DSC ADRIA’s Post

I’m happy to share that our paper “Continual Domain Randomization” was accepted at the IROS 2024 conference in Abu Dhabi. I want to thank the co-authors Sayantan Auddy, Mohammadhossein Malmir, Justus Piater, Alois Knoll and Nicolás Navarro-Guerrero, whose contribution and support made this possible. Summary: Domain randomization is often used in sim2real transfer of reinforcement learning policies for robotic tasks, but the combined randomization of many parameters might increase the task difficulty and the convergence to a suitable policy. We propose Continual Domain Randomization (CDR) to address this problem and to provide a more flexible training process in simulation. CDR combines domain randomization with continual learning to enable sequential training on a subset of randomization parameters at a time and subsequent sim2real transfer. A preprint of the paper, as well as the code and supplementary video are available at https://lnkd.in/dh-BHtQ5

🚀 Excited to share our latest research! Our study introduces a multimodal map-building method that integrates Multiple Object Tracking (MOT) and Gaussian Process Regression (GPR) to revolutionize how dynamic objects are represented in autonomous navigation systems. 👉 Explore the research in detail: - Paper: https://lnkd.in/g7_czruq - Video: https://lnkd.in/gj6AyMKD 🔑 Key Highlights: - Developed a novel framework that combines SLAM-based geometric mapping with dynamic object tracking to create a Target Activity (TA) Map. - Proposed a grid-based regression model to represent dynamic object activity levels, enabling efficient path planning in dynamic environments. Validated the approach in complex scenarios, such as libraries and public squares, using custom real-world datasets.

If you're seeking valuable insights and innovative ideas in data science, we have something special for you! Whether you missed the event or just want more, DSC ADRIA 24 talks are here for you! Today we’re sharing Miloš Košprdić’s talk: Verif.ai: A Trustworthy Scientific Generative Search Engine Milos Kosprdic presents Verif.ai, an innovative open-source scientific question-answering system designed to deliver referenced and verifiable answers. This pioneering system employs a dual-stage process, combining robust information retrieval techniques with fine-tuned generative models and advanced verification engines to ensure answer accuracy and reliability. By integrating semantic and lexical search techniques over scientific papers and utilizing a Retrieval-Augmented Generation (RAG) module, Verif.ai generates claims with references to source materials, bolstering answer credibility. Additionally, a Verification Engine cross-checks generated claims against original articles, identifying any potential inaccuracies. This speech by Milos Kosprdic was held on May 23rd at DSC Adria 2024 in Zagreb. Don't miss next year's DSC ADRIA 25 – it's set to be the biggest AI&Data conference in Croatia! 🤫 For the entire video just click the link below  https://lnkd.in/dyKUrqXD #ai #datascience #generative #search #engine #verifAI #ml #dscadria

Our great presenter Miloš discusses the Verif.ai project - a system uniquely designed for question-answering with verification in the biomedical field.

📢 #HighlyCited 1.  Automated Arabic Long-Tweet Classification Using Transfer Learning with BERT https://lnkd.in/geADXYmk 2. Edge-YOLO: Lightweight Infrared Object Detection Method Deployed on Edge Devices https://lnkd.in/gFm-W_WK 3. Adaptive Driver Face Feature Fatigue Detection Algorithm Research https://lnkd.in/g34YmR-j 4. A Financial Time-Series Prediction Model Based on Multiplex Attention and Linear Transformer Structure https://lnkd.in/gXMCPWZu

A few days ago I had the honor to present the joint work together with Nick Rossenbach: „On the Effect of Purely Synthetic Training Data for Different Automatic Speech Recognition Architectures“ at the SynData4GenAI 2024 workshop at #Interspeech2024 on Kos Island. In the follow-up work of our last paper (https://lnkd.in/ek83mucz) we looked at different ASR architectures and compared their behavior on synthetic only data. Through varying model parameters we found that the indicated overfitting shown in the train loss does not reflect in the final ASR performance. Furthermore, we saw that pre-trained speaker embeddings did only help the model in specific cases, while the baseline embeddings already showed a decent performance in this low resource scenario. If you are interested give the paper a read at: https://lnkd.in/ezRECAvP

Towards autonomous shotcrete construction: semantic 3D reconstruction for concrete deposition using stereo vision and deep learning Paper ID 156 - ISARC 2024 by Patrick Schmidt IAARC https://lnkd.in/gVphrU9E

🚀 Transforming Object Detection with DETR 🔍 The DETR (DEtection TRansformers) model is a game-changer in object detection, leveraging transformers and CNNs to simplify the process and boost accuracy. By eliminating the need for traditional post-processing steps like anchor boxes and non-maximum suppression (NMS), DETR introduces a more efficient, end-to-end detection pipeline. In my latest blog, I break down DETR's transformer-driven architecture, explain how it achieves state-of-the-art performance, and discuss why it’s a significant leap forward in computer vision. Let’s explore how DETR is reshaping the future of object detection! #AI #ObjectDetection #DETR #DeepLearning #Transformers #MachineLearning #CVPR

This paper derives a #pilot-#aided (#PA) #Bayesian #learning (#PA-#BL) framework for the #estimation #of #the #Terahertz (#THz) #MIMO channel that relies exclusively on the pilot beams transmitted. Since the framework designed can successfully operate in an #ill-#posed #model, it can verifiably lead to reduced pilot transmissions in comparison to conventional methodologies. The above paradigm is subsequently extended to additionally incorporate data symbols to derive a #Data-#Aided (#DA) BL #approach that performs joint data detection and #channel #state #information (#CSI) estimation. They will demonstrate that it is capable of improving the #dual-#wideband #channel’s estimate, despite further reducing the training overhead. The #Bayesian #Cramér-#Rao #bounds (#BCRLBs) are also obtained for explicitly characterizing the lower bounds on the #mean #squared #error (#MSE) of the PA-BL and DA-BL frameworks. ----Abhisha Garg, Prof. Suraj Srivastava, @Nimish Yadav, Aditya K Jagannatham, Lajos Hanzo More details can be found at this link: https://lnkd.in/g4tJVZ9W

Interesting approach to use iterative micro-segmentation, combination and a smoothing technique to generate the final segment to detect subtle junctures where topics shift and can be chunked meaningfully.