THIS EVENT HAD TO BE CANCELLED AT SHORT NOTICE, WILL BE RESCHEDULED.

A get-to-know INI colloquium on the topic of large language models with interesting talks by:

• Jun.-Prof. Dr. Tatjana Scheffler -- Chair of Digital Forensic Linguistics, Ruhr University Bochum

• Prof. Dr. Bilal Zafar -- Chair of Computing and Society, Ruhr University Bochum

• Prof. Dr. Meeyoung Cha -- Scientific Director of Max-Planck Institute for Security and Privacy, Bochum

Followed by a panel discussion and networking coffee.

If you plan to attend, please register here.

Abstract:

In the past two decades, the shift towards a distributed computing paradigm led our smart systems
to become more and more interconnected. These systems need to elaborate increasingly amount of
data while featuring low-power operation, area efficiency, and ability to interact with the external
world in real time. Memristive technology, with its unique characteristics and capabilities, holds great
promise for the design of such cognitive systems. The potential for energy-efficient and parallel
computing, combined with the ability to integrate complex neural and synaptic dynamics within a
single device, provides avenues for high-performance hardware implementations. Moreover, by
offering volatile and non-volatile memory in a small footprint, enabling dense integration, and
facilitating in-memory computing, memristive technology presents advantages that, if correctly
combined with CMOS technology, can extend the functionality of current artificial intelligent systems.
In this talk, we discuss the challenges and the opportunities to realise memristive neuromorphic
computing by developing novel hardware architectures and learning algorithms specifically tailored
to best exploit the intrinsic properties of memristive technology. Indeed, we show that memristive
technology offers vast potential, but its effective utilization relies on the synergetic development of
memristive devices, circuits, and algorithms to create performing hardware cognitive systems.

Bio:

Dr. Erika Covi is Assistant Professor at the Zernike Institute for Advanced Materials and the
Groningen Cognitive Systems and Materials Centre, University of Groningen (the Netherlands).
She received her PhD in Microelectronics in 2014 from the University of Pavia (Italy), where she
worked on designing integrated systems for the characterisation of memristive devices. She worked
at the National Research Council (CNR) of Italy, at Politecnico di Milano (Italy), and as Senior
Scientist at NaMLab gGmbH, Dresden (Germany), where she wass the leader of the Cognitive
Systems group. She was awarded with an ERC Starting Grant for the project MEMRINESS on the
development of memristive neurons and synapses for neuromorphic edge computing in 2021.
Her research interests lie at the intersection of emerging devices, circuit design, and brain-inspired
computing. More specifically, they focus on exploiting the intrinsic physical characteristics of
memristive devices to reproduce computational primitives of the brain in mixed neuromorphic-
memristive systems.

Ms. Ivonne Lee is the new psychosocial counselor at Ruhr-University Bochum, having taken up her position in December 2023. With a strong focus on prevention, Ms. Lee provides valuable support in cases of workplace conflicts, addiction risk and counseling, psychological stress, and life crises. She emphasizes the importance of mental health for employees, noting that since we spend so much time at work, it is crucial to create a supportive environment. Her approach is guided by the principle of "love it, change it, or leave it," highlighting the need to identify which factors we can influence and which we cannot. She also conducts seminars on relaxation, mindfulness, changing perspectives, and conflict management. Ms. Lee is dedicated to fostering interpersonal connections and is eager to engage with employees personally.

Abstract:

We present a proximal policy optimization (PPO) agent trained through curriculum learning (CL) principles and meticulous reward engineering to optimize a real-world high-throughput waste sorting facility. Our work addresses the challenge of effectively balancing the competing objectives of operational safety, volume optimization, and minimizing resource usage. A vanilla agent trained from scratch on these multiple criteria fails to solve the problem due to its inherent complexities. This problem is particularly difficult due to the environment's extremely delayed rewards with long time horizons and class (or action) imbalance, with important actions being infrequent in the optimal policy. This forces the agent to anticipate long-term action consequences and prioritize rare but rewarding behaviours, creating a non-trivial reinforcement learning task. Our five-stage CL approach tackles these challenges by gradually increasing the complexity of the environmental dynamics during policy transfer while simultaneously refining the reward mechanism. This iterative and adaptable process enables the agent to learn a desired optimal policy. Results demonstrate that our approach significantly improves inference-time safety, achieving near-zero safety violations in addition to enhancing waste sorting plant efficiency.

