You are here


Industrial handwriting recognition solutions.  Since 2003, the MINDS group is enjoying a lively collaboration with PLANET intelligent systems GmbH, a northern Germany located SME that develops text recognition solutions for administrative, police, and postal customers. The collaboration is focussed on developing Echo State Networks algorithms for handwriting recognition, and has led to fielded systems for postal address recognition. PLANET regularly supports the MINDS team by stipends for PhD students and student projects.


Quadratic observable operator models.  Funded by the DFG (contract JA 1210/1-1&2, 2005-2009), this project developed a number of statistically and computationally efficient learning algorithms for observable operator models (OOMs), and established a theory of quadratic OOMs. These describe stochastic processes by linear operators, where probabilities are encoded by the squared norms of vectors, which creates an intriguing analogy to the formalism of quantum mechanics. The results were published in three articles in Neural Computation.


Observable operator networks. Funded by the DFG (contract JA 1210/5-1, 2009-2012), the objective of this project is to extend observable operator models (OOMs) to multivariate, continuous-valued time series, for example EEG recordings. The approach is to (i) introduce input-output OOMs (IO-OOMs), (ii) connect numerous such IO-OOMs into a causal dependency graph, where each node is one IO-OOM which receives input from its graph-parent IO-OOMs, (iii) identify from data the causal graph structure by applying the concept of Granger causality. A key to make this work is the computational efficiency of IO-OOM learning algorithms, which makes it possible to explore a very large number of candidate causal relationships.


ORGANIC (Self-Organized Recurrent Neural Learning for Language Processing) was a European Collaborative Project (IST-231267, 2009-2012) whose mission was to establish neurodynamical architectures as viable alternative to statistical methods for speech and handwriting recognition. Coordinated by the MINDS group at Jacobs University, it comprised six European research groups from reservoir computing, cognitive neuroscience, and speech and handwriting technology. The MINDS group contributed layered neural learning architectures based on Echo State Networks.


AMARSi (Adaptive Modular Architectures for Rich Motor Skills) was a European Collaborative Project (IST-248311, 2010-2014) which aims at a qualitative jump in robotic motor skills toward biological richness. Coordinated from the COR Lab at the University of Bielefeld, Amarsi joined 10 partners from robotics, neural computation and computational neuroscience, the motion sciences, and cognitive science. The MINDS group contributed models of neurocontrollers and led the architectures workpackage.

NeuRAM3 (Neural Computing Architectures in Advanced Monolithic 3D-VLSI Nano-Technologies) is a European H2020 Project (2016-2018) which develops a novel neuromorphic VLSI chip architecture and fabrication technology. Coordinated by the Commisariat à l’énergie atomique et aux énergies alternatives (CEA) and uniting 9 partners (among them gloablly leading chip manufacturers), the consortium aims especially at a low power consumption (reduction by a factor of 50 compared to current technology) and flexible configurability. In this project MINDS develops novel neural learning architectures on the basis of conceptors, for extreme robustness against noise, parameter drift, and coping with low numerical accuracy.