Table of Contents
For decades, the premise of man versus machine has been the plot of countless science-fiction novels, TV shows, and movies. Most of these play on people’s fear of machines becoming self-aware or thinking and learning for themselves.
Technology has come a long way and machines haven’t taken over… yet. But they are learning. Does that mean the future will unfold as Rod Serling and Isaac Asimov imagined? To get a better understanding of the real-life science behind the fiction, I turned to Full-Stack Developer and machine learning (ML) expert Petr Bruna.
It’s a quiet Friday afternoon. I message Petr that I've arranged our interview in Toyen, one of several meeting rooms at Salsita—all of which are named after Czech artists (though, looking back now, the room named after Josef Čapek would've been the most appropriate given the focus of Petr's work).
Toyen is cozy and feels the least like an actual meeting space. And, since Petr was apprehensive about being featured, I wanted him to be as comfortable as possible.
I got the initial conversation flowing by talking about guitars, music, and some of his favorite bands while he sat with his oversized green mug. I wasn’t sure if it was mainly for drinking or keeping his hands warm. Either way it seemed to do the trick.
LEAN, MEAN LEARNING MACHINE
Petr studied physics and computer modeling in science and technology. He received his master’s degree during his five years at university. As computer modelling can be used with machine learning, it came in handy for his future endeavors. Although the main aspect was physics, there was a strong focus on how to build these models.
According to Nvidia, “Machine Learning at its most basic is the practice of using algorithms to parse data, learn from it, and then make a determination or prediction about something in the world.”
Petr has been interested in machine learning since his time at university. When he first heard about it, he thought it was cool. Petr jokes that you just input data and it will do some crazy magic. At that time, it was a shiny, new topic; a great concept to tap into.
“It is always fascinating what can be done with simple ideas,” says Petr. “When it works properly, it’s simple to explain and you can do a lot of things with it, which is the same thing that got me interested in physics.”
As Petr continued, I could see his eyes light up with excitement. He had put down his comfort mug and was gesticulating with both hands, emphasizing his points. It was great to see him start to dig into what he loves talking about. This enthusiasm and excitement are what separates OK teachers from great teachers. “Petr can describe the basic principles with great examples and makes it easy to understand such a complex topic,” says Full-Stack Developer Alex Khrapko.
LET’S DO LUNCH
The machine learning lunches started with Salsita Founder and CEO Matthew Gertner. “I wanted to drive home the fact that we are serious about ML and start to encourage knowledge sharing,” says Matthew. The first four or five lunches were dedicated to discussing interesting papers and how to apply them at Salsita.
Unfortunately, the group never got past the first paper. The one Matt had chosen to discuss was too advanced and some of the people attending hadn’t learned the basics of ML yet. After the class, Petr asked if they’d be interested in learning the basics from him. They were.
“A course turned out to be a better way to get everyone up to speed both on the basics and on more advanced techniques,” says Matthew. During the first two weeks of December 2018, Petr prepared slides on ML basics. Since then, he’s been teaching twice a month to anyone who wants to join.
“Many of us have AI degrees but Petr's knowledge and experience are unmatched in the company.” – Full-Stack Developer David Kuboň
BASICS, PRESENTATIONS, AND HOMEWORK
During the first few months, the ML lunches had a steady group of 10. It grew a bit more but has started to decline since. Petr attributes this drop off to busy schedules and “maybe the homework” he says with a smile. “Petr gave the ML lunches structured content, which is why I enjoy them more now,” says Salsita CTO Roman Kašpar.
As the class changed, so did the material. The beginning was about theory and the basics. Now the lunches are being taught from a more practical point of view. Since the lunches move progressively, Petr says it may be difficult for some people to join in now, having missed out on vital information and basic ML theory.
Don’t fret though. Petr keeps his presentations regularly updated, so anyone can see them. He prepares everything in Google Docs and adds more when needed. The group also has a Slack channel for continuous discussion. “The presentations aren’t recorded or formal, and the slides are just the tip of what we talk about” says Petr. “I encourage people to interrupt me with questions.” His presentation has 104 slides and is growing every month.
Petr, like most teachers, spends a lot of time putting together his lesson plan. It took longer to understand the basic concepts at first, so those lessons were spread out over a couple of months. That’s not the case now. What once took two hours to plan between lunches is now up to six to eight hours. “That part is becoming a bit of a chore,” Petr says.
MOVING BEYOND THE THEORY
With a lot of theory behind them, the group is now focusing on the math behind the basic building blocks, how to combine them together, what they are good for, and how they can be used. From here, they speak about commonly used architectures for neural networks and the properties and ideas behind them.
There isn’t an end date in sight for these lunches, but Petr has goals beyond passing on his knowledge. He wants them to be able to create their own models and architecture and be able to modify them as they’d like; to use it in real life. “I can talk a lot, but there needs to be a practical application for them,” says Petr. Currently Petr is the only one of the ML team, but a couple of others will soon be joining him. They have a background in ML and will also help to tackle augmented reality and other projects.
When asked about how to gain practical experience outside of these lunches, Petr put it simply: get data. Getting data is the most difficult part about ML. You can’t do anything without it and the smaller the data set, the harder it is to do. Petr says you can have beautiful models, but if you don’t have good data, who cares?
PRACTICAL USES FOR MACHINE LEARNING
Petr applied ML for the proof of concept for a project for a large perfume retailer. He went into their store and after 30 minutes he had taken 1200 photos of various perfume bottles. He’s now compiling this data into a simple Android app and seeing if it will behave as it should in real life.
“Petr’s extremely knowledgeable and he puts a lot of time and effort into his presentations with many illustrations and examples. He’s also been spending a lot of time showing us real code and explaining how it works. He even gives us homework so we can get our hands dirty.” – Salsita CEO Matthew Gertner
Machine learning is being used by companies to comb through massive amounts of data and analyze it quicker than any person could. “There is a lot of work to gather the data, but that’s common across any project,” says Petr. “That’s why it’s helpful to have a broad spectrum of tasks.”
A common way companies use ML is by recommending services, music, or movies, based on previous purchases and selections. Netflix’s recommendation engine was given a makeover when they held a $1 million contest to make it more accurate. Amazon is another site you’ll see this recommendation engine in action. Their message reads, “These recommendations are based on items you own and more.”
As far as ML’s future, Petr feels it isn't advancing fast enough.“Probably because a lot of people would lose their jobs,” he says. Jobs have tended to migrate to low-wage areas where employing humans is actually cheaper than automation, but as wages go up in those areas we'll see automation taking over.
Petr says the same thing will happen to AI within the next 10 years. More and more people are getting involved with AI, which makes it spread faster. AI is very narrow. Every model can only focus on one thing and can’t do anything else. The focus will be to make it broader so it can be used across different tasks.
Another thing regarding ML and AI is that it will become common to see ML-powered apps running on smartphones, as many companies are starting to realize. They are making chips that can accelerate the computation quite significantly without consuming much energy. Chances are, in a couple of years, those chips will be everywhere.
As for the ML lunches, Petr continues to teach anyone who’s eager to learn. He’s always looking for people interested in the ML classes, and not just for the free lunch. The first year of machine learning just wrapped up. The attendees are excited to put their knowledge to use in practical applications, and we’re looking forward to growing our ML team and making their skills and experience available to our clients.