A5003202878- Domain Adaptation and Few-Shot Learning
- Terahertz technology and applications
- Adversarial Robustness in Machine Learning
- Machine Learning and Algorithms
- Neural Networks and Applications
- Human Pose and Action Recognition
- Explainable Artificial Intelligence (XAI)
- Multimodal Machine Learning Applications
- Semiconductor Quantum Structures and Devices
- Superconducting and THz Device Technology
- Generative Adversarial Networks and Image Synthesis
- Advanced Neural Network Applications
- Anomaly Detection Techniques and Applications
- Infrastructure Maintenance and Monitoring
- Radio Frequency Integrated Circuit Design
- Ethics and Social Impacts of AI
- Machine Learning and Data Classification
- Non-Destructive Testing Techniques
- Intelligent Tutoring Systems and Adaptive Learning
- Video Surveillance and Tracking Methods
- Geophysical Methods and Applications
- Machine Learning and ELM
- Sparse and Compressive Sensing Techniques
- Epigenetics and DNA Methylation
- Autonomous Vehicle Technology and Safety
Technical University of Darmstadt
2021-2024
Goethe University Frankfurt
2012-2023
Hessian Center for Artificial Intelligence
2022
Goethe Institute
2019-2021
Flint Institute Of Arts
2019
Goethe Institut
2019
Frankfurt Institute for Advanced Studies
2017
SKF (Germany)
1975
Current deep learning methods are regarded as favorable if they empirically perform well on dedicated test sets. This mentality is seamlessly reflected in the resurfacing area of continual learning, where consecutively arriving data investigated. The core challenge framed protecting previously acquired representations from being catastrophically forgotten. However, comparison individual nevertheless performed isolation real world by monitoring accumulated benchmark set performance. closed...
This paper reports on field-effect-transistor-based terahertz detectors for the operation at discrete frequencies spanning from 0.2 to 4.3 THz. They are implemented using a 150-nm CMOS process technology, employ self-mixing in n-channels of transistors and operate well above transistors' cutoff frequency. The theoretical description device by Dyakonov Shur is extended order describe impedance, responsivity, noise-equivalent power novel detection concept, which couples signal drain. approach...
Recognition of defects in concrete infrastructure, especially bridges, is a costly and time consuming crucial first step the assessment structural integrity. Large variation appearance material, changing illumination weather conditions, variety possible surface markings as well possibility for different types to overlap, make it challenging real-world task. In this work we introduce novel COncrete DEfect BRidge IMage dataset (CODEBRIM) multi-target classification five commonly appearing...
Learning continually from non-stationary data streams is a long-standing goal and challenging problem in machine learning. Recently, we have witnessed renewed fast-growing interest continual learning, especially within the deep learning community. However, algorithmic solutions are often difficult to re-implement, evaluate port across different settings, where even results on standard benchmarks hard reproduce. In this work, propose Avalanche, an open-source end-to-end library for research...
We present Queer in AI as a case study for community-led participatory design AI. examine how and intersectional tenets started shaped this community's programs over the years. discuss different challenges that emerged process, look at ways organization has fallen short of operationalizing principles, then assess organization's impact. provides important lessons insights practitioners theorists methods broadly through its rejection hierarchy favor decentralization, success building aid by...
We demonstrate for the first time applicability of antenna-coupled field-effect transistors detection terahertz radiation (TeraFETs) multi-spectral imaging from 0.76 to 4.25 THz. TeraFETs were fabricated in a commercial 90-nm CMOS process and noise-equivalent powers 59, 20, 63, 85 110 pW/√(Hz) at 0.216, 0.59, 2,52, 3.11 THz, respectively, have been achieved. A set has applied raster-scan transmission reflection pellets sucrose tartaric acid simulating common plastic explosives. Transmittance...
Modern deep neural networks are well known to be brittle in the face of unknown data instances and recognition latter remains a challenge. Although it is inevitable for continual-learning systems encounter such unseen concepts, corresponding literature appears nonetheless focus primarily on alleviating catastrophic interference with learned representations. In this work, we introduce probabilistic approach that connects these perspectives based variational inference single autoencoder model....
We present experimental and theoretical results on subharmonic mixing in field-effect transistors high above the transistor cutoff frequencies <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$f_{T}$</tex></formula> . Analytical expressions for heterodyne are derived considering different coupling conditions. They have to ensure that charge density oscillations excited transistors' channels by signals...
