My Experience at Khipu AI 2019
About Khipu
Khipu, the Latin American conference in Artificial Intelligence, was inspired by Deep Learning Indaba, a meeting that’s been happening since 2017 in Africa. It’s last edition took place in November at the Facultad de Ingeniería, Universidad de la República in Montevideo, Uruguay. It was the first time that Latin America hosted an AI event of such scale and importance. Hundreds of researchers and professionals from all over the continent joined together with the purpose of empowering research in the field. An effort that was absolutely necessary.
To make it happen, a large number of people and sponsors had to get involved. It is well known that being a researcher in a developing country is not easy. In Brazil, for instance, a master’s student monthly salary offered by federal government is equivalent to $350 dollars and for a Ph.D, $520. This makes the academic career very unattractive, to say the least. The emotional and psychological pressure are huge. Very often students get frustrated with their work. They feel alone, they feel the social pressure and they feel undervalued. Not to mention that they always have to answer to the “When are you going to get a real job?” question.
Spending a week surrounded by so many interesting people was certainly motivating for each one of us. We felt part of a community. Moreover, it was also very important to be around so many successful researchers. They shared their stories and gave us many advices. A succesful academic career seemed like a less impossible goal and also totally worthy and rewarding. It inspired me to continue with my studies and, more importantly, to share my knowledge with the ones that need it. I felt the importance of sharing my experience with my local community and also of being closer to it. This is how I hope to repay to what I learned at Khipu.
Program
Khipu’s “official” routine happened from 08:30 to 19:00, but had many unofficial events. The schedule basically consisted of theoretical lectures interspersed with Spotlights and Sponsor’s Talks until late afternoon. Afterwards, we had parallel practical sessions in which we had the opportunity to choose between two topics or to participate in the Hackaton.
The program was pretty intense, so I’ll just highlight my favourite moments. To see the whole schedule and have access to videos and slides, please access Khipu’s program.
The first day of Khipu was mostly dedicated to Machine Learning and Deep Learning fundamentals. My favourite part of this day was the practical session of Optimisation for Deep Learning. Of course, I can’t say that I’ve mastered this subject, but I felt relieved because now I feel more confident about hyperparameter tuning for optimisation.
Like I said earlier, Khipu had many unofficial events. The organizers worked very hard to entertain us with multiple surprises. The first day of Khipu ended with a Tango performance at Anfiteatro del Edificio Polifuncional José Luis Massera. It was awesome, and the week was just starting.
My favourite parts of Day 2 were Kyunghyun Cho’s lecture on Recurrent Neural Networks and the panel “How to Write a Great Research Paper” with Nando de Freitas, Claire Monteleoni, David Lopez-Paz and Martin Arjovsky.
The panelists shared precious advices: the tips went from text stylistics to editing in collaboration and guidelines. They also shared examples of great papers to inspire us.
The best moments of Day 3 were related to the topic of Reinforcement Learning. Unfortunately, I arrived at Khipu knowing very little about this subject, but Nando presented an awesome lecture and I got very interested in studying it more deeply. The practical session afterwards fit like a glove.
To finish Day 3, one of Khipu’s major sponsors, Tryolabs, threw a great party at Plaza Mateo Rooftop & Bar.
It was amazing to enjoy such a beautiful sunset and get together with so many great scientists. Those feelings of solitude were far distant by now.
After such amazing three days, I thought it would be very hard for Khipu to continue to surprise us — but I was very wrong.
At Day 4, Chelsea Finn gave a fascinating talk about Robotics and Continuous Control. I’ve always been interested in robots, but my studies took me in a quite distant direction. So it was good to learn more about this subject.
David Lopez Paz gave one of my favourite talks ever. His presentation held my attention from the beginning. I highly encourage you to watch the video. In this talk, David guides us through the history of causality and how it relates to correlation.
To finish Day 4, Google AI hosted the event Women in AI with the panelists Sandra Avila, Chelsea Finn, Maria Simon, Giulia Pagallo, Guillermo Moncecchi and Nando de Freitas. Remember what I said about that feeling of isolation that researchers often feel? I believe it is much worse when it comes to female researchers. The lack of representative female figures makes it feel like this career is not for us. But the room was filled by amazing women and this motivated me even further.
Now it’s time to talk about the last day of Khipu. It’s tough to choose the best moments of such an incredible day. I’ll start by saying that I was looking forward to watch the parallel session on Advanced NLP with Oriol Vinyals Video, Jorge Pérez Video, Luciana Benotti Video and Lucia Specia Video; and I wasn’t disappointed. It was one of the most important sessions to me since this is my research area. This session basically summarized all the hot topics on NLP right now.
Next, Oriol Vinyals presented the exciting project AlphaStar: StarCraft II using multi-agent RL (Video). By now, I must say that if Khipu would’ve been just this morning, I’d be totally satisfied. But it went on!
During lunch I had the chance to present my poster. It was a great opportunity to talk about my master’s research. Khipu happened two weeks before my defense, so it was the nicest way of concluding this stage of my life. But sadly I didn’t have much time to see other posters that were being presented along with mine.
After lunch we had to choose between two parallel sessions: i) AI for Social Good with Jeff Dean Video, Danielle Belgrave Video, Cecilia Aguerrebere Video Slides, Alejandro Noriega Campero Video, Guillermo Sapiro Video and ii) Life of a ML Startup with Mario Guajardo, Agustina Sartori, Martín Alcalá,
Thiago Cardoso, Matthieu Jonckheere.
I chose to watch AI for Social Good and I was pleased to see that so many researchers are working so hard on social problems.
And finally, the last talk at Khipu: Deep Learning to Solve Challenging Problems with Jeff Dean (Video). Jeff is currently the lead of Google AI and most of his talk's been structured around a publication from 2008 by the U.S. National Academy of Engineering in which there was a list of Grand Engineerging Challenges for 21st Century. With this, Jeff introduces this list of 14 problems and mentions that Google is currently working on 10 of them. Nevertheless, he selected 5 of these to share their progress with us, namely: i) Restore and Improve Urban Infrastructure, ii) Advance Health Informatics, iii) Engineer better medicines, iv) Reverse-Engineer the brain, v) Engineer the tools for scientific discovery. During his talk, Jeff also communicates that it would be very interesting to see what researchers outside of Google could do with more computational resourcers. And then, once more we've been surprised by Khipu. We got free access to cloud TPUs to support our research. Unbelievable.
After the talk, we've been invited to an awesome closing party with dinner and music from a local band. It happened at Club Uruguay, a club located in a historical neighbourhood in Montevideo (Ciudad Vieja). It was incredible.
I feel so grateful for this marvelous week and city. Uruguay entered to my list of favourite places in the world. I met a lot of great people. I connected to my local community of AI researchers. I watched very important talks and got numerous priceless advices. I presented my research to Latin America (something that would have been quite unpredictable two years ago). I had a lot of fun. And most importantly, I came back to Brazil very inspired and with renewed energy to continue with my enthusiasm for AI.
Thank You:
First I want to thank my supervisor Marcelo Barra, who first introduced me to the AI field. I'll be forever grateful for this.
Secondly, I want to thank Renato Vicente (a.k.a my boss at Experian Latam Datalab) for the incentive and support. His recommendation letter was very important for my participation at Khipu. Thank you for the trust.
I also want to thank Khipu organizers and sponsors for making this possible. I’ll never forget this week.
