IRIS.AI Your Science Assistant
Research Discovery and Open Science with Artificial Intelligence
What if scientific research could be incrementally better? Better of course is hard to define. Perhaps it would be more accurate to say different, in the sense that scientific research to some extent has been closed off behind paywalls for those that can afford to access (rich universities and R&D). The library has been the go-to-place for navigating research for some time.
More researchers are using new tools to navigate research. One of these tools is Iris.ai with the mission statement:
Democratizing science. AI tools and decentralized applications for an open, transparent and unbiased world of science.
Story of Iris.ai
In this changing landscape of research, Iris.ai currently seems a very likely protagonist.
Iris in the Greek mythology is a goddess bringing messages from the gods to the people. She is also known as one of the goddesses of the sea and the sky. Iris links the gods to humanity and travels with the speed of wind from one end of the world to the other. If you have been briefly within the scientific field you may understand the metaphor and its disturbing presence in the current way research is being conducted. Science is almost like an inference from the gods and divine intervention (money) needed to get a message across in the scientific field.
The landscape of research has been dominated by a series of large publishers with larger profit margins than Google (30–40% at time according to the Guardian). Rsearch has been limited by the unreasonably high prices which has made it available only to an elite few. Science has for the last few decades been stuck behind ‘paywalls’, an arrangement whereby users have to subscribe to gain access to a website. This was perhaps an upgrade from the previous expense of libraries, yet maintained a steep pricing. Indeed a continuing trend into the digital transformation of parts of society, despite digital divides present today.
There is a hopeful change to this with universities saying no to large publishers, however there are also severe repercussions as Elsevier pulls back subscriptions due to arising disagreements. As an example now in 2019 a German institute is hitting a paywall 10,000 times per day after their subscriptions were pulled. While the University of California in March 2019 dropped its $10 million a year subscription to Elsevier. Let us however jump back a few years.
Waves of open research were already rolling when the idea of a science assistant was conceived during Singularity University’s Global Solutions Program at NASA Ames Research Park the summer of 2015. Their challenge was to bring forth an idea that could impact the lives of 1 billion people. A common frustration was the access to scientific knowledge. Shortly after their meeting and after finding the early vibe as a startup Iris.ai posted their first blog post ‘hello world!’, a classic programming joke used to refer to most tutorials in different programming languages.
We are thrilled and excited about this journey we embarked on this previous summer. Until now, we haven’t had much to say to the public — we’ve kept in close contact with our co-creation partners and potential investors, but besides that kept our heads decently low, with a small exception of some teasers here and there and a participation in “Ideas from Europe” in December.
-January the 8th 2016, official Iris.ai blog
Shortly, 2 days after the first post, a video from an idea pitch was released to the Iris.ai blog. The video was a pitch to Ideas from Europe, and I made a transcript from the video that you can read underneath.
I believe that we have already discovered the solution to most of our pressing problems. These solutions come in the shape of puzzle pieces. A research paper here, a TED talk there, a conference presentation in Chinese; some scribbles on the back of a napkin. They’re scattered and there’s no way to bring them all together. The abundance of information and knowledge we have as a human species is unprecedented, but our human brains do not have sufficient processing power to process it all. In short we cannot read and understand everything. Every single day 3000 research papers are published within science, technology and medicine. 3000, not to mention the millions that are already out there, we cannot process it all.
So, these papers end up in the digital equivalent of a dusty old drawer. Half of all papers published are read by less than five people globally. What if we had a brain a really big and really powerful brain that could read and understand all of this research, well that is what we are building with Iris.ai. Iris is an artificial intelligence that will read any scientific text, extract the concepts, cluster them for context, match them against the full body of knowledge and present it to a user in a way even a non-technical users understand. In the short run iris.ai will soon be able to help entrepreneurs and innovators navigate this science. In a five year perspective Iris will be able to connect the dots herself and puzzle the pieces together herself responding to specific requests. And in a ten year perspective Iris will be able to not only train human-beings in science, but also other artificial intelligences and robots.
