Andrea Vedaldi

4 papers accepted for presentation at ECCV 2024, to be held in Milan, Italy from September 29 to October 4, 2024. Thanks to postdocs, PhD students and our collaborators! 1- Xinyu Yang, Hossein Rahmani, Sue Black, Bryan Williams, "Weakly Supervised Co-training with Swapping Assignments for Semantic Segmentation". Code and paper are available at https://lnkd.in/eCXrAsDr. The proposed method, CoSA, is the first single-stage weakly supervised semantic segmentation approach to outperform all existing multi-stage methods including those with additional supervision, surpassing existing baselines by a substantial margin. 2- Feixiang Zhou, Bryan Williams, Hossein Rahmani, "Towards Adaptive Pseudo-label Learning for Semi-Supervised Temporal Action Localization". Code and paper will be available soon 3- Zhengbo Zhang, Li Xu, Duo Peng, Hossein Rahmani, Jun Liu, "Diff-Tracker: Text-to-Image Diffusion Models are Unsupervised Trackers". Code and paper will be available soon. 4- Xiaofei Hui, Qian Wu, Hossein Rahmani, Jun Liu, "Class-Agnostic Object Counting with Text-to-Image Diffusion Model". Code and paper will be available soon. #ECCV2024

128

2 Comments

🚀 Exploring Extrinsic Hallucinations in LLMs! 🚀 The article delves into the causes and solutions for hallucinations in large language models (LLMs). 🌟Key highlights: - Causes of Hallucinations: Addressing issues in pre-training data and the challenges of fine-tuning with new knowledge. - Detection Methods: Utilizing techniques like retrieval-augmented evaluation and sampling-based detection to identify hallucinations. - Anti-Hallucination Techniques: Implementing strategies such as Retrieval-Augmented Generation (RAG), fine-tuning for factuality, and employing retrieval methods to enhance accuracy. https://lnkd.in/g3j5kU5b For more updates, follow Rohit Agarwal! #AI #LLMs #MachineLearning #TechInnovation #DataScience #ModelAccuracy #FactualAI #HallucinationDetection

4

Stuff we've been up to in the last several months that reviewers also liked: With Kamron Zaidi (EngSci 2T4), we trained a neural network to identify moves in chess that people view as brilliant. (Brilliant move ≠ good move. An obvious checkmate is good but not brilliant). The system predicts moves to be identified as brilliant based on the shape of the game tree: long thin paths are probably more brilliant. Also, moves are predicted to be perceived as brilliant if a strong system rates them higher but a weak system rates them lower. https://lnkd.in/g-jCdzR6 (International Conference on Computational Creativity 2024) With Amy Saranchuk (EngSci 2T4), we identify an intriguing property of the self-supervised pre-training process when using the rotation task: the performance of the system seems to have a periodic response to the rotation angle (makes sense) with the specific response showing very different and sometimes contradictory patterns of response to the rotation angle (mysterious). (At the Conference on Robots and Vision Workshops 2024). https://lnkd.in/gMaUMiyw With Juho Kim (EngSci 2T4), we leverage large-scale analysis using LLMs to show that Southern US honour culture https://lnkd.in/gWhB8buP) leaves its traces in internet flame wars on reddit: redditors geolocated to the US South are statistically more likely to retaliate (verbally) in response to personal attacks on the internet. (International Conference on Computational Social Science 2024) With Yiping (Sandra) Wang and Cameron Smith (EngSci 2T4), we propose synthetic datasets for exploring how computer vision networks classify images when colour is an important cue. (Usually the shape is much more important) (At the Conference on Robots and Vision Workshops 2024). Preliminary preprint: https://lnkd.in/gpgnYQUj With Jason Zuo (EngSci 2T4), we explore how Transformers can be trained on tasks that require position information without positional encodings. Intruiguingly, we find that correlations of the activations in the Transformer layers indicate that even untrained Transformers contain position information https://lnkd.in/gaKkMigS (At the Meeting of the North American Chapter Association of Computational Linguistics Student Research Workshop 2024) With Tracy Qian (EngSci 2T4) and building on work with Max Piasevoli , we show that the critic component of a WGAN can be used for model selection when fitting hierarchical models https://lnkd.in/gQVjS33K (At the Statistical Society of Canada Annual Meeting 2024) With Kexin Chen (MIE M.Eng '24), we've trained interpretable LLMs on reddit data to identify common patterns in posts of people with suspected Lyme disease and reported mental health symptoms (at the AI in Medicine 2024 conference)

