Hybrid Intelligence Research Laboratory
We build the next generation of intelligent machines that learn, explain, and predict the world as humans do
HINT Lab is a hybrid intelligence research laboratory with the long-term vision of using AI technology to expand human capabilities beyond natural limitations, thereby, enhancing cognition, amplifying intelligence, and boosting human productivity.

We build autonomous machines that can understand the real world around us just like humans. That means not only recognizing objects, but also being able to analyze situations, understand their real meaning, and then communicate this understanding to humans in a natural way.

* hybrid: augmented natural and artificial intelligence
Research Groups and Areas
We take inspiration in models from different fields of natural intelligence science, including computational linguistics, neuroscience, psychology, and cognitive science. Then, we combine these models with machine learning methods and virtual simulations to teach machines to reason and make decisions in an unsupervised manner.
Robotics, Planning, and Vision Group
Neuroscience studies show that high-level features of human intelligence are based on primitive but essential, functional building blocks shared with all mammals. These include but are not limited to an intuitive recognition of 3D spaces, attributes and relations between objects, real-time mental, physical simulations, and planning. Building computational models of these functions is the primary imperative of the robotics and vision group.
Current research activities
Learning to Dream: Online Time-Constrained General Domain Model Learning, Planning, and State Tracking Using Mental Simulations for Prediction and Explanation
While current ML and RL systems require millions of examples to learn successful acting strategies, humans show exceptional "one-shot" learning abilities. For example, when two-year-olds encounter a new problem, they quickly recognize the problem's configurations and successfully solve puzzles in a few trials after observing a few solution scenarios. Many findings, proposed by cognitive science and getting empirical support from neuroscience, promise the explanation.

First is the idea of learning as the rapid model building that promotes the effective reuse of already learned models, and next is the idea of mental simulations—the ability of mammals to simulate action outcomes in the mind, without resorting to slow trial and error in reality. Building on these ideas, we create a new breed of ML algorithms and study their properties in a variety of domains.
Real-Time Scene Understanding via Digital Reconstruction and Simulation
Semantic reasoning and the recognition of dynamic 3D scene situations go beyond current deep learning systems by providing full digital twin copies of the perceived world and its dynamics, including physical kinematic systems and human dynamic behavior, allowing such high-level capabilities as an explanation, prediction, and communication for human-computer interaction and autonomous decision making.
Computational Dexterity: An Artificial Hand with Fine Motor Skills for Real-Time Precise Interactions with Solids, Deformables, and Substances
Human fine motor skills are a fascinating example of natural intelligence that is unmatched in its sophistication and dexterity by any other living thing on earth. With the sense of touch, humans can feel, explore, grasp, move, mold, intertwine, and relate other objects to one another, as well as make free gestures and convey information. Studies show that a direct link exists between high-level cognitive abilities and fine motor skills due to shared areas of the brain that are involved in both the processing of motor information and cognitive tasks.

Our goal is to build working models of hand that can precisely interact with most real-world material: solids, deformable objects, and substances.
Language, Logic, and Relational Reasoning Group
Despite the recent success of supervised machine learning methods in speech recognition, no progress has been shown in the automatic understanding of the meaning and logic of language. Unlike speech recognition, natural language understanding (NLU) struggles with feasible computational representations and unsupervised learning methods.

At the same time, cognitive and neuroscience field studies of language bring insight into the mental models the brain uses for language comprehension. The findings hint at the importance of imagination facilities, mental simulations, innate psychology representations and reasoning. By building on these insights, the group works on computational models of language and knowledge representation, fusing machine learning methods with constraint-solving and relational logic reasoning approaches to bring new ideas to the field and achieve human-level language understanding.
Current research activities
Multi-Agent Belief-Based Rational Communication for World Knowledge-Sharing and Coordination by Imagination-Equipped Agents
We study computational models of human communication as a dual process of conveying an internal state of mind by a speaker and a reverse process of inferring that state by a listener. While the human brain information space is measured in petabytes, conveying information via language reduces throughput to tens of bytes per second, yet people successfully encode thoughts in language and can restore meaning from it.

Such effectiveness is achieved by mechanisms of shared world belief space, shared world domain models, and the ability to recreate in mind new configurations of models "constrained" by language but never perceived before by participants.
Abstract Conceptual Systems in Human and Machines - Missing Layer Between Language and Vision
Human language is inherently abstract. Over 70% of all lexicon is used in the abstract sense. However, representation and learning of abstract concepts remain a mystery for the linguistic and ML community. The current body of research is scarce and mainly avoids the subject, but instead focuses on more mentally accessible concrete language cases where word learning is supported by perceiving examples of objects in the world.