The cool thing about this is that you no longer need to have a PhD to solve a real problem. Seven years ago I was a theatre student, today I am a serial tech entrepreneur and the CEO and co-founder of Iris that is the world we are living in and the wave we are riding on. We are a young startup we have only been around at four months. Here’s the thing we have immense amount of traction we started at Singularity University, we’re collaborating with Bare Healthcare, Alta University, Chalmers University, Stockholm Resilience Centre and about a 100 entrepreneurs to co-create the product. The reason advances in artificial intelligence and the projected ones makes the tech feasible, we have a prototype that will be out during the year; it is a secret, but we have been admitted to a very prestigious accelerator stemming from Silicon Valley. Iris.ai is a product that needs to be built, it is not going to be easy, but we are the right team to make it. It is time to puzzle the pieces of this puzzle together. It is time for the artificial intelligence to take all of the knowledge we have, make it accessible to our limited brains and help us implement the solutions we already have.
– Anita Schjøll Brede talking at Ideas from Europe in December 2015.
The prestigious accelerator that Anita mentioned during her pitch was 500 Nordics. This started January the 18th in 2016. Not long after the first images of their solution started appearing.
At the time in another blog post another team member was attempting to explain how it worked. Arguing that the algorithm was designed to learn like a human brain. It does reading of texts and then does elaborate frequency analysis over the text (so, yes text analysis with machine learning techniques). This here described to be done through a voronoi treemap.
-Post on the 15th of February by Victor Botev.
So a frequency analysis taking into consideration the ‘dynamics’ of the words in the text and their context. They started with text analysis on TED talk transcripts and were in the beginning unable to approach scientific research papers. At the time they felt the need to get someone to help training ‘her’. In early 2016 Iris held a talk at TEDxBinnenhof.
In March they held a talk at Chalmers University both in classes and research institutes. Discussions were as described by Iris.ai about the time wasted doing mapping studies and literature reviews. Mapping studies are useful for identifying blind spots. Their argument at the time was to help people with less time to spend on research to get an overview, and perhaps it could help scientists in the future as it became more advanced. The four co-founders of Iris.ai were then Anita Schjøll Brede, Jacobo Elosua, Maria Ritola and Victor Botev.
Jacobo wrote an important text the 31st of March on Why AI Wants to Be Open. At this point they were already asking who would reap the benefits of the progress. As I have mentioned in a previous post on inequalities and AI this is a question that most AI companies I have seen do not deal with this to the extent perhaps necessary. Then again the premise for Singularity University was to make a solution that would impact the lives of one billion people, it follows a clear narrative. In this he had a clear critique of how intellectual property is managed:
That traditional IPR system has come under growing criticism over time, but as professor James Robinson –co-author of the highly refreshing and thought provoking essay ‘Why Nations Fail: The Origins of Power, Prosperity, and Poverty’– reiterated at a recent conference, that patent-based system could be categorized as an inclusive economic institutions largely beneficial to society as a whole. It had the merit of aligning effort and reward.
Merit according to Jacobo had become more fuzzier and complex in the world of science. He was asking how to recognize the role and value of AI trainers. As a startup how could they get a critical mass of researchers to train the algorithm. As described in a post by Victor shortly after they were struggling to find out how to teach a 4th grader university-level language.
Anita was importantly as well questioning the sex of the tool/application/assistant/company. Why are most of the applications in the field of AI female? She was mentioning the failure of Microsoft Tay and calling out the ‘brogrammers’ in Silicon Valley sharing the ambition that Iris.ai should not be servile, yet become a scientist in her own right. Again, they needed AI trainers. In April 2016 they were launching an AI crowd training platform. AI trainers were asked to look at TED talks and see the results from Iris to:
- Validate the concepts
- Make corrections
- Suggest new concepts.
That May they started their collaboration with Our Future Health an e-health conference. In July Iris.ai got its first pre-seed investors, angel investors.
