Resource
2024 EN
Kerem Mertoğlu · Yusuf Şalk · Server Karahan Sarıkaya
+6 more
Creation of new annotated public datasets is crucial in helping advances in3D computer vision and machine learning meet their full potential for automaticinterpretation of 3D plant models. In this paper, we introduce PLANesT-3D; anew annotated dataset of 3D color point clouds of plants. PLANesT-3D iscomposed of 34 point cloud models representing 34 real plants from threedifferent plant species: \textit{Capsicum annuum}, \textit{Rosa kordana}, and\textit{Ribes rubrum}. Both semantic labels in terms of "leaf" and "stem", andorgan instance labels were manually annotated for the full point clouds. As anadditional contribution, SP-LSCnet, a novel semantic segmentation method thatis a combination of unsupervised superpoint extraction and a 3D point-baseddeep learning approach is introduced and evaluated on the new dataset. Twoexisting deep neural network architectures, PointNet++ and RoseSegNet were alsotested on the point clouds of PLANesT-3D for semantic segmentation.
Resource
2024 EN
Chengbo Zang · Mehmet Kerem Turkcan · Gil Zussman
+2 more
We present a novel data-driven simulation environment for modeling traffic inmetropolitan street intersections. Using real-world tracking data collectedover an extended period of time, we train trajectory forecasting models tolearn agent interactions and environmental constraints that are difficult tocapture conventionally. Trajectories of new agents are first coarsely generatedby sampling from the spatial and temporal generative distributions, thenrefined using state-of-the-art trajectory forecasting models. The simulationcan run either autonomously, or under explicit human control conditioned on thegenerative distributions. We present the experiments for a variety of modelconfigurations. Under an iterative prediction scheme, the way-point-supervisedTrajNet++ model obtained 0.36 Final Displacement Error (FDE) in 20 FPS on anNVIDIA A100 GPU.
Resource
2024 EN
Mehmet Kerem Turkcan · Yuyang Li · Chengbo Zang
+3 more
We introduce Boundless, a photo-realistic synthetic data generation systemfor enabling highly accurate object detection in dense urban streetscapes.Boundless can replace massive real-world data collection and manualground-truth object annotation (labeling) with an automated and configurableprocess. Boundless is based on the Unreal Engine 5 (UE5) City Sample projectwith improvements enabling accurate collection of 3D bounding boxes acrossdifferent lighting and scene variability conditions. We evaluate the performance of object detection models trained on the datasetgenerated by Boundless when used for inference on a real-world dataset acquiredfrom medium-altitude cameras. We compare the performance of theBoundless-trained model against the CARLA-trained model and observe animprovement of 7.8 mAP. The results we achieved support the premise thatsynthetic data generation is a credible methodology for training/fine-tuningscalable object detection models for urban scenes.
Resource
2024 EN
Kerem Cekmeceli · Meva Himmetoglu · Guney I. Tombak
+3 more
Neural networks achieve state-of-the-art performance in many supervisedlearning tasks when the training data distribution matches the test datadistribution. However, their performance drops significantly under domain(covariate) shift, a prevalent issue in medical image segmentation due tovarying acquisition settings across different scanner models and protocols.Recently, foundational models (FMs) trained on large datasets have gainedattention for their ability to be adapted for downstream tasks and achievestate-of-the-art performance with excellent generalization capabilities onnatural images. However, their effectiveness in medical image segmentationremains underexplored. In this paper, we investigate the domain generalizationperformance of various FMs, including DinoV2, SAM, MedSAM, and MAE, whenfine-tuned using various parameter-efficient fine-tuning (PEFT) techniques suchas Ladder and Rein (+LoRA) and decoder heads. We introduce a novel decode headarchitecture, HQHSAM, which simply integrates elements from twostate-of-the-art decoder heads, HSAM and HQSAM, to enhance segmentationperformance. Our extensive experiments on multiple datasets, encompassingvarious anatomies and modalities, reveal that FMs, particularly with the HQHSAMdecode head, improve domain generalization for medical image segmentation.Moreover, we found that the effectiveness of PEFT techniques varies acrossdifferent FMs. These findings underscore the potential of FMs to enhance thedomain generalization performance of neural networks in medical imagesegmentation across diverse clinical settings, providing a solid foundation forfuture research. Code and models are available for research purposes at\url{https://github.com/kerem-cekmeceli/Foundation-Models-for-Medical-Imagery}.
Resource
2024 EN
Tathagata Srimani · Robert Radway · Masoud Mohseni
+2 more
The vast majority of 21st century AI workloads are based on gradient-baseddeterministic algorithms such as backpropagation. One of the key reasons forthe dominance of deterministic ML algorithms is the emergence of powerfulhardware accelerators (GPU and TPU) that have enabled the wide-scale adoptionand implementation of these algorithms. Meanwhile, discrete and probabilisticMonte Carlo algorithms have long been recognized as one of the most successfulalgorithms in all of computing with a wide range of applications. Specifically,Markov Chain Monte Carlo (MCMC) algorithm families have emerged as the mostwidely used and effective method for discrete combinatorial optimization andprobabilistic sampling problems. We adopt a hardware-centric perspective onprobabilistic computing, outlining the challenges and potential futuredirections to advance this field. We identify two critical research areas: 3Dintegration using MOSAICs (Monolithic/Stacked/Assembled ICs) and the concept ofIllusion, a hardware-agnostic distributed computing framework designed to scaleprobabilistic accelerators.