This is a trial balloon for a new format of the INI seminar, in which some of the INI professors lead a discussion about a broad topic of interest to all INI members. Each professor will give a brief statement about the question raised, that is expected to trigger a discussion among the professors with the active participation of the audience. 

Abstract:

In sequential decision-making problems, identifying a subset of alternatives
that guarantee the best outcome is challenging, especially when there is a
structural dependency among the rewards associated with each action. Besides
the individual action's reward, it is necessary to learn the causal
relationships to improve the decision-making strategy. To solve this two-fold
learning problem, we have developed the 'combinatorial semi-bandit framework
with causally related rewards'. In this framework, we model the causal
relationships among actions as a directed graph in a stationary structural
equation model. To achieve this objective, we propose a policy that learns the
network's topology to determine the causal relationships and simultaneously
uses this knowledge to optimize the decision-making process.
In this talk, I start with a brief explanation of sequential decision-making.
After that, I describe our framework, explain its applications, and present
our solution together with the results of the analysis. I conclude the talk by
discussing the extensions of our framework into non-stationary settings and
delayed feedback.

Abstract:

In this talk, we argue that any AI decision-making system must consider the uncertainty it may face in the real world. In particular, an AI system must account for, be robust against, or reduce uncertainty. We highlight our general vision of foundational and application-driven research aiming for safe and reliable AI. We present a number of research highlights that all follow our general scheme of neurosymblic AI: Tightly integrated learning and verification techniques that essentially (1) learn a model of a system, (2) verify the model, (3) use a rigorous verification result to improve the learning.

Look here for background information. 

Abstract:

Memories are typically studied only at encoding and retrieval. However, their life between these two bookends is essential, as it determines whether memories will be forgotten, remembered, or become disruptive and intrusive. In this presentation, I will share the findings of a multi-year project in which we applied a formal model of memory to infer and describe the trajectory of memories before they are accessed. The central parameter of the model, which determines the speed of forgetting, is remarkably stable within individuals and can be decoded by analyzing features of their resting-state brain activity. Moreover, the model can be successfully applied to clinical populations, successfully identifying individuals with amnestic cognitive impairment and accounting for changes in the hippocampus size in individuals suffering from PTSD.

Join Zoom Meeting:
https://ruhr-uni-bochum.zoom.us/j/64423067014?pwd=VXA3MncxY3pyTURXLzNrcEtZSm9kZz09
Meeting ID: 644 2306 7014
Passcode: 246479

As DeepMind has revolutionized the AI side of planning in combinatorially large problems, our lack of understanding of how humans plan in such situations has come into stark focus. The cognitive science of chess, once promising, is now virtually extinct. Planning tasks widely used in the field nowadays don’t require much thinking ahead. I will show that it is possible to study human multi-step planning in tasks of intermediate complexity while maintaining experimental tractability and computational modelability. I will describe a series of experiments on a game that we call four-in-a-row -- a variant of tic-tac-toe or Go Moku (五子棋). Inspired by best-first search, we built a heuristic computational model of human play in this game and fitted it to move-level data. The model predicts moves in unseen positions, decisions in unseen tasks, eye fixation patterns, mouse movements, and response times. Moreover, the model allows us to computationally characterize the effects of expertise and time pressure. Linking back to the chess literature, I will discuss how experts differ from novices in remembering game positions and move sequences. Finally, I will describe parallel results from a very large online data set, a project on the development of planning, and ongoing projects on the neural basis of multi-step planning.

This Colloquium will take place on Zoom.

Link: https://ruhr-uni-bochum.zoom.us/j/68859099254?pwd=Ukgxc0gzZDB5MFhsQkJob2VveEVjUT09
Meeting ID: 688 5909 9254
Passcode: 057211

Abstract:

Flexibly storing and updating memory content is a fundamental ability of animals in which prefrontal cortex (PFC) is believed to play a major role as its activity can reflect information that must be temporarily maintained to realize the current goal. However, in our flexible spatial working memory task, where rats were required to navigate from different starting points and via multiple routes to one of multiple possible goal locations, we found no evidence that the well-controlled delay period contained current-goal-specific memory. This suggests that under conditions where memory needs to be employed flexibly, alternative storage mechanisms exist, or other brain areas serve to store this information. PFC activity instead categorized important spatial locations according to their meaning in the task and displayed neural representations that reflected the geometry of the maze. Furthermore, we found that elapsed time was encoded at starts and goals and that this time code was invariant to different start and goal locations. Despite the tendency of individual neurons to show mixed selectivity, i.e. to be selective for multiple task-relevant variables, we found that subsets of neurons had functional preferences for time or space. This structured selectivity may facilitate complex behaviors by efficiently generating informative representations of multiple variables.

