In this post, we discuss the limitations to the traditional data collection approach and illustrate how to use crowd-acting, an innovative…
Nov 30, Twenty Billion Neurons
Machines today can identify objects in images, but they are unable to fully decipher the most important aspect: what’s actually happening…
Oct 10, Twenty Billion Neurons
How we built a robust gesture recognition system using standard 2D cameras
Sep 14, Twenty Billion Neurons
A bespoke storage format for deep learning on videos
Sep 8, Valentin Haenel
Understanding what neural networks see when classifying videos
Aug 22, Raghav Goyal
How we placed third in the 2017 ActivityNet challenge.
Aug 11, Twenty Billion Neurons
Wide Range of Applications
Use cases range from human gesture recognition to security video monitoring
Custom Datasets
We own some of the largest industry datasets for intelligent video analysis
Runs in Real-time
Our software extracts meaning from continuous video streams in real-time
Flexible Licensing
Licensing options range from one-off royalty to subscription models

Our mission

When humans perform tasks and solve problems, they rely heavily on their common sense knowledge about the world. A detailed understanding of the physical world is however still largely missing from current applications in artificial intelligence and robotics. Our mission is to change that. We are developing new, ground-breaking technology that allows machines to perceive the world like humans.

How it works

Our technology can analyze human actions and extract real-time information from video streams. We build deep learning systems using our large video datasets about common world situations and then fine-tune them to specific use cases with minimal effort.

A novel database

One of the limiting factors for advancing video understanding is the lack of large and diverse real-world video datasets. To circumvent this bottleneck, we have built a scalable crowd-acting™ platform and have created some of the largest industry video datasets for training deep neural networks.

A unique approach

Our deep neural networks are pre-trained on our datasets of crowd-acted videos. The datasets contain short video clips that show a wide range of physical and human actions. We then transfer the capabilities of our trained network to contribute to specific video applications.


Step one

We pre-train deep neural networks on a foundational dataset for understanding physical actions. A large amount of annotated video data is required so that the models develop an internal representation of how objects interact in the real world.


Step two

We transfer this intrinsic knowledge to solve problems of high complexity that rely on an understanding of these fundamental physical concepts. The data requirements are drastically lower, allowing us to solve a large variety of video use cases.

Use Cases

Our machine learning systems excel at deciphering complex human behavior in video

Gesture recognition
Gesture Recognition
Automatic detection of dynamic hand gestures for human-computer interaction
Social robotics
Personal Home Robots
Visual scene understanding for domestic robots interacting with humans
Elderly care
Elderly Fall Detection
Automatic detection of accidental falls among elderly people at home
Aggression detection
Aggression Detection
Automatic detection of aggressive behavior in public transport systems
Theft detection
Theft Detection
Automatic detection of suspicious activity like theft or shoplifting in stores
Video based ad suggestions
Video-based Ad Suggestions
Media system that automatically places ads based on the video's content
Textual video search
Textual Video Search
Media system that allows users to search and discover video content
Video content moderation
Video Content Moderation
Media system that automatically detects and removes offensive video material
Our Data Factory

Our Data Factory

Through crowd-acting™, we are constantly growing our large-scale datasets that help machines to see and understand the world.

Learn more
Data factory

Our Core Datasets

Tailored to the needs of product groups and industrial R&D labs

Our core dataset is used to teach machines common sense and basic physical concepts
The world's largest video-based dataset for reading dynamic hand gestures
Human Actions
Our scene understanding dataset is used to detect human behavior and complex actions in context

About us

We are a technical team that is re-defining how machines understand our world

Dr. Christian Thurau
Chief Biz Dev Officer & Co-Founder
Roland Memisevic, PhD
Chief Scientist & Co-Founder
Dr. Ingo Bax
Chief Technology Officer & Co-Founder
Valentin Haenel
VP of Engineering
Susanne Westphal
A.I. Engineer
Moritz Müller-Freitag
Chief Product Officer
Raghav Goyal
A.I. Engineer
Joanna Materzynska
Junior A.I. Engineer
Héctor Marroquin
Crowdsourcing Supporter
Farzaneh Mahdisoltani
A.I. Researcher
Waseem Gharbieh
A.I. Researcher
Guillaume Berger
Research Intern
Till Breuer
A.I. Engineer
Melanie Metzner
People Manager, Executive & Finance Assistant
Florian Letsch
A.I. Engineer
Nahua Kang
Product Intern
Sarah Rose
HR Manager, Executive & Finance Assistant
Check our Open Positions


Yoshua bengio
Yoshua Bengio
Montreal Institute for Learning Algorithms
Nathan benaich
Nathan Benaich
Peter yianilos
Peter N. Yianilos
CEO Edgestream Partners
World map

Contact Us

Send us a message

Contact information


Twenty Billion Neurons GmbH
Oppelner Str. 26/27
10997 Berlin

+49 30 5564 3880 |


Twenty Billion Neurons Inc.
111 Queen Street East
South Building, Suite 450
Toronto, Ontario, M5C 1S2

+1 647 256 3554 |