The main focus of this position is to develop novel AI systems and methods for robot applications: Dexterous robot grasping, Human-robot learning, Transfer learning – efficient online learning. The role offers close cooperation with other institutes, universities, and numerous industrial partners, a self-determined development environment for own research topics with active support for the doctorate research project, flexible working hours, and work in a young, interdisciplinary research team.

3 PhD positions on « Physics Aware Deep Learning » are available at Sorbonne University, Paris, Fr. The positions are: 1) Machine Learning augmented model design of unsteady reactive flows – application to a scramjet combustor, 2) Domain generalization and transfer learning for deep learning data coherence of climate data sets, 3) Physics Based Deep Learning of Surrogate Models for Fluid Flow Simulation. Application to sustainable space missions.

The Post Doc for Sustainable, Reliable and Robust AI throughout the Life Cycle in the Context of Condition Monitoring will be part of an interdisciplinary team in pre-development/corporate research and develop AI-supported solutions for condition monitoring of machines and processes. The aim is to ensure that these solutions are robust, practical and suitable for long-term use. Using methods such as transfer learning, domain adaptation, drift detection and (data and knowledge-driven) hybrid modelling, the postdoc will design and implement workflows for machine learning. The postdoc will also work on real machines and processes to collect data and validate the developed procedures.

The focus of the PhD position in the Autonomous Intelligent Systems group at the University of Münster will be on Deep Reinforcement Learning for the control of locomotion in robots. The aim is to develop biologically-inspired principles that enable more efficient learning mechanisms for adaptive behaviour. The position will focus on model-free and model-based learning approaches for the control of robots based on biological principles such as decentralization, modularization, and hierarchical organization. The architecture will be applied to robots, e.g., the Unitree Go1, in multiple and increasingly more difficult tasks that require transfer learning. The candidate will also have the opportunity to extend this work towards a multi-agent setting or XAI to make decision-making more transparent. The position is tied to working towards a doctorate.

Flexible computation is a hallmark of intelligent behavior. Yet, little is known about how neural networks contextually reconfigure for different computations. Humans are able to perform a new task without extensive training, presumably through the composition of elementary processes that were previously learned. Cognitive scientists have long hypothesized the possibility of a compositional neural code, where complex neural computations are made up of constituent components; however, the neural substrate underlying this structure remains elusive in biological and artificial neural networks. Here we identified an algorithmic neural substrate for compositional computation through the study of multitasking artificial recurrent neural networks. Dynamical systems analyses of networks revealed learned computational strategies that mirrored the modular subtask structure of the task-set used for training. Dynamical motifs such as attractors, decision boundaries and rotations were reused across different task computations. For example, tasks that required memory of a continuous circular variable repurposed the same ring attractor. We show that dynamical motifs are implemented by clusters of units and are reused across different contexts, allowing for flexibility and generalization of previously learned computation. Lesioning these clusters resulted in modular effects on network performance: a lesion that destroyed one dynamical motif only minimally perturbed the structure of other dynamical motifs. Finally, modular dynamical motifs could be reconfigured for fast transfer learning. After slow initial learning of dynamical motifs, a subsequent faster stage of learning reconfigured motifs to perform novel tasks. This work contributes to a more fundamental understanding of compositional computation underlying flexible general intelligence in neural systems. We present a conceptual framework that establishes dynamical motifs as a fundamental unit of computation, intermediate between the neuron and the network. As more whole brain imaging studies record neural activity from multiple specialized systems simultaneously, the framework of dynamical motifs will guide questions about specialization and generalization across brain regions.

My group pioneered the development of a novel intracortical brain computer interface (iBCI) that decodes muscle activity (EMG) from signals recorded in the motor cortex of animals. We use these synthetic EMG signals to control Functional Electrical Stimulation (FES), which causes the muscles to contract and thereby restores rudimentary voluntary control of the paralyzed limb. In the past few years, there has been much interest in the fact that information from the millions of neurons active during movement can be reduced to a small number of “latent” signals in a low-dimensional manifold computed from the multiple neuron recordings. These signals can be used to provide a stable prediction of the animal’s behavior over many months-long periods, and they may also provide the means to implement methods of transfer learning across individuals, an application that could be of particular importance for paralyzed human users. We have begun to examine the representation within this latent space, of a broad range of behaviors, including well-learned, stereotyped movements in the lab, and more natural movements in the animal’s home cage, meant to better represent a person’s daily activities. We intend to develop an FES-based iBCI that will restore voluntary movement across a broad range of motor tasks without need for intermittent recalibration. However, the nonlinearities and context dependence within this low-dimensional manifold present significant challenges.

Social learning and analogical reasoning both provide exponential opportunities for learning. These skills have largely been studied independently, but my future research asks how combining skills across previously independent domains could add up to more than the sum of their parts. Analogical reasoning allows individuals to transfer learning between contexts and opens up infinite opportunities for innovation and knowledge creation. Its origins and development, so far, have largely been studied in purely cognitive domains. Constraining analogical development to non-social domains may mistakenly lead researchers to overlook its early roots and limit ideas about its potential scope. Building a bridge between social learning and analogy could facilitate identification of the origins of analogical reasoning and broaden its far-reaching potential. In this talk, I propose that the early emergence of social learning, its saliency, and its meaningful context for young children provides a springboard for learning. In addition to providing a strong foundation for early analogical reasoning, the social domain provides an avenue for scaling up analogies in order to learn to learn from others via increasingly complex and broad routes.