An introductory talk (In Person/Hybrid) by Prof. Dr. Robert Schmidt at the start of new year 2023. This would be one hour slot, as opposed to the regular 30 min INI talks. 

Abstract:

The activity of a population of cells encodes instantaneous sensory, cognitive, and motor aspects of animal behavior. On the other hand, rules governing the behavior are likely encoded in the communication between cells/circuits. Addressing brain interactions is, however, hampered by recurrent connectivity (loops) and common inputs. For example, recurrent circuits reverberate previous activity states causing long-term correlations that undermine any mathematical scrutiny of underlying interactions. To quantify communication principles regardless of spurious correlations, I have developed a theoretical and experimental approach for estimating an "impact function" between populations of neurons. To realize this approach, I temporally fuse ultra-fast optogenetic suppression and simultaneous large-scale extracellular recordings in awake animals. In this presentation, I will describe the foundations of the approach followed by results from cell-type specific recurrent circuits in the primary visual cortex of the mouse

I will talk about why it is importart to gain insight into how AI models form their decisions. I will then present some of the most popular methods that are being used to make ML models more transparent and comprehensible.

Abstract:

The ability of the brain to represent the external world in real-time is impacted by the fact that neural processing takes time. Because neural delays accumulate as information progresses through the visual system, representations encoded at each hierarchical level are based upon input that is progressively outdated with respect to the external world. This is particularly relevant to the task of localizing a moving object – because the object’s location changes with time, neural representations of its location potentially lag behind its true location. It has therefore been proposed that the visual system utilizes the predictive nature of motion to extrapolate moving objects along their trajectory. Burkitt and Hogendoorn (2021, https://doi.org/10.1523/JNEUROSCI.2017-20.2021) showed how spike-timing dependent plasticity (STDP) can achieve motion extrapolation in a two-layer, feedforward network of velocity-tuned neurons, by shifting the receptive-fields of second-layer neurons in the opposite direction to a moving stimulus.

The current study extends this work by implementing two changes to the network to bring it more into line with biology: expansion of the network to multiple higher layers to better reflect the depth of the visual hierarchy, and introduction of the time constants associated with biological neural processing. We examine the degree to which STDP can facilitate compensation of neural delays across six layers, showing that the relative contribution of the highest layers increases at higher velocities. We also explore the effect of additional delays imposed on the network by the integration time of the membrane potential.

This is a first instance of a new category of INI Colloquia in which INI group leaders will present aspects of the research program of their teams. This will extend the INI Seminar presentations of group leaders to a more in depth presentation that will take up to an hour. In this first instance, I will give an overview over our research program in the motor domain by taking you through the neural dynamic architecture of movement generation that links perception and cognition to motor control. 

Nowadays more and more data of moving entities such as cars, animals, and people is being collected. These data can be analysed for example to detect movement patterns, similarity, or type of movement. I will discuss geometric algorithms for doing so, in particular for computing similarity and segmentation.

Prof. Maike Buchin is our colleague in the Faculty of Computer Science at RUB. Her colloquium is part of a series of talks aimed at getting to know each other within our new Faculty

Michael Kamp heads a new Junior Research Group funded by the "Institut für künstliche Intelligenz" of the Essen University Clinic. He will be housed at the INI and we are eager to learn more about his research. This Colloquium gives an introduction into his line of work. See here for more information. 

