More Big Techs started to join a game. They actively recruited stellar people from academia with big chunks of money. Professors can hold both a position in a school and a lab industry. Those deciding not to join also started up their own companies. The wave had risen.
Back to Vietnam again, I was thinking about the next destination. That would be Canada. However, my skill set was not strong enough to dive deep into challenges last year, so I need to be in industry to boost my learning rate. I thought about Big Tech and some hot startups. I had made some connections to recruiters. Then, I got an offer from a startup based in Seattle - they had an office in Vietnam with a team connected to a national university.
When I was still wondering if this should be the right place, a software engineer from the Valley appeared and joined a company. His former team was a high growth startup later acquired by a Big Tech. He returned to the country because of an obligation from his PhD Fellowship.
In startup there were a lot of white label tasks. We did not have enough people. My lead was pretty interested in practicing an engineering framework shared among previous team members from Google. He did a lot of experimentation with several big tech frameworks from scratches. The learning curve was pretty rough for me especially when trying to debug in javascript and scala and meeting deadlines. It was such a huge pain on the ass for me but gosh that was exactly what I wanted.
At that time, Udacity had their first open AI course for a small group of users. I was selected and stacked all of my free time with courses while trying to replicate a tech vibe that I learned from San Francisco. With collaborations from other organizations, we opened a small series of events.
A graduate student whom I met at a party last year became a scientist working for a Big Tech in a Valley. Sometimes we kept in touch and I tried to ask him to help me connect with the Professor in Canada. I was too shy to talk to him in-person. One day, I summoned all of my courage and sent him a cold email. Two weeks later, he answered. Later, we started to exchange emails.
Before 2015, there were a few engineering frameworks to work on Deep Learning. The best known is Theano developed by University of Montreal. In parallel, there were a few more by other core academic groups and top Big Techs. After the breakthrough in 2015, most of the team in Theano joined Google, then Tensorflow was borned in January 2016. Another descendant, Torch with its own language to tackle the limitations of Theano, later made several fruitful results compared to its sibling. In 2017, PyTorch (a python version of Torch) was acquired by Facebook. Since then, the open source code was maintained by them.
One day, the scientist sent me a message. He struggled to find people and how difficult it was for him to assemble a team for a medical application. He had a few GPUs and was happy to share. Along with another team member, we gathered a group to design the first course based on TensorFlow - which was known by a small number of practitioners in the Valley. That later became a community class. The first cohort has 14 students. Only one woman. We invited speakers who were young academic graduates in Japan and the US to come and share knowledge.
At the end of 2017, there was an open virtual conference, one of the first of its kind. At that time, the idea was streaming several-hour and sessions at one place and one time. Both participants and speakers did not have to travel. I personally could not travel and became their users. Unexpectedly, a few days after the event, organizers sent me an invitation to join a brand new challenge. They said it was to celebrate a ceremony of the world's largest co-working space called Fstation. The challenge was about mining medical data provided by a top industrial lab in dermatology. In addition, every participant is sponsored by Azure - a cloud platform of Microsoft. It was a great opportunity to test my technical skills, so I signed up.
Different from various competitions in data science, there was only one submission with no leaderboard. The result will only be informed by organizers at the end of the competition. This is a blinded submission compared to several competitions in that time.
I tried both source code from Tensorflow and Pytorch. It was still a pain on the ass while working with Tensorflow and GPU setup. There were several different versions of Tensor that required different configurations. Documentation was not clear and one tiny wrong setup could could an entire process. It was a much more pleasant experience working with PyTorch. Several sleepless nights and sudden suspension are exchanged by a few minutes. However, some of the largest models are maintained and shared by the Tensorflow community.
Experiments showed that the larger the performance, the higher accuracy would be. At the end of the competition, I decided to jump into the source code of Tensorflow. Before that I only worked with pre-built Python packages. My knowledge with C++ and Javascript was underwhelming. There was a stagnation despite how long I trained as a model, so I decided to give it a try. However, it was super lucky that most of the source code is written and deployed by Google in the language that I was able to comprehend while working for the previous startup. Sweats and sleepless nights were paid off. I jumped into the top 1% of more than 300 teams that year.
Finally I was ready to jump into the next level after two years. If someone asked me about working on frontier tech looked like, it was all about documentation, debug, debug and debug.