Deep learning Indaba 2017


All that I can say about Deep learning Indaba is wow! Never in my life have I been surrounded by so many brilliant people!

Held at the University of the Witwatersrand from Sunday the 10th of September 2017 until Friday the 15th of September 2017, it was meant to be more of a refresher course in machine learning and provided a platform for data scientists to meet and mingle.

First day: For the love of math!

The first day had a stark contrast between work and play.

Opened by Marc Deisenroth explaining the mathematics for machine learning set the tone for the rest of the Indaba. All machine learning is based off of core mathematical concepts and if one wishes to innovate algorithms in this field, one should have a deep understanding of the math behind it all.

The after party caught me by a surprise. I never in my wildest dreams did I imagine scientists would dance at a party! It just goes to show, this field is still very young and so are most of the people at the conference.

Second day: Hung over, but surviving. On to deep neural nets!

Some of us partied a little bit too hard…Nonetheless the show must go on!

Adam Habib the Vice Chancellor of the University of Witwatersrand opened the second day. He truly inspired me to try and change the world through utilizing artificial intelligence.

After that Nyalleng Moorosi and Willie Brink gave a lecture on the fundamentals of machine learning.

Lunch I went to talk to some of the IBM guys. Their research is amazing. They are experimenting with quantum computers. Scary stuff…Less scary is the cancer detection research. They are simply amazing.

One of the most interesting characters that I had the pleasure of learning from was Yann Dauphin who is a Facebook AI research scientist. He has been all over the world! He draws knowledge from so many places. He showed us how the layers of neural networks interact as he demonstrated the abstractions that happen at each node. This gave me a fundamental insight about how I should reason when building models.

Third day: Nando de Freitas!!!

Nando de Freitas from Google’s Deepmind is a legend in the field of AI. He is one of the most passionate men I have ever met. When he speaks you are just a little bit intimidated by his level of intelligence. He did a presentation on convolutional neural networks. In this presentation he explained how the features are abstracted within the network layer by layer.

The second lecture of the day was on recurrent neural networks presented by Stephan Gouws and Richard Klein. I must say the math behind recurrent neural nets left me a little bit dizzy but the practicals helped me understand this structure better!

Onto the poster sessions! Here I could see how artificial intelligence is instrumental in bringing research forward. One of the more interesting projects that were presented was the use of pattern detection to detect new galaxies. This might not sound impressive, however the research needed to create the machine learning approach took a year and is able to find new galaxies in the span of 7 minutes! Just to put this into perspective, usually this would take astronomers between 1 to 3 years to make such an observation! Another project that caught my eye, is one in which a researcher simulated the brain functions of a worm. One of these days we will have digital creature models on which we can test the effects that medicine and other products will have on biological creatures…

Fourth day: Information overload

Three lectures were presented. Probabilistic Reasoning presented by Konstantina Palla, Unsupervised Learning by Alta de Waal and Deep Generative Models with Ulrich Paquet. So many things to learn! If it weren’t for frequent breaks I don’t think that I would have survived.

Fifth day: Last day of learning + Retro Party

Reinforcement learning is such a vast field it had to be split up into two lectures. The initial lecture was presented by Vukosi Marivate and Benjamin Rosman and the second by George Konidaris. These lectures were somewhat interactive since reinforcement learning is applied to robotics. Demos of robots learning how to balance and doing specific tasks always peak my interests.

The last lecture of the day was titled “Applications of Machine Learning in Healthcare” which was presented by Danielle Belgrave. She stated that one of the greatest challenges in this field is to get access to data and how to interpret results gained from models.

The closing event at Retro’s Braamfontein offices was in a sense magical. Words cannot describe so I will let the pictures of the party do the talking.

Closing day: Moral obligations and challenges for innovators

Artificial intelligence is the forefront of innovation. With that comes some consequences. People that will lose jobs as they are replaced by machines, privacy can get violated as we need more access to information. The way of doing business is changing and those who aren’t willing to integrate AI will get left behind.