The cell assembly hypothesis postulates that neurons coordinate their activity through the formation and repetitive co-activation of groups. We assume that assembly activity is expressed by the occurrence of precisely timed spatio-temporal patterns (STPs) of spikes emitted by neurons that are members of the assembly. We developed a method that is capable to detect significant STPs in massively parallel spike trains (SPADE, Torre et al, 2016; Stella et al, 2019). SPADE identifies and counts repeating patterns using Frequent Itemset Mining. Then the detected patterns are evaluated for their significance through comparison to patterns found in surrogate data that contain the same changing firing rates of the neurons. Various surrogate techniques can be used to evaluate significance, and their correct choice is crucial to ensure that by-chance patterns are not classified as significant (Stella et al, under revision). Using SPADE we evaluated parallel spike data recorded (by a 100 electrode Utah array) in pre-/motor cortex of monkeys performing a reach- to-grasp task (Riehle et al, 2013). We indeed find significant STPs specific to behavioral context. Typically STPs are formed by 3-6 neurons, and have a maximal temporal extend of 60ms. Neurons involved in STPs are not clustered on the cortex, but may be far apart (up to 3.6mm). Often we find neurons that are involved in multiple STPs. Currently, we explore if a network model based on synfire chains (SFCs) is capable of explaining the patterns we find in experimental data.

Cognition is the result of computations over representations in the brain to yield behavior. But what are the parts of the brain that perform those computations, and how do they relate to cognition? In this talk, I will outline two different views on those parts and their relationship to thought. The Sherringtonian view describes networks of neurons that individually perform computations for cognition. On this view, the representations are the spikes sent between neurons and the computations are the ways that individual neurons transform those signals. In contrast to the Sherringtonian's emphasis on single neurons and similar biophysical details, the Hopfieldian view describes dynamical objects like state spaces, manifolds, and trajectories for cognition. On this view, the representations are attractors in neural state spaces and the computations are the trajectories through or transitions between them. I will illustrate both views, discuss their relationship to levels of explanation and levels of analysis, and argue that the Hopfieldian view has greater promise for explaining the most challenging aspects of cognition.

Software engineering is at an inflection point. The traditional and largely manual construction of software is challenged by novel trends and technologies -- including rapidly changing markets, artificial intelligence or continuous delivery (e.g., in-field updates of cars). Developers need to frequently experiment with new ideas and quickly customize software for different markets, environments or hardware platforms, while optimizing for non-functional properties, such as security, performance or energy consumption.

In this talk, I will present my group’s contributions towards automating the engineering of software systems. I will discuss how we combine foundational and applied research, often in collaboration with industry, to advance the scientific and engineering principles of software engineering. My focus will be on two application domains: variant-rich systems and robotics control systems.

Thorsten Berger is a Professor in Computer Science at Ruhr University Bochum in Germany. His research focuses on automating software engineering for the next generation of intelligent, autonomous, and variant-rich software systems -- exploring new ways of software creation, analysis, and evolution. Thorsten Berger received the PhD degree in computer science from the University of Leipzig in Germany in 2013, worked as a Postdoctoral Fellow at the University of Waterloo in Canada and the IT University of Copenhagen in Denmark, and then an Associate Professor jointly at Chalmers University of Technology and the University of Gothenburg in Sweden.

This talk is based on my latest publication. Find the paper's abstract below:

Video-based traffic sign recognition is a key ability of autonomous vehicles but a demanding challenge due to the enormous number of classes and natural conditions in the
wild. We address this problem with a fully automatic close-to-life image-to-image translation technique for traffic sign substitution in natural images. The work is intended
as data augmentation technique and allows for training rare or unavailable traffic sign classes, or otherwise uncommon cases in visual traffic sign detection and classification. To this end, we extend our previous data generation model and propose a rendering pipeline to create convincing traffic sign images with realistic background and camera recording artifacts.
Experiments are conducted by exchanging traffic sign classes on different parts of the German Traffic Sign Recognition Benchmark (GTSRB). We demonstrate that the pipeline
is well-suited for generating representative images of unseen traffic sign classes. A baseline image classification setup trained on real data shows an overall performance similar to being trained with a comparable number of artificial data samples.

In this talk I will discuss a recent reddit discussion started by Bengio on a (in my opinion) old problem.

I am going to briefly summarize a new research program that investigates the relationship between dynamic field theory, spiking neurons, and neuromorphic hardware. I will also show some (very) preliminary simulation results.

A first attemt to interact in a video conference format about a scientific topic relevant to many at the INI. A short starter presentation will set the topic, followed by ample time for interaction. 