Resource
2024 EN
Wei Liu · Kerem Delikoyun · Qianyu Chen
+7 more
Off-axis digital holographic microscopy is a high-throughput, label-freeimaging technology that provides three-dimensional, high-resolution informationabout samples, particularly useful in large-scale cellular imaging. However,the hologram reconstruction process poses a significant bottleneck for timelydata analysis. To address this challenge, we propose a novel reconstructionapproach that integrates deep learning with the physical principles of off-axisholography. We initialized part of the network weights based on the physicalprinciple and then fine-tuned them via weakly supersized learning. Our off-axishologram network (OAH-Net) retrieves phase and amplitude images with errorsthat fall within the measurement error range attributable to hardware, and itsreconstruction speed significantly surpasses the microscope's acquisition rate.Crucially, OAH-Net demonstrates remarkable external generalization capabilitieson unseen samples with distinct patterns and can be seamlessly integrated withother models for downstream tasks to achieve end-to-end real-time hologramanalysis. This capability further expands off-axis holography's applications inboth biological and medical studies.
Resource
2024 EN
Kerem Karatas · Bunyamin Ozkal · Abdullah H. Cosar
+1 more
In this research, we have fabricated a Ta/HfO$_{2-X}$/Mo-based single-cellmemristor, a unique configuration worldwide. The synaptic behaviour of Tantalumand Molybdenum electrodes on an HfOx-based memristor device has beeninvestigated. HfO$_{2-X}$ (15 nm) was grown using the Pulsed Laser Deposition(PLD) method, and electrodes were fabricated using a sputtering system andphotolithography method. The metal oxide stoichiometry was ascertained viaX-ray photoelectron spectroscopy (XPS). Long-term Potentiation (LTP) and pairedpulsed facilitation (PPF) characteristics, which play a significant role in thelearning processes of artificial neural networks, have been successfullyobtained. Current-voltage measurements and retention tests were performed todetermine the SET and RESET states of the device in appropriate ranges. Theresults show that this memristor device is a strong candidate for ArtificialNeural Network (ANN) applications.
Resource
2024 EN
Kubra Duran · Lal Verda Cakir · Mehmet Ozdem
+2 more
6G networks are envisioned to enable a wide range of applications, such asautonomous vehicles and smart cities. However, this rapid expansion of networktopologies makes the management of 6G wireless networks more complex and leadsto performance degradation. Even though state-of-the-art applications onnetwork services are providing promising results, they also risk disrupting thenetwork's performance. To overcome this, the services have to leverage what-ifimplementations covering a variety of scenarios. At this point, traditionalsimulations fall short of encompassing the dynamism and complexity of aphysical network. To overcome these challenges, we propose the GenerativeAI-based Digital Twins. For this, we derive an optimization formula todifferentiate different network scenarios by considering the specific keyperformance indicators (KPIs) for wireless networks. Then, we fed this formulato the generative AI with the historical twins and real-time twins to startgenerating the desired topologies. To evaluate the performance, we implementnetwork and smart-city-oriented services, namely massive connectivity, tinyinstant communication, right-time synchronization, and truck path routes. Thesimulation results reveal that our approach can achieve 38% more stable networkthroughput in high device density scenarios. Furthermore, the generatedscenario accuracy is able to reach up to 98% level, surpassing the baselines.
Resource
2024 EN
Kubra Duran · Mehmet Ozdem · Kerem Gursu
+1 more
The dramatic increase in the number of smart services and their diversityposes a significant challenge in Internet of Things (IoT) networks:heterogeneity. This causes significant quality of service (QoS) degradation inIoT networks. In addition, the constraints of IoT devices in terms ofcomputational capability and energy resources add extra complexity to this.However, the current studies remain insufficient to solve this problem due tothe lack of cognitive action recommendations. Therefore, we propose aQ-learning-based Cognitive Service Management framework called Q-CSM. In thisframework, we first design an IoT Agent Manager to handle the heterogeneity indata formats. After that, we design a Q-learning-based recommendation engine tooptimize the devices' lifetime according to the predicted QoS behaviour of thechanging IoT network scenarios. We apply the proposed cognitive management to asmart city scenario consisting of three specific services: wind turbines, solarpanels, and transportation systems. We note that our proposed cognitive methodachieves 38.7% faster response time to the dynamical IoT changes in topology.Furthermore, the proposed framework achieves 19.8% longer lifetime on averagefor constrained IoT devices thanks to its Q-learning-based cognitive decisioncapability. In addition, we explore the most successive learning rate value inthe Q-learning run through the exploration and exploitation phases.
Resource
2024 EN
Arif Kerem Dayi · Sitan Chen
LoRA has emerged as one of the de facto methods for fine-tuning foundationmodels with low computational cost and memory footprint. The idea is to onlytrain a low-rank perturbation to the weights of a pre-trained model, givensupervised data for a downstream task. Despite its empirical sucess, from amathematical perspective it remains poorly understood what learning mechanismsensure that gradient descent converges to useful low-rank perturbations. In this work we study low-rank fine-tuning in a student-teacher setting. Weare given the weights of a two-layer base model $f$, as well as i.i.d. samples$(x,f^*(x))$ where $x$ is Gaussian and $f^*$ is the teacher model given byperturbing the weights of $f$ by a rank-1 matrix. This generalizes the settingof generalized linear model (GLM) regression where the weights of $f$ are zero. When the rank-1 perturbation is comparable in norm to the weight matrix of$f$, the training dynamics are nonlinear. Nevertheless, in this regime we proveunder mild assumptions that a student model which is initialized at the basemodel and trained with online gradient descent will converge to the teacher in$dk^{O(1)}$ iterations, where $k$ is the number of neurons in $f$. Importantly,unlike in the GLM setting, the complexity does not depend on fine-grainedproperties of the activation's Hermite expansion. We also prove that in oursetting, learning the teacher model "from scratch'' can require significantlymore iterations.