By Eduard Bolmosov
Today we had an amazing presentation by Jason Yosinski, Co-Founder & Machine Learning Scientist at UBER AI where he focused on understanding the technology of machine learning.
Jason’s speech was very interesting and started with a graph showing the exact phase we have now in tech implementation.
Jason says that the one which allowed to raise machine learning was a computation power which increases every year, and the amount of data which people learn to collect. All these factors made it possible to build different models for any dataset you would get. And as in any science there is a capability to understand the technology by the people: how do we get it? How do we use it? Is that technology used for good or for bad?
The tech understanding develops in a different tempo, it all started when the understanding was higher than the actual ability to perform. Later on the computation and datasets were developing and meeting the point of efficiency for the technology, but nowadays it is getting more inefficient because the tech has started to overtake our understanding. And it is the great interest for a dedicated people who work with machine learning.
Nowadays we pick the phone and say — “Hello, Google” and this phrase opens for us all the power of computer abilities to search, filter, and give feedback for a request. We also build robots which learn the way they orientate in a room by learning on their faults. It was actually a great experiment for a self-learning robot. First it was taught to move forward, left, right and back as it is in kids programming tools. But in several days the unit learned to walk in a different style, the robot thought that it had some limbs disabled from operation, it was actually crawling, then dancing. And all that the computer made by itself relying on the date and patterns it had been provided at magnificent speed 9x the speed of the coding ability of an average programmer. And what? Machines are getting smarter and they are already beating us in many fields, now it is coding and programming for learning purposes.
The session took some time to review a valuable part of the Machine learning, computer vision.
Jason showed the slide describing what neurons actually are, and how they work for defining images
The neural network consists of different layers of sets of neurons. Every layer and every neuron in it is responsible for a specific template. For example you could make a layer defining colors, contrast, black and white intense. The layer of neurons define objects, like if there is a face presented, or a line, of text. So the tech works when the image of cat (for example) passes initial layers of colors and contrast, then goes to the layer defining pattern for animals, then moves into the layer of cat type and in the end it knows the exact type of a cat. It is incredible! How many types of cats can you name? 5–10, but machine can now define almost every type. It is all about the dataset.
The data is the cornerstone of the technology, who gets the more data gets the more precise labeling and in the end models.
We also need to take into consideration the regulation. Jason thinks that giving rights to decide whether to provide data or not is all their rights. At the same time it is important to solve the problem of having data monopoly for giants of IT markets. As there shouldn’t be the time when 2–3 companies in the world can rule it just by having access to the peoples data.
In conclusion we talked about the hype cycle for machine learning.
Jason thinks that in some time the hype and demand for the new specialist would decrease as the tech would get widely spread all over the world and become like programming, people who were coding previously just would learn how to build neural, and this would be a rise for ML in the 21st century!
It was an absolutely interesting presentation and I hope many people have a chance to visit one like this to get more understanding of machines, their abilities and more importantly our abilities to rule them and guide them to the better world —and not the dark time of Terminator antiutopia series.