Descending motor control signals in humans may not directly specify any biomechanical movement variables, such as muscle activation, movement direction or speed. Rather, these signals may control movement indirectly, through intermediate neurophysiological parameters. One such parameter for movement control could emerge at the spinal level of motoneuron membrane potential. I will illustrate the physical and physiological principles underlying such parametric control and present experimental results on the time-course modulation of this neurophysiological parameter in human reaching movement.

The talk is more neuro-oriented than some other work by Christian Igel, has animations
and no equation, and is based on:

- Mathias Perslev, Michael Hejselbak Jensen, Sune Darkner, Poul Jørgen Jennum, and Christian Igel. U-Time: A Fully Convolutional Network for Time Series Segmentation Applied to Sleep Staging. Advances in Neural
Information Processing Systems (NeurIPS 2019), accepted
- Mathias Perslev, Erik Dam, Akshay Pai, and Christian Igel. One Network To Segment Them All: A General, Lightweight System for Accurate 3D Medical Image Segmentation. In: Medical Image Computing and Computer
Assisted Intervention (MICCAI), LNCS 11765, pp. 30-38, Springer, 2019
- Thorbjørn Louring Koch, Mathias Perslev, Christian Igel, and Sami Sebastian Brandt. Accurate Segmentation of Dental Panoramic Radiographs with U-Nets. In: IEEE International Symposium on Biomedical Imaging
(ISBI), pp. 15-19, IEEE Press, 2019

 it is needed for the Sicherheitsbegehung on October 28

Prof. Spencer gives a presentation on how he uses ideas from DFT to analyze neural imaging data in the DFT summer school. The talk is open to all. 

This colloquium talk by Dietmar Heinke is given in the context of the DFT Summer School but is open to all INI members. 

Temporal-difference learning is widely adopted and has been successfully applied to a number of problems as well as used to model animal learning. However, they are mainly based on neural pathways involved in reward-seeking behaviour since little is known about punishment-driven
learning and less still about the combined effects of both types of reinforcement on learning. In this talk, I will discuss the implications for machine learning and robotics of this incomplete model of reward-based learning and possible strategies to deal with it.

I will motivate and explain the math and the intuition of low-rank matrix approximation, which is a general technique that appears frequently as a building block in machine learning and optimization algorithms.

I will review the free energy principle proposed by Karl Friston and colleagues and discuss its influence on research in neuroscience and behavior. 

This hands-on tutorial will give you a chance to try out concepts of neural dynamics in our software CEDAR.

We will host the tutorial in the robotics lab (NB 02/77). There is no need to bring your laptop as you will be able to work on our computers.

The tutorial will take about an hour.

(This is also still subject to change as we plan everything.)

I will give a short introduction to dynamic field theory (DFT) to lay the foundation for the hand-on DFT tutorial that Jan Tekülve and I will be offering after lunch.

Prof. Suddendorf will give a talk in a joint Colloquium session of the Institute für Neuroinformatics and the Center for Mind and Cognition.

Abstract:

"The human ability to travel mentally in time and to consider diverse future possibilities has increasingly become a topic of considerable research attention. Here I will review recent studies from our laboratory, examining the nature and development of foresight, from the essential capacity to conceive of alternatives, to strategic applications, such as deliberate practice aimed at shaping future skills. I maintain that complex prospection has been a prime mover in human evolution and is a key to understanding human dominance on the planet"

The central RUB office for worker safety will inform us about the current status of exposure to asbestos in our rooms. Dr. Wiegand will give a presentation and answer questions.

Denkstunde für Rolf Würtz

Wir werden PD Dr. Rolf Würtz am 19. Dezember 2018 ab 12 Uhr 30 bis ca. 14 Uhr in einer kleinen Feier gedenken. Die Gedenkstunde findet im Veranstaltungszentrum der Ruhr-Universität statt, also im Gebäude der Mensa der RUB auf der Ebene 04, Saal 1. 

Wir haben eine Reihe von Freunden und Kollegen von Rolf gebeten, in kurzen Ansprachen, verschiedene Aspekte von Rolfs Leben und Wirken zu beleuchten. Einige kurze musikalische Einlagen begleiten das Programm. Im Anschluss gibt es ein paar Happen zu Essen und Gelegenheit zum Austausch unter den Gästen.

