A talk of ILLS will be held on Thursday, November 24 at 12:00 in hybrid mode.

Title: Information Analysis and Methods for Representation Learning and Data-Driven Structural Detection
Jorge F. Silva, Professor a University of Chile, Santiago de Chile

Abstract:
Machine learning and Information Theory are two broad research areas with strong connections. In this presentation, we will cover two topics that explore de use of information-theoretic measures in learning. On the first topic, we will present results that show how adequate the adoption of mutual information is for predicting the operational quality of a transformation (or encoder) in classification. These results offer new insights into adopting information measures in machine learning, like mutual information and cross-entropy. For the second topic, we will discuss the idea of information sufficiency, representing a model’s latent structure, and explore a non-parametric data-driven method to detect this type of structure from data. We will focus on the learning-decision task of testing independence using a non-parametric mutual information estimator. We present non-asymptotic and asymptotic results that support the advantage of this approach and elaborate on applications for learning data structure and model-change detection.

References:
Silva and Tobar, “On the Interplay between Information Loss and Operation Loss in Representations for Classification,” in AISTATS2022.
Gonzales et al. “Data-Driven Representations for Testing Independence: Modeling, Analysis and Connection with Mutual Information Estimation,” IEEE Trans. on Signal Proc., 70, 2022.

Short-Bio:
Jorge F. Silva (Senior Member, IEEE) is an Associate Professor in the Electrical Engineering (EE) Department at Universidad de Chile and Principal Investigador with the Advanced Center of Electrical and Electronic Engineering in Valparaiso-Chile. Jorge F. Silva received an M.Sc. and Ph.D. in Electrical Engineering from the University of Southern California (USC), Los Angeles, CA, USA, 2005 and 2008, respectively. Jorge F. Silva was a Research Assistant with the Signal Analysis and Interpretation Laboratory (SAIL), USC, during 2003–2008 and was also a Research Intern with the Speech Research Group, Microsoft Corporation, Redmond, in 2005. He received the Outstanding Thesis Award 2009 for Theoretical Research of the Viterbi School of Engineering, the Viterbi Doctoral Fellowship 2007-2008, and Simon Ramo Scholarship 2007-2008 USC. He was an Associate Editor for the IEEE Transactions on Signal Processing from 2006 to 2008.

* In person: ETS-LIVIA, room A-3600.
* Zoom link: https://cnrs.zoom.us/j/96338640901?pwd=MkNlT0FFS1c1T2Z6c0dManpLc3l1dz09
* Meeting ID: 963 3864 0901

The next LIVIA seminar will be held on Thursday, August 18 at 12h00 in hybrid mode.

Title: Semi-Weakly Supervised Object Detection by Sampling Pseudo Ground-Truth Boxes
by Akhil Meethal, Ph.D. candidate at the LIVIA

Abstract:
Semi- and weakly-supervised learning have recently attracted considerable attention in the object detection literature since they can alleviate the cost of annotation needed to successfully train deep learning models. State-of-art approaches for semi-supervised learning rely on student-teacher models trained using a multi-stage process, and considerable data augmentation. Custom networks have been developed for the weakly-supervised setting, making it difficult to adapt to different detectors. In this paper, a weakly semi-supervised training method is introduced that reduces these training challenges, yet achieves state-of-the-art performance by leveraging only a small fraction of fully-labeled images with information in weakly-labeled images. In particular, our generic sampling-based learning strategy produces pseudo-ground-truth (GT) bounding box annotations in an online fashion, eliminating the need for multi-stage training, and student-teacher network configurations. These pseudo GT boxes are sampled from weakly-labeled images based on the categorical score of object proposals accumulated via a score propagation process. Empirical results on the Pascal VOC dataset, indicate that the proposed approach improves performance by 5.0% when using VOC 2007 as fully-labeled, and VOC 2012 as weak-labeled data. Also, with 5-10% fully annotated images, we observed an improvement of more than 10% in mAP, showing that a modest investment in image-level annotation, can substantially improve detection performance.

https://arxiv.org/abs/2204.00147

* In person: ETS-LIVIA, room A-3600.

The next LIVIA seminar will be held on Thursday, June 23 at 12h00 in hybrid mode.