55

1 Comment

Our paper on quantum computed Green's functions has been published! 🎉 https://lnkd.in/e6aeasRZ

74

6 Comments

Sharing our recent work on model editing Tl;dr - a novel evaluation protocol that measures the efficacy and impact of model editing in long-form generative settings. Video: https://lnkd.in/dAae3rn6 Paper: https://lnkd.in/d5gcvDBg The current evaluation of model editing is limited to measuring the performance using a few tokens after an input prompt. A successful model edit must result in a consistent change in knowledge beyond the next few words and the current sentence. In this work, we design a protocol (LEME)  for evaluating the consistency in the generation of models over long text after being edited. Our results highlight new insights into the limitations of current model editing methods such as lexical cohesion and topical drift issues. #naacl2024, #generativeAI, #NLP, #Deeplearning

78

6 Comments

A new lecture in the Course: "Introduction to Quantum Computing", has been published: https://lnkd.in/dAbjGRHk Lecture 14: The CHSH Game In this lecture we introduce nonlocal games, and in particular the CHSH game. It is one of the Bell tests that show the proper difference between quantum information and the counterpart classical realm. We game protocol is defined and we show that any classical strategy has at most 75% of winning. However, by going quantum, particularly, sharing and entanglement-bit between Alice and Bob, an 85% chance of winning is achieved. The instruction language is Arabic, however, all the material, slides, examples, definitions, concepts, etc., are all given in English.

10

Top Important LLMs Papers for the Week from 01/07 to 07/07 Link: https://lnkd.in/dwRnhisr Large language models (LLMs) have advanced rapidly in recent years. As new generations of models are developed, researchers and engineers need to stay informed on the latest progress. This article summarizes some of the most important LLM papers published during the First Week of July 2024. The papers cover various topics shaping the next generation of language models, from model optimization and scaling to reasoning, benchmarking, and enhancing performance. Keeping up with novel LLM research across these domains will help guide continued progress toward models that are more capable, robust, and aligned with human values. ➡ Do not miss the 60% discount on To Data & Beyond yearly subscription: https://lnkd.in/gmwFUqYr

2

In this new paper we explore the use of classical programming techniques in implementing the quantum lattice Boltzmann method in the Intel Quantum SDK. https://lnkd.in/gpDSFe6b

13

📢 Thrilled to share our latest paper on #LLM safety: "Are Large Language Models Really Bias-Free? Jailbreak Prompts for Assessing Adversarial Robustness to Bias Elicitation" (✍️ w/ Giada Cosenza, Alessio Orsino, and Domenico Talia). Our research offers an in-depth evaluation of the true resilience of widely used LLMs against biases and stereotypes across various scales. ** Key contributions ** - 🔍 Identification of the most prevalent biases in the responses of the latest LLMs, assessing their implications for model #safety, #robustness, and #fairness. - 📊 Detailed examination of how LLMs respond to bias elicitation prompts, including whether they decline, debias, or favor stereotypes and counterstereotypes. - 🛠️ Comprehensive experiments using advanced #jailbreak techniques to challenge LLMs and assess their vulnerability to adversarial prompting. Our findings emphasize the urgent need to improve alignment and mitigation strategies for creating more robust, fair and inclusive AI systems. Dive into the details of our study here: 👉📄 https://lnkd.in/dQ92U3HJ #AI #LLM #SustainableAI #AISafety #AIFairness #AdversarialRobustness #Jailbreak