Abilities and skills in performing dynamic manipulations by humans appear in course of life-long training and experimentations. As a result, most of humans can readily manipulate or learn quickly how to manipulate external objects and environments without a firm grip or without re-grasping for re-orientating or moving. Here one can think of grasping and moving a wet soap when cleaning hands, of controlled rolling basketball ball on a palm in throwing it to the goal, think of manipulating knife in cutting soft materials etc. We can learn how to handle various objects or media for the purpose even them can slide or roll being in contact.

The talk is focused on challenging problems in modern robotics with emphasis of the discussion shifted to model-based approaches for developing some of human-like functionalities for robots. The intentions are examined by developing a solution for performing a rolling of passive objects (disc) on a robotic hand. Necessary mathematical concepts and arguments for the task are presented. Importance, relevance and scalability of the reasoning are supported by successful experimental studies on the robot.

The cognitivist approach to cognition that emerged in the mid-1950's exploited the newly invented computer as a literal metaphor for cognitive functions and operation, using symbolic information processing as its core model for cognition. In spite of the later appearance of the so-called "emergent" approaches such as connectionism and embodied dynamicism, the main tenet of cognitivism, namely that features of the environment are represented by internal states within a cognitive agent, is still deeply entrenched in our mindset. Such pervasive "representationism" forces us to interpret the neural data in a certain way. For example, any correlation between internal (e.g. neural activity) and external (e.g. features of the environment) states are automatically interpreted as a "code". In this talk, I will try to use simple metaphors to stir the debate on the issue of representation.

The medial temporal lobe (MTL) is well-known to be essential for declarative memory processing. According to the traditional view, the MTL exclusively subserves mnemonic processes. However, in the last two decades results from a growing body of studies have suggested that the MTL may also play a critical role in high-level perception (perceptual-mnemonic hypothesis). I will briefly discuss those studies and then present a computational model providing an alternative explanation for the putative role of the MTL in perception.

Prof. Dr. Katharina Henke & Dr. Simon Ruch (University Bern) will give two talks in a joint Colloquium of the INI and of the SFB 874.

Titles: Conscious versus unconscious memory for episodes (K.Henke) & Implicit relational vocabulary encoding during sleep is bound to slow-wave peaks (S. Ruch)

Time: Monday, July 16th, 4-6pm

Venue: NB 3/57

Host: Sen Cheng

Abstracts: 

K. Henke: Conscious versus unconscious memory for episodes

Episodic memory builds on the rapid encoding of flexible what-where-when associations. When we are aware of our episodic memories, they originate from consciously witnessed events. But episodic memories may also originate from an unconscious processing of subliminal or unattended events. I’ll report experiments on the supraliminal (conscious) and subliminal (unconscious) encoding of movies that put the idea of unconscious episodic encoding to a critical test and that examine the hypothesis of a larger memory capacity for unconscious versus conscious episodic memory. Information load was manipulated in 3 steps with 1, 3 or 9 consecutive movies presented for encoding. Retention of movies was assessed using a two-alternative forced-choice task with accuracy as explicit and reaction time as implicit measure of memory. This task records the retention of relational inferences participants made while watching a movie. Reaction times revealed successful long-term retention of inferences made on both subliminal and supraliminal movies. These results support the notion of an unconscious form of episodic memory. Accuracy caught memory of supraliminal movies alone. Retrieval performance remained robust with high load in the subliminal condition but collapsed in the supraliminal condition both in terms of accuracy and reaction times. We assume that unconscious versus conscious episodic encoding yielded sparse and segregated memory representations with little overlap, which may provide for less interference and a larger capacity. Just as the strength of conscious episodic memories comes with a limited capacity, so the weakness of unconscious episodic memories appears to come with a large capacity. The hippocampus supported both the conscious and unconscious encoding and retrieval of flexible what-where-when associations.

 

Abstract: 

I'll talk about two of my recent works on deep probabilistic models: Bayesian Hypernetworks and Neural Autoregressive Flows.  Bayesian Hypernetworks apply deep generative models (specifically normalizing flows) to represent an approximate posterior for another deep network.  Neural Autoregressive Flows generalize some previous works on normalizing flows, which are a flexible technique with numerous applications including generative modeling and variational inference, and are an essential component of Google's text-to-speech engine.

The high level motivation of this research is to investigate the potential of Bayesian deep learning as a tool for building safe, well-calibrated ML systems, and these works combined provide a variational inference method which is in principle able to represent the true posterior of a deep network, although optimization and scalability remain as practical challenges.  

