Bridging the gap between academic research and at-scale development
Currently, there is a significant gap between specialized academic research on deep learning and the at-scale application of these concepts to solving real problems. At Merantix, our goal is to aggregate concepts and technical research and use them to create scalable and relevant deep-learning solutions. In order to reach these goals, our team is closely connected to researchers at major academic institutions, with whom we collaborate on various topics. We recognize the importance of bringing groundbreaking research on deep learning to life and we strive to realize with our constantly evolving products and solutions.
Applied Research Topics
How software agents ought to take actions in an environment so as to maximize a suitable notion of cumulative reward.
Provides a deep learning framework which can also model uncertainty. BDL can achieve state-of-the-art results, while also understanding uncertainty.
Debugging neural networks can be challenging, even for experts. With millions of parameters involved, even one small change can cause errors. At Merantix, we develop methods to visualize neural nets to make this process smoother.
Learning continuously and adaptively about the external world and enabling the autonomous, incremental development of ever more complex skills and knowledge.
Compression of deep neural networks for memory-efficient, compact feature representations becomes a critical problem when these networks are deployed on resource-limited platforms.
Well-annotated datasets can be prohibitively expensive and time-consuming to collect. Our research explores the usage of larger albeit noisier datasets that can be more easily obtained.
Our team is experimenting with the modularization of machine-learning pipelines in order to achieve maximum efficiency throughout complex processes.
Experimentation is a crucial part of our work at Merantix. We are constantly improving experimentation cycles to be able to run more and more experiments.
Initiator and Member of
Research Member of SECREDAS Project
Merantix was selected by the European Commission to take part in a Horizon 2020 project on Cyber Security for Cross Domain Reliable Dependable Automated Systems (SECREDAS). Within SECREDAS Meratnix is a core contributor developing robust and secure image segmentation algorithms, that function reliably in corner-case situations as well as when faced with real-world adversarial attacks. Horizon 2020 is the financial instrument implementing the Innovation Union, a Europe 2020 flagship initiative aimed at securing Europe's global competitiveness. Seen as a means to drive economic growth and create jobs, Horizon 2020 has the political backing of Europe’s leaders and the Members of the European Parliament. The goal is to ensure Europe produces world-class science, removes barriers to innovation and makes it easier for the public and private sectors to work together in delivering innovation.
VDA (German Association of the Automotive Industry)
Member of the VDA Lead Initiative for Safeguarding Autonomous Vehicles
Merantix is a member of the VDA Lead Initiative on safeguarding autonomous vehicles, which aims to build both technology and industry standards to enable the scaled development and deployment of autonomous systems. The German federal government of has set the goal of ensuring Germany’s pioneering role in self-driving technology. Working closely with all major German OEMs and Tier 1 suppliers, Merantix is contributing to make detection and segmentation algorithms robust. Furthermore we are collaborating to develop semantics and standards for safe autonomous vehicles as well as concepts for scalable testing environments.
Recently in our paper discussion sessions
Recent Advances in Recurrent Neural Networks (2017)
Hojjat Salehinejad, Sharan Sankar, Joseph Barfett, Errol Colak, and Shahrokh Valaee
ReNet: A Recurrent Neural Network Based Alternative to Convolutional Networks (2015)
Francesco Visin, Kyle Kastner Kyunghyun Cho, Matteo Matteucci, Aaron Courville, Yoshua Bengio
World Models: Building generative neural network models of popular reinforcement learning environments (2018)
David Ha, Jurgen Schmidhuber
Continual Lifelong Learning with Neural Networks: A Review (2018)
German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, Stefan Wermter
End-to-end differentiable learning of protein structure (2018)