In a widely popular analogy by Turing Award Laureate Yann LeCun, machine intelligence has been compared to cake - where unsupervised learning forms the base, supervised adds icing, and reinforcement is cherry on top. We expand this 'cake that intelligence' from simple structural metaphor full life-cycle of AI systems, extending it sourcing ingredients (data), conception recipes (instructions), baking process (training), tasting selling (evaluation distribution). Leveraging our...
Continual learning (CL) is the sub-field of machine concerned with accumulating knowledge in dynamic environments. So far, CL research has mainly focused on incremental classification tasks, where models learn to classify new categories while retaining previously learned ones. Here, we argue that maintaining such a focus limits both theoretical development and practical applicability methods. Through detailed analysis concrete examples - including multi-target classification, robotics...
We present an analysis of predictive uncertainty based out-of-distribution detection for different approaches to estimate various models' epistemic and contrast it with extreme value theory open set recognition. While the former alone does not seem be enough overcome this challenge, we demonstrate that goes hand in latter method. This seems particularly reflected a generative model approach, where show posterior recognition outperforms discriminative models outlier rejection, raising...
Current deep learning methods are regarded as favorable if they empirically perform well on dedicated test sets. This mentality is seamlessly reflected in the resurfacing area of continual learning, where consecutively arriving data investigated. The core challenge framed protecting previously acquired representations from being catastrophically forgotten. However, comparison individual nevertheless performed isolation real world by monitoring accumulated benchmark set performance. closed...
Continual learning is a subfield of machine learning, which aims to allow models continuously learn on new data, by accumulating knowledge without forgetting what was learned in the past. In this work, we take step back, and ask: "Why should one care about continual first place?". We set stage examining recent papers published at four major conferences, show that memory-constrained settings dominate field. Then, discuss five open problems even though they might seem unrelated sight, will...
What is the state of art in continual machine learning? Although a natural question for predominant static benchmarks, notion to train systems lifelong manner entails plethora additional challenges with respect set-up and evaluation. The latter have recently sparked growing amount critiques on prominent algorithm-centric perspectives evaluation protocols being too narrow, resulting several attempts at constructing guidelines favor specific desiderata or arguing against validity prevalent...
We developed an end-to-end pipeline for brake light transition detection based on cognitive theories of anomaly-detection and model systems engineering principles. Inspired by theory, we decompose the visual input stream into submodalities color, shape, intensity motion which is closely coupled with a graphical model. A memory module that contains priors populated exploitation knowledge from specifications simulation. High-fidelity 3D-Simulations have been created to populate memory, whereas...
We report on an order-of-magnitude enhancement of sensitivity CMOS-transistor-based THz detectors. At 2.54 THz, 3.13 and 4.25 responsivity values 336 V/W, 308 230 V/W optimum noise-equivalent-power 63 pW/√Hz, 85 110 pW/√Hz are obtained.
The quest to improve scalar performance numbers on predetermined benchmarks seems be deeply engraved in deep learning. However, the real world is seldom carefully curated and applications are limited excelling test sets. A practical system generally required recognize novel concepts, refrain from actively including uninformative data, retain previously acquired knowledge throughout its lifetime. Despite these key elements being rigorously researched individually, study of their conjunction,...
A dataset is confounded if it most easily solved via a spurious correlation which fails to generalize new data. We will show that, in continual learning setting where confounders may vary time across tasks, the resulting challenge far exceeds standard forgetting problem normally considered. In particular, we derive mathematically effect of such on space valid joint solutions sets tasks. Interestingly, our theory predicts that for many datasets, correlations are ignored when tasks trained...
Identification of cracks is essential to assess the structural integrity concrete infrastructure. However, robust crack segmentation remains a challenging task for computer vision systems due diverse appearance surfaces, variable lighting and weather conditions, overlapping different defects. In particular recent data-driven methods struggle with limited availability data, fine-grained time-consuming nature annotation, face subsequent difficulty in generalizing out-of-distribution samples....
Medical image distributions shift constantly due to changes in patient population and discrepancies acquisition. These distribution result performance deterioration; deterioration that continual learning aims alleviate. However, only adaptation with data rehearsal strategies yields practically desirable for medical segmentation. Such violates privacy and, as most approaches, overlooks unexpected from out-of-distribution instances. To transcend both of these challenges, we introduce a...