I'll give a brief introduction to normalizing flows and Bayesian deep learning before describing our contributions.  Time permitting, I will also discuss some of the other open questions for this research agenda.

"A neural dynamic model for the perceptual grounding of spatial and movement relations"

The talk will be in German.

Abstract: 

This talk will cover two different areas of research on decoding non-invasive brain signals: (1) the use of big data neuroimaging and genetics resources for decoding brain disorders, and (2) the use of electroencephalographic recordings to control assistive technology devices by means of a Brain-Computer Interface.

Psychiatric disorders are among the main contributors to morbidity and disability around the world, yet in contrast to other fields of medicine, clinical decision making in psychiatry still today largely relies on symptom descriptions rather than biology. The talk will discuss how novel big data resources of neuroimaging and genetics data may be utilized toward biology-informed precision medicine in psychiatry and how this can lead to an increased understanding of processes underlying brain maturation and dysfunction.

Another field of research – Brain Computer Interfaces – attempts to decode electroencephalographic signals from the brain of individuals with severe motor paralysis for controlling assistive technology devices. The talk will describe developments of such systems as an alternative communication channel as well as for wheelchair control.

Each of the fields has progressively advanced over recent years, yet faces numerous challenges still. Can they learn from each other?

Time: Wednesday, June 27th, 12-1pm

Venue: NB 3/57

Hosts: Sen Cheng & Johannes Lederer (Fakultät für Mathematik)

Enlighten the brain - advances in optical imaging approaches

In this talk we analyze video signals and try to maximize micro movements in video signals to make them visible. These micro movement are typically undetectable and can’t be seen by human being. We utilize several frequency and transform domain tools to maximize difference in frame pixels and try to detect any minor change in frames and try to magnify this difference to yield video signals with movement maximization. This movement maximization is achieved by magnifying some specific frequency bands by a multiplication factor and reconstructing back the video signal after some manipulation and modification to make these micro movements seen and observable. These micro movements can later be utilized in different applications such as Medical Imaging, Structural Engineering, Mechanical Engineering, Physical Feature Analysis and Industrial Engineering. In our work we introduce our own model for micro movement magnification and compare it with very recent state of the art work (el. Freeman.) and show the tradeoffs.

Sensory neural circuits are thought to efficiently encode incoming signals. Several mathematical theories of neural coding formalize this notion, but it is unclear how they relate to each other and whether they are even fully consistent. I will present a unified framework, based on the ‘Information bottleneck’, that encompasses and extends previous proposals. With this framework, we can highlight key tradeoffs faced by sensory neurons. For example trading off future prediction versus efficiently encoding past inputs leads to qualitatively different predictions for neural responses to natural visual stimulation. Finally I will discuss how this approach could be used as a first step to theoretically explaining the observed diversity of neural responses. [see http://www.pnas.org/content/115/1/186.long, for related publication]. SFB 874

Vorstellung der Bachelorarbeit

Autonomous navigation is very active research area in robotics that is searching for alternative methods to provide intelligent ways to robots navigate in unknown environment, which defines the so called simultaneous localization and mapping (SLAM) problem. Outcomes provided by the neuroscience research indicates that hippocampus play a role in individuals’ space navigation when sequences of movements episodes of an individual can be used to construct a map of an unknown environment, which at the end provides a good source of inspiration to design SLAM solutions (here called Biomorphic SLAM). On the other side, neuroscience can profit from biologically plausible robotic solutions to test the hypothesis about how the biological structures really work. Therefore, this loop interleaving robotics and neuroscience advancements benefits both fields of knowledge. From the robotics’ perspective, the gain in using the Biomorphic SLAM becomes even more valuable when multi-robots systems are the target. Nowadays, Cloud-Robotics Systems are the state of the art in multi-robots systems, demanding solutions for a number of problems, in which the cooperative SLAM (C-SLAM) is a highlight. In C-SLAM, all robots perform SLAM in the same environment, but each of them just manages to cover a small portion of that environment. The goal of a C-SLAM solution is to provide a merge of their partial maps so that the system a whole manages to build a complete map. This talk discusses the application of the Biomorphic SLAM approach in the C-SLAM problem in the scope of Cloud-Robotics Systems. In this context, the state of the collaborative work between UFRGS and RUB will be presented.