Title: Privacy-Preserving Data Sharing and Security in Machine Learning Models
by Prof. Mohammadhadi Shateri, Department of Systems Engineering

Abstract: These days, many people admire the great effects and the importance of AI in different applications including healthcare, social media, transport, and so forth. As the two main components of any AI approach one can name the “learning model” and the “data”. The focus of the recent studies has been mostly on boosting the efficiency of the AI approaches by improving the current models or developing more efficient learning algorithms and collecting data samples. Although important, the fact that both the learning model and the process of collecting/sharing datasets can leak sensitive information about the users, received less attentions in the literature. In this talk, the privacy issues regarding the (machine) learning models and data sharing are discussed in terms of the current attack/defense mechanisms. Some practical examples in applications such as smart meters would be presented and several challenges and the current focus of research will be discussed.
Bio: Mohammadhadi Shateri received the Ph.D. in electrical engineering from McGill University, Montreal, Canada in 2021. He continued his work with McGill as a postdoctoral researcher until he joined École de technologie supérieure in June 2022 as an assistant professor. His research interests include machine learning, security of (machine) learning models, and secure data sharing with applications in health and smart grids, among others. He won several scholarships for supporting his research including MEDA (McGill engineering doctoral award), MGS (Manitoba graduate scholarship education and advanced learning, Province of Manitoba), and UMGF (University of Manitoba graduate fellowship).

* In person: ETS-LIVIA, room A-3600. Please confirm your presence if you attend in person.

The next LIVIA seminar will be held on Thursday, March 3rd at 12h00 by Zoom.

Title: Negative evidence for weakly supervised learning
by Soufiane Belhabi, postdoctoral fellow at LIVIA

Summary:
Class Activation Mapping (CAM) methods have recently gained much attention for weakly-supervised object localization (WSOL) tasks. They allow for CNN visualization and interpretation without training on fully annotated image datasets. CAM methods are typically integrated within off-the-shelf CNN backbones, such as ResNet50. Due to convolution and pooling operations, these backbones yield low resolution CAMs with a down-scaling factor of up to 32, contributing to inaccurate localizations. Interpolation is required to restore full size CAMs, yet it does not consider the statistical properties of objects, such as color and texture, leading to activations with inconsistent boundaries, and inaccurate localizations. As an alternative, we introduce a generic method for parametric upscaling of CAMs that allows constructing accurate full resolution CAMs (FCAMs). In particular, we propose a trainable decoding architecture that can be connected to any CNN classifier to produce highly accurate CAM localizations. Given an original low resolution CAM, foreground and background pixels are randomly sampled to fine-tune the decoder. Additional priors such as image statistics and size constraints are also considered to expand and refine object boundaries. Extensive experiments1, over three CNN backbones and six WSOL baselines on the CUB-200-2011 and OpenImages datasets, indicate that our F-CAM method yields a significant improvement in CAM localization accuracy. F-CAM performance is competitive with state-of-art WSOL methods, yet it requires fewer computations during inference. Additional experiments and ablations were conducted on histology datasets with a focus on negative evidence. Results showed the benefits of our method compared to state-of-the-art methods.

Papers:
https://arxiv.org/abs/2109.07069
https://arxiv.org/abs/2201.02445

The next LIVIA seminar will be held on Wednesday, September 28 at 12h00 in hybrid mode.

Title: Few-Shot Object Detection in Aerial Images
by Pierre Le Jeune, Ph.D. candidate at the L2TI, University Sorbonne, Paris

Abstract: Object Detection is a challenging task in computer vision. Recently, deep learning-based methods overcame classical algorithms both in terms of quality and speed. However, deep learning requires large annotated training sets to achieve such performance. Few-Shot Learning (FSL) aims to overcome this shortcoming by learning more efficiently on scarce data. While FSL was vastly explored in the literature, Few-Shot Object Detection (FSOD) only became a topic of interest very recently. Most authors develop and benchmark their methods on natural images and nothing guarantees the transfer of their performance on other kinds of images. This work focuses on applying FSOD to aerial images. First, we review the definition of FSOD and several existing methods to address this task. A performance analysis is done on aerial and natural images to understand the challenges of using such methods on aerial images. In light of this analysis, we propose a novel attention mechanism. It specifically targets small objects which appear to be extremely difficult to detect in the few-shot regime. Finally, we question the relevance of Intersection over Union (IoU) as a criterion for box similarity and suggest a scale-dependent version: Scaled-IoU which agrees better with human perception.

Bio: Pierre LE JEUNE is a PhD student at L2TI laboratory, University Sorbonne Paris Nord while working at COSE company. He received the M.Sc. degree in Mathematical Modelling and Computation from Danish Technical University (Copenhagen) and the M.Sc. in engineering from Centrale Nantes. His current research interests include Few-Shot Learning, Computer Vision and Deep Learning

* In person: ETS-LIVIA, room A-3600.