AUBG Students Dive into Artificial Intelligence with Centroida and The Hub
October 03, 2017
Artificial intelligence (AI) does not sound like a fantasy anymore, as it is slowly becoming a part of our lives. On Sept. 28 the Andrey Delchev Auditorium filled with young programmers, who came to find out how deep learning methodologies and AI can be used to construct neural networks.
Stefan Lazov and Rim Mustafin, software engineers at Centroida and students at AUBG, presented on behalf of the company, together with The Hub.
“Deep learning is completely different from ordinary programming concepts we are familiar with,” Lazov started out. “We used to think about programming as studying algorithms to solve a problem, but such approach was limited with the human factor.”
He pointed out that now the traditional approach is now being replaced with machine learning, artificial neural networks, reinforcement learning, and deep learning. Later each of the categories was introduced with examples.
“In reinforcement learning, we are approaching machines as we would approach humans. It is fascinating how it works,” said Mustafin. He gave an example of a cleaning robot, which is programmed to receive a negative score for walking around, and a positive one for picking up trash. Just as little kids learn what is good or bad, the robot will learn to walk around faster, and clean better.
The presentation continued with insights on what Centroida is working on. Lazov shared the machine learning and deep learning libraries they work with, such as Theano, Caffe, and TensorFlow. Adversarial samples, inputs to a machine learning that cause the model to make a mistake, are another area of interest for the company.
“At Centroida we managed to design some adversarial samples to fool famous deep learning classifiers,” Lazov said, demonstrating their work on the screen.
He showed pictures of a wolf and an elephant, and their adversarial samples. They looked identical at first sight and the classifier was supposed to only recognize the original images. But the adversarial samples created by Centroida misled the classifier, making it focus on the background of the image or mistake an elephant for a bird.
The presentation finished with an active Q&A session, where Lazov and Mustafin shared more of the research and projects they are working on at Centroida.