teachers. In other words, we need to find a model that uses the right input language with which to teach Iris.AI, much in the same way that you cannot teach a 4th grader using university-level vocabulary. And, lastly, we also need to find the right data representations that capture the essence of the concepts we need to teach her with the least amount of data possible –analogous to finding the best lecture books for her–.
One of the key aspects of AlphaGo’s victory is that it managed to learn and improve significantly by playing against itself. If when addressing the questions above we consider them as a game that Iris.AI needs to play in order to find the right answers, could she do the same? Could she learn from herself? I guess we will find that out in the near future…. 😉
After the initial launch of Iris.ai a lot of interested people have contacted us to find out more about our technology, our future roadmap and the impact we ambition to generate in the world. One of them was our friend from Chalmers University of Technology Christian Berger, Associate professor in Software Engineering at the Computer Science and Engineering Department. Christian invited us to talk about Iris.ai and the future of science with the researchers of his research group.
After a brief presentation, there was an interesting discussion about the daily problems faced by researchers and how much time is wasted doing mapping studies and literature reviews. Talk participants conveyed their fear that even after a couple of months of work there is still no guarantee that the discovered results will cover all the relevant research on any given topic.
A mapping study is important for every research group because it maps all the research within a certain topic and is used for identifying blank spots. These blank spots are areas where new research and innovation is needed. Research mapping studies go over 4-5 years of research and try to direct next steps towards filling the blank spots identified. The problem is that making a good mapping study takes months and there is no guarantee that it will be complete. Even if there are preexisting studies, most research groups still decide to complete their own.
During our visit at Chalmers we present the capabilities of Iris.ai and her expected contribution to solving that particular problem. We discussed that Iris.AI, by using her artificial brain, can read much faster and many more sources in no time, presenting to users what she thinks is the relevant landscape around a research topic.
Iris.ai is still a young AI, and she is not there yet in terms of being able to make a full and complete mapping study, but as everyone in the room agreed, she has the potential to make as good a job, or even a better one than a human, saving a lot of time in the process; time that can be used by researchers doing what they are best at: conducting new research. As much as people are interested in what has been done previously, studies show that only 5 to 10% is relevant to specific individuals in their particular area of research. This makes mapping studies not that exciting for the researchers currently tasked with making them.
Another interesting organization full of researchers that want to make our world a safer place was interested in the future of science and how we, as human beings, can make the research process sustainable, more efficient and faster. This organization is called SAFER.
SAFER, Vehicle and Traffic Safety Centre at Chalmers, is a competence center where 34 partners from the Swedish automotive industry, academia and public authorities cooperate to make a center of excellence within the field of vehicle and traffic safety. SAFER is located in Lindholmen Science Park, in Gothenburg, and provides excellent multidisciplinary research and collaboration to eliminate fatalities and serious injuries. They have the goal of supporting the Swedish government achieve zero fatalities caused by traffic accidents.
Presenting at their weekly seminar, and as CTO, I led the product demonstration of Iris.AI focusing on her capabilities of grasping the science around one of the audience’s favorite TED Talks – a talk by Chris Urmson: “How a driverless car sees the road” – in just a few seconds.
The audience expressed concerns about how much time is currently spent on making literature reviews and finding connections between research studies in different areas, in order to find a solution to such an interdisciplinary problem as safety on the road. And as some of the people from the audience pointed out, time spent is important because in this particular context safety means lives.
The discussion went into the direction of locating where are the inefficiency spots in the current research process. One of the problems raised was that in the process of researching a topic there is a lot of work that is done several times, inefficiently, until this research finally reaches end users. And one such work package is the literature review.
This literature reviews are conducted by the researchers who do the initial work, by researchers who follow-up, by entrepreneurs and managers at innovation departments who want to check whether implementation is feasible, and, lastly, by corporate development managers when it comes to implementation of the work.