We argue that most basic abstract concepts (both nouns and verbs) can be learned in the same manner and by using the same facilities as concrete words but using self-introspection facilities of an agent as instances and proposing the computational model of a process.
Algorithmic Management and Decision Making
Combining the power of artificial intelligence, natural language understanding, and virtual online workforce, we research and build the next generation of end-to-end automation solutions. They will replace the old automation software stacks by providing a single artificial agent governing all operations in organizations on behalf of stakeholders by collecting business data, making effective decisions, coordinating the teams, and communicating with people in natural language.
Current research activities
Algorithmic Teams: Real-Time Human Teams Supervision and Coordination by AI Agents in VR Simulations and Real-World Scenarios
We study mixed human-automation cooperation in complex physical, time-pressured or high-risk work scenarios, where the AI dispatcher is given real-time control and supervision over human task force teams. The system performs a contextual assessment of the situation by receiving a sensory data stream of the environment from its sensors and verbal communication with a team member, and then accomplishes inter-team coordination by planning and communicating actions for each team member.
Algorithmic Organizations: Distributed Autonomous On-Demand Organizations Governed by an Algorithm
Modern organizations possess an ancient legacy of hierarchical traditions of governance structures that are rigid and slow to change. Business processes are usually initiated and designed by hand and are prone to human error and oversimplification due to complexity. The problem worsens when organizations grow in headcount.

We propose a fully autonomous system able to perform usual management duties such as hiring and planning in a fully automatic, unsupervised way. The system is capable of on-demand scaling by automatically hiring experts in real time using online labor markets and reconfiguring workflows on demand by splitting complex work into microtasks performed through crowdsourcing platforms such as Amazon Mechanical Turk.
Open Fellowship Program and Admission
We invite people across the globe who want to leave their mark in AI science. Whoever you might be — a mature data scientist, machine learning professional, junior software developer, or just a student — you are welcome to join us on-site or remotely.
Your own research
Those of you who are mature scientists are proposed to drive research and suggest your own topics. You will be given an assistance team and you will be free from marketing and technological costs and tasks. We will help you with research and publishing. This allows you to focus on science and implement your ideas easily.
Great team
You will be surrounded by bright minds from various fields of science and from all over the world. Your leader will be a professional mentor that will help you go through the research, provide feedback and help gather new information relevant to your research. Experience shows that it extremely boosts professional growth.
Easy time management
You can try researching without risks related with the necessity to change your job. Even a 10 hours/week contribution will give you enough answers on whether you like this, whether you feel free with such a process and whether you want to dig deeper and tie your life with a science. But also this can remain just your hobby.
Unique experience
The experience of taking part in new-fashioned online distributed research is not only fun, but will widen your working skills in this new reality.
Science papers
Taking part in preparing of science papers. Your name will be mentioned in paper, what may skyrocket your career.
Open to new ideas
If you have some thoughts on our fields of research, you can test them. We will provide you with all that you need, including assistance
Do you want to join us?
Our Vision: Roadmap to Intelligent Machines

Best-performing AI systems are based on a set of assumptions about the current environment (i.e., one that has predefined models). However, the main problem here is that engineers cannot describe the whole world because of its complexity.

So, each algorithm solves problems only in a very specific and narrow domain, where it knows the rules of the game. As an example, an algorithm that can win against the best professional chess players can't recognize human faces, as it knows nothing about humans.
Our approach is entirely different. By combining ideas from different scientific areas—from cognitive science to cutting-edge machine learning methods—we build a technology that, in the first instance, can understand the real world without predefined, specific models but can also learn new models rapidly from interaction with the world and with people.

This approach allows us to move one step closer to the 'real' AI systems, which are capable of general problem-solving and planning in unspecified domains.

Eugene Legkiy
Co-founder and Chief Science Officer
Ukraine-born Israeli. Math-savvy, serial tech entrepreneur. Early-stage tech startup investor and advisor. AI \ ML practitioner, visiting MIT, and MIPT scholar. Eugene founded and led a number of companies till successful exit, one being the largest AI-tech company in the Russian advertising market: Segmento.

Programming from the age of 11 and falling in love with the idea of AI from then on, he spent more than 20 years pursuing AI research and development endeavors.

Combining knowledge from artificial intelligence, neuroscience, and cognitive science, Eugene leads and defines the agenda for fundamental and applied research activities at HINT Lab.
Anton Popovich
Co-founder and Chief Technical Officer
Russian-born tech leader and entrepreneur. Experienced software developer, team leader, system architect, and project and product manager with multi-year experience as a chief technical executive in a number of AI startups.

Started programming at the age of 6. A graduate of the Computer Science faculty of a technical university. Multiple prize-winner in algorithmic programming contests. Inventor of an algorithm in prime strings theory ("Finding Lexicographically Least Cyclic Shift of a String with Linear Complexity") that outruns all existing counterparts in speed.

Anton and Eugene first met and worked together from day one at an AI-tech company, Segmento, where they created a new breed of AI technology for the advertising market. Combining system architecting and team management experience, Anton leads technology implementation in HINT Lab.
Contact Us