We are happy to announce that we have officially closed our pre-seed investment round, having raised $350,000. This will bring us way beyond our next launch and has allowed us to double our team over the last few months. Tharald, Sean & Stina, Thomas, Sjur & Sjur, Øyvind and Philipp — we are thrilled that you believe in us and excited to have you join us on our journey.
-Anita Schjøll Brede on Iris.ai 15th of July 2016
This can be confirmed on Crunchbase. They got the resources to kick off further engagement. From having a so called minimum viable product (MVP) to launching Iris 2.0. In the Autumn with new UX, neural network (7% improved results) and supervised learning. Iris.AI then addressed natural language processing through the in-house implementation of a novel neural model.
In their natural language processing they went from LDA model (Latent Dirichlet Allocation) a generative statistical model often used to classify text to a non-semantic neural topic modelling.
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… 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.
Iris had started to do scithons an interesting mix between the hackathon, a design sprint-like event in which computer programmers and others involved in software development together with a diverse scientific community. In December 2016 Iris.ai was selected as one of the 13 most promising early-stage companies from more than 600 applicants around the world to present at TechCrunch Disrupt in London. They also held a scithon there and defined the concept.
Jacobo started being hired out to companies to help apply a combination of human and AI to solve open-ended science challenges.
In January 2017 Mark Zuckerberg and Priscilla Chan’s $45 billion philanthropy organization made its first acquisition. The Chan Zuckerberg Initiativeis acquiring Meta, an AI-powered research search engine startup, and will make its tool free to all in a few months after enhancing the product. TechCrunch quotes Meta co-founder and CEO Sam Molyneux writing that “Going forward, our intent is not to profit from Meta’s data and capabilities; instead we aim to ensure they get to those who need them most, across sectors and as quickly as possible, for the benefit of the world.” The Toronto-based startup was funded with $7.5 million by investors, including Rho Canada Ventures and HIGHLINEvc. They decided to make the product free.
In February 2017 Iris covered this as great news for the entire academic community. Maria congratulated in a blog post writing that they were working in different parts of the research process: “Meta helps researchers keep track of the most recent articles in the medical field by using citations graphs, our AI Science Assistant focuses on reading and understanding the contents of research papers across different fields.” The acquisition of Meta certainly brought to light the attention on AI within research and Iris made it to the Forbes list of most innovative companies 2017 within artificial intelligence, this only 18 months into running their startup.
They started training from research paper abstracts in early 2017 with another goal.
Our next goal is to gather and inject a trained dataset of 5000 paper abstracts to the algorithm. With those inputs we aim to improve the connections in the neural nets of Iris.AI by approximately 10 %.
-The New AI Training tool is here
Iris.ai had started new partnerships in health and running more scithons creating research maps such as these:
Iris joined Founders Factory in London in April 2017, being one of the first AI startups in their portfolio. The Scithon was taking the shape to be a model that could be implemented and run in different cities. Still focused on the medical field, Iris held a few scithons in collaboration with Stryker (a Fortune 500 medical technologies firm).
Another year another version of Iris.
They said less about how they had upgraded their models, however it seemed an exciting development. At around this time they launched their own token, tokenizing science. Again repeating their value-driven goal of changing the science industry and the problem of paywalls and the high profit margin of large publishing houses.
For us to truly make impact in the world, it is not enough to build some great tools, we need to disrupt and uproot an entire industry. We can not do that on our own — it’s a grassroots challenge. We need your help.
-Anita Schjøll Brede on the 13th of December 2017
They were taking a step towards a Knowledge Validation Engine, a core feature of the AI Scientist, around this time introducing the new focus feature for premium users. Thereby visualising the time reduced through using Iris.ai through a visual display. A video introducing the tool can be seen here:
What was not mentioned on the blog, was that Iris had closed another round of funding. Their Seedround had three investors for a total of $2,000,000 in funding the 1st of December 2017.
Early 2018 Jacobo Elosua made one of his first posts on Medium: For Open Science, but up a different path exploring the movement of open science. The article ended up with Iris’s solution Project Aiur:
With Project Aiur, we now have the goal of creating an open, community-governed AI Engine for Knowledge Validation. In the brave new world we envision, any knowledge seeker should be able to input a scientific text, be it an existing research paper or a self-written problem statement, and query the system, Aiur, to get a number of related outputs, including a validation of the input’s hypotheses and building blocks against all of the world’s existing science.
