“There is no development without research. To make our efforts in exploring new lands in machine learning more structured, deepsense.ai has launched its Research and Development Hub,” explains Tomasz Kułakowski, CEO of deepsense.ai.
deepsense.ai’s Research and Development Hub consists of four full-time researchers supported by a technical team. Supporting deepsense.ai’s business goals, the Hub has a fixed research budget funding projects that bridge the gap between enterprise and academia.
“Sometimes AI research projects seem distant from our everyday experience. But it’s actually just the opposite. The research we are doing today will power the autonomous cars and sophisticated prosthetic limbs of the future“, says Kułakowski. “Projects based on Atari games give us benchmarks on applying new techniques in computational-heavy projects done in more heavyweight simulators.”
The team has conducted numerous successful research projects in cooperation with global leaders including Google Brain, Intel, a leading car manufacturer as well as top universities and scientific institutions around the world.
deepsense.ai is currently focused on model-based reinforcement learning and transferring models from a simulated environment to the real world (sim2real). Reinforcement learning is a driving technology behind the large strides that have been made in AI, including recent success in Go, Chess, Dota 2 and StarCraft II.
Model-based reinforcement learning and learning in simulation address the challenge of collecting voluminous real-world data. That challenge does not exist in board games such as Go, Chess and e-sports, where one can collect an equivalent of millions of years of human gameplay and learn from such extensive datasets. It is, however, a major obstacle for real-world applications of reinforcement learning, such as robotics.
In a recent project – Model-Based Reinforcement Learning for Atari – deepsense.ai set out to improve the quality of video models used in model-based reinforcement learning. The family of neural networks developed in this project are the state of the art in model-based reinforcement learning. The researchers were invited to present their work at Oxford University, Google Brain and DeepMind.
Another project consists of training in Unreal Engine 4 and deployment on a real car. The objective of this research is to assess the feasibility of training a fully functional autonomous driving agent in simulation with only a minimal amount of real-world data. The project is being conducted in cooperation with a leading global car manufacturer.
“Long-term investment in research work is a distinguishing feature of deepsense.ai. Over the past three years, we have written a number of papers and presented at important ML venues including NeurIPS. All of this is done with a firm conviction in mind: machine learning is changing and developing so fast that today’s research can be tomorrow’s gold standard,” says Henryk Michalewski, head of deepsense.ai’s R&D team.
More of the research and development projects deepsense.ai has undertaken can be found on its recently published R&D Hub subsite.
Delivering the future
Apart from scientific work, the R&D Hub will make AI research more accessible and comprehensible for non-engineers and people who are unfamiliar with data science.
“Our aim is to provide readers with as much information on our work as possible. Research should not be a hermetic game. It is about building a beneficial future,” comments Kułakowski. “I am proud that deepsense.ai is contributing to the development of this fascinating discipline.”