24

#TMLR Proud of the work of our amazing team members Amirhosein Ghasemabadi and Kamran Janjua on #ImageRestoration.  Our paper, titled "CascadedGaze: Efficiency in Global Context Extraction for Image Restoration", has been accepted at TMLR. We present an innovative non-attention-based architecture that is not only computationally efficient but also competitive with SOTA methods on tasks such as real and Gaussian image denoising and single image deblurring, while using fewer MACS.  Paper: https://lnkd.in/gEhEZcyy Code: https://lnkd.in/gu8Y5zSZ Congrats to the team, Prof. Di Niu, and our collaborators Also, Our team is #recruiting. The position will be posted later this week. #huawei #huaweitechnologies #computervision #cv #artificialintelligence #imagerestoration #deblurring #denoising #deeplearning

28

4 Comments

Read #RNCP latest paper accepted for publication in IEEE Journal of Quantum Electronics that introduces the use of four-wave mixing (FWM) in photonic neuromorphic computing. FWM is a complex nonlinear process attributed to Kerr effect and can be triggered in highly nonlinear fibers and silicon photonics nano-waveguides. The paper shows how it can be leveraged to provide arbitrary nonlinear activation functions and complex polynomial processing as an extreme learning machine modality in specific nonlinear tasks with very high performance. Many thanks to Kostas Sozos, Stavros Deligiannidis and my colleague Charis Mesaritakis, Ph.D. This work was funded by our HFRI project #NOOKHFRI. #neuromorphiccomputing #photonics #nonlinearoptics

46

Check out the latest research on high-performance graph processing from a hardware/software co-design perspective! This paper dives deep into innovative strategies to optimize computational efficiency. https://lnkd.in/gUXumc5x 廖小飞 Hai Jin Long Zheng Zhiyuan Shao

A new publication titled “Deep Learning-Based Superpixel Texture Analysis for Crack Detection in Multi-Modal Infrastructure Images” is now available on NDT MDPI. #ndt #imaging #deeplearning #textureanalysis #cracks #multimodalimages #infrastructure

7

If you have an interest in Deep Learning and Computer Vision, we invite you to submit your high-quality papers to our special issue. #deeplearning #computervision #AI

16

📢 I am thrilled to share that our paper "Forget More to Learn More" has been accepted at the European Conference on Computer Vision #ECCV2024, a premier computer vision conference, to be held in Milan, Italy.  📝 Summary: We propose a novel episodic learning strategy for domain adaptation using limited-to-no annotation by identifying and unlearning domain-specific features and prioritizing domain-agnostic ones. Initially, we focus on learning to discern domain-specific discriminative representations, which is subsequently followed by the process of unlearning them. This is accompanied by domain-agnostic feature mining using a Gaussian-guided latent alignment strategy. 📅 Stay tuned for the release of our code, coming soon alongside the camera-ready version of our paper. 🙃 Beyond the celebrations lie the unseen challenges of academic research—countless experiments, meetings, and the persistent fatigue that often go unmentioned. This year, the acceptance rate was ~18%, signifying that nearly 82% of researchers face the daunting task of restarting their rigorous efforts. What are your thoughts, fellow researchers? Feel free to connect with me for further discussions! #ECCV2024 #ComputerVision #DomainAdaptation #WeakSupervision #EpisodicLearning

211

3 Comments

https://lnkd.in/dvrt3AaT via @researchgate https://lnkd.in/dvrt3AaT Securing and validating AI-ML systems are major concerns for the industry and the society. Our paper on the state of the art reached 2500 reads in a few weeks.. check it out for the notion and a survey of the #security4ai issues Chan Yeob Yeun

19

The recent paper by Yilei Chen "Quantum Algorithms for Lattice Problems " (https://lnkd.in/dWH-vryZ) has attracted a lot of attention in the past few days, with numerous posts discussing the implications for post-quantum based candidates. The paper was updated two days ago and an error was found: "Update on April 18: Step 9 of the algorithm contains a bug which I don't know how to fix. See Section 3.5.9 ( Page 37) for details".

10

2 Comments

Check out this great preprint on how to train foundation models on neural activity from multiple different brain regions https://lnkd.in/d3ijbNd5

2