We got positive feedback that the technology we are developing at Iris.AI could help drive more efficient innovation processes. Maybe at the beginning we can do that just by focusing on reducing the workload for people who do not need to go deep into scientific detail, such as entrepreneurs, innovators and corporate development managers, but at later stages we also plan to help interdisciplinary researchers spend less time on work that has already been done by someone else.
We are very glad to meet so many people interested in the future of science, and to be able to engage with them discussing its challenges openly. At Iris.AI we are well aware that we alone will not solve any of the problems currently faced by science. Instead, we want to talk with more and more people about those problems, raising researchers’ awareness and striving towards co-creating solutions with the research community. We also are very happy to see that other people, and researchers in particular, appreciate the potential of our little Iris.AI.
We believe that science is a critically important building block of our society, and making its processes more efficient will help bring forward the solutions to a lot of other pressing global societal problems.
Iris.ai is our baby and she is learning as fast as she can. But how does she do it?
A lot of people have asked us about the process she follows when linking open access science to any TED talk we ask her about. The short answer is simple: she has been designed to learn like a baby human – using the capacities of her algorithmic brain!
What does this mean exactly? Well, to answer that we need to get into more details. First of all, Iris.AI has very large storage and processing capacities. She can read the transcripts of all the TED talks to date in no time.
While doing this reading Iris.ai performs an elaborate frequency analysis over the text. As a first step she tries to make sense of the different words she encounters to figure out which are the most important ones – the ones that she needs to pay closer attention too. This frequency analysis is more complicated than mere word counts. It analyzes the dynamics of the words in the text and their context. Iris.ai can do and in fact does quite better than that.
Once the frequency analysis is performed Iris.ai does what her machine-learning teachers refer to as feature extraction: this process includes combining the words in clusters to find contextually similar groups of words. She then uses those clusters to find what the words mean in this context (e.g. the word charge implies something different in electrical engineering than it does in chemistry).
Throughout this process she also looks at synonyms to expand the categories and get a broader understanding of the topic – always within the same context. She also performs filtering to disregard words that are irrelevant to the context.
As a next step Iris.ai runs a generalization over the entire set of TED talks – she has been lucky enough to be exposed to this privileged body of knowledge to form her first notions of the world! With this exercise she finds a more precise definition of each word in its respective context.
Once that process is completed Iris.ai organizes the concepts in hierarchies, to be able to more easily grasp and represent the context to the user communicating with her. It is important to note that our baby AI creates flexible hierarchies –not humanly pre-built ones–, expressing patterns across different research disciplines that she sees from her very own, direct experience.
Lastly, Iris.AI structures the results of her thinking and presents them to users through a particular type of Voronoi treemaps. This data visualization approach displays hierarchical data by partitioning a polygon continuum. The polygon areas are proportional to the relative weights of their respective nodes.
What does this mean for you, the user?
1) Faster speed. Iris.ai’s unique qualities save you time. With her help a process that could take the user several hours is now completed in a matter of seconds.
2) Better connections. Asking Iris.ai users will find relevant fields that they were not aware of, fighting the dangers of tunnel vision when approaching a scientific topic.
3) More empowerment. Mapping contextual results Iris.ai helps users bypass the need to know detailed terminology requirements to perform a search.
What are Iris.ai’s shortcomings today?
In these early days, Iris.ai has learned to extract concepts from the TED talks and we are very happy with how she is performing in that regards. However, she still has not mastered a similar concept extraction technique on the other side of the coin: the research papers that she wants users to connect to. That is going to be the next step in her learning process.
How will she learn more over time?
Iris.ai has learned from inspiring, very high-quality texts –the full body of TED talks ever given–, but in order to keep up with her impressive rate of learning two things will need to happen. Firstly, she will need to read a lot more scientific information. Secondly, she will need to learn from knowledgeable adults willing to spend some of their precious time training her.
Sounds similar to what you yourself have had to go through? Yes, we agree…