During the winter months, Finland doesn’t sleep — quite the opposite. Last weekend approximately 1,500 hackers from across the globe descended upon the Otaniemi campus of Aalto University to take part in Junction, Europe’s biggest hackathon. The general idea was straightforward: form a team, and build something awesome in three days.
Nordcloud didn’t need to think twice about participating. After some brainstorming, we chose to host a deceptively simple challenge in the Artificial Intelligence track: summarise long-form news articles into easy-to-digest summaries and headlines.
Despite the simple brief, accurate text summarisation is one of the hardest problems in machine learning. Solving it well requires some real creativity, ingenuity, beefy hardware, and intimate familiarity with some pretty advanced learning algorithms. Before the hackathon, we worried we had given the teams an impossible challenge.
Turns out that our initial apprehensions were completely unfounded. A total of ten teams participated in the challenge and the results were nothing short of spectacular. On behalf of the challenge mentors and judges, I can safely state that we were blown away by the quality and creativity of the final submissions. Some of the brightest minds from the University of Helsinki, Aalto University, MIT, UC Berkeley and other prestigious schools partook in the challenge, producing some exemplary work. One team made a trending news reader that automatically clusters several articles and uses AI to create collective summaries and machine-generated headlines free of journalist & publication bias; another used word embeddings to produce a shortlist of pros and cons for any product on Amazon.com. Some teams ran tone analysis to provide users with AI-powered summaries based on mood (anyone up for some upbeat news in the morning?). One team used AI to determine if a news article can be considered trustworthy. A special mention goes to the team that built an AI that summarises text in the tone and mannerisms of a politician.
Mentors and judges from SC5 & Nordcloud watching and taking notes at the demo session.
In all, a broad set of machine learning techniques were on display, ranging from K-means and K-NN clustering to word embeddings to recurrent neural networks with attention mechanisms. Some teams even implemented, from scratch, bleeding-edge neural network architectures that have only very recently been introduced in scientific circles (see  for one such example). It’s a great example of how crucial basic research in machine learning is, and why it’s important for industry and academia to work together to advance the field. New breakthroughs and best practices make their way into production systems extremely quickly, more so than in any other discipline I can think of.
The three days of Junction flew by, and although many teams (and mentors) got little to no sleep over the course of 72 hours, everyone I saw around Dipoli was energised and smiling throughout the weekend, despite a false fire alarm and subsequent evacuation at 5am in the morning of the submission deadline. If that doesn’t inspire you, I don’t know what will.
A special thanks to all Junction personnel and Nordcloud challenge personnel for their exceptional work, and to our partners at Amazon Web Services for providing the much-needed CPU and GPU grunt needed for training such advanced AI.
 Vaswani, Ashish, et al. “Attention Is All You Need.” arXiv preprint arXiv:1706.03762 (2017), URL: https://arxiv.org/abs/1706.03762