Thus combining open-source with blockchain (buzzword bingo!) in a way an experiment in open science. Yet in a way possibly a useful way to utilize blockchain as community distribution, starting to distribute Aiur tokens in May 2018.
The project was launched with a website https://projectaiur.com/ and a whitepaper focused on knowledge-validation (released in June 2018). They started the campaign by distributing the currency through an ‘Aiur Airdrop Campaign’ to researchers and students if they decided to participate. After which followed some technical specifications on the Aiur token distribution in June.
White paper: A white paper is an authoritative report or guide that informs readers concisely about a complex issue and presents the issuing body’s philosophy on the matter. It is meant to help readers understand an issue, solve a problem, or make a decision.
Anita explained that Project Aiur was a not-for-profit endeavor from the Iris.ai founders. However, that it was a vital component for the success of Iris.ai that Project Aiur exists. Yet discussing also the balance between commercial interests and impact as challenging. In a blogpost Jacobo described the progress in three milestones:
- Focus on extracting the main structural elements contained therein: i.e. Problem, Solution, Evaluation and Result descriptions
- Use this extracted information to build knowledge graphs
- Incorporate information from images, tables, graphs, etc. (i.e. ‘black data’) into these graphs
The goal being to increase transparency and accountability. Additionally attempting to solve some of the most commonly cited issues: (1) information overload; (2) access barriers; (3) reproducibility issues; (4) built-in biases; and, (5) incentive misalignment. This launch led Iris to coming into contact with thousands of researchers and keeping the conversations going on Telegram, and in their Facebook group.
Then in 2019 came Iris version 4.3 with the integration of PubMed, a free search engine accessing primarily the MEDLINE database of references and abstracts on life sciences and biomedical topics. This added millions of papers to the possible search results. There were other changes to the upgraded version:
- The ability to edit a problem statement after creating an exploration map.
- Clarifying language in the Focus tool UI.
- Improved keyword extraction for more relevant results.
- Improved reliability and security.
In March 2019 during the climate protests around the world Iris.ai announced that they wanted to fulfil one of the goals stated in the economic forum of making information more useful. Additionally sharing different maps of research while asking what actions to do with all the research. Around this time joining an initiative to collaborate in the Nordics (Anita being based in Norway) together with other startups in the area. This initiative was called the Nordic AI Alliance.
Towards the end of April 2019 Iris announced that they would be ending their AI training program arguing that they needed to take a step back and review what they had learnt from the process. Iris 5.0 was launched the 9th of May 2019 with a relevance score feature making it easier to navigate different cells as well as getting an easier visual overview of relevance.
In the month of May Iris released another blog article about training stating that:
Iris.ai’s learning is largely unsupervised but has used the training platform as a form of validation of her assumptions as an AI. With the help of our AI Trainer community, we were able to look at her assumptions against real, human inputs to ensure she was working as effectively as possible. With your help, we collected over 8,000 validation points against which to test Iris.ai. Thank you! Of course, the training platform as it existed before today used questions and methods that validated the previous algorithms. Now that new algorithms are in place, it is time to reassess the training platform to determine how our training community can assist in validating Iris.ai’s brand new algorithms. For the remainder of this year, we’ll be working on a revamped training system that will serve Iris.ai and our community of users and trainers even better as our little AI scientist grows.
So that was the story of Iris so far as of June 2019.
I hope you enjoyed the read and hopefully we can discuss more the implications of science assistants going forward.
This is day 16 of my project #500daysofAI.
What is #500daysofAI?
I am challenging myself to write and think about the topic of artificial intelligence for the next 500 days with the #500daysofAI.
This is inspired by the film 500 Days of Summer where the main character tries to figure out where a love affair went sour, and in doing so, rediscovers his true passions in life.
Hope you enjoyed the read!