Publications associées à Robustness of Classifiers to Universal Perturbations: A Geometric Perspective

Understanding generalization and robustness in modern deep learning

In this thesis, we study two closely related directions: robustness and generalization in modern deep learning. Deep learning models based on empirical risk minimization are known to be often non-robust to small, worst-case perturbations known as adversari ...

EPFL2024

ZigZag: Universal Sampling-free Uncertainty Estimation Through Two-Step Inference

Nikita Durasov, Nik Joel Dorndorf, Minh Hieu Lê

Whereas the ability of deep networks to produce useful predictions on many kinds of data has been amply demonstrated, estimating the reliability of these predictions remains challenging. Sampling approaches such as MC-Dropout and Deep Ensembles have emerge ...

2024

Seeking the new, learning from the unexpected: Computational models of surprise and novelty in the brain

Alireza Modirshanechi

Human babies have a natural desire to interact with new toys and objects, through which they learn how the world around them works, e.g., that glass shatters when dropped, but a rubber ball does not. When their predictions are proven incorrect, such as whe ...

EPFL2024

Incorporating Projective Geometry into Deep Learning

Michal Jan Tyszkiewicz

In this thesis we explore the applications of projective geometry, a mathematical theory of the relation between 3D scenes and their 2D images, in modern learning-based computer vision systems. This is an interesting research question which contradicts the ...

EPFL2024

Topologically Better Delineation of Curvilinear Structures

Doruk Oner

Curvilinear structures are frequently observed in a variety of domains and are essential for comprehending neural circuits, detecting fractures in materials, and determining road and irrigation canal networks. It can be costly and time-consuming to manuall ...

EPFL2023

Confidence Matters: Applications to Semantic Segmentation

Prabhu Teja Sivaprasad

The successes of deep learning for semantic segmentation can in be, in part, attributed to its scale: a notion that encapsulates the largeness of these computational architectures and the labeled datasets they are trained on. These resource requirements hi ...

EPFL2023

Stereo Confidence Estimation via Locally Adaptive Fusion and Knowledge Distillation

Pascal Frossard, Seungryong Kim

Stereo confidence estimation aims to estimate the reliability of the estimated disparity by stereo matching. Different from the previous methods that exploit the limited input modality, we present a novel method that estimates confidence map of an initial ...

IEEE COMPUTER SOC2023

Deep Learning for Localized-Haptic Feedback in Tactile Surfaces

Camilo Hernandez Mejia

Touchscreens are nowadays the preferred choice for user interfaces in consumer electronics. Significant technological advances have been made in terms of touch sensing and visual quality. However, the haptic feedback offered by commercial products is still ...

EPFL2023

Multienzyme deep learning models improve peptide de novo sequencing by mass spectrometry proteomics

Hamed Khakzad

Author summaryIn recent years, the application of deep learning represented a breakthrough in the mass spectrometry (MS) field by improving the assignment of the correct sequence of amino acids from observable MS spectra without prior knowledge, also known ...

PUBLIC LIBRARY SCIENCE2023

Deep Learning Generalization with Limited and Noisy Labels

Mahsa Forouzesh

Deep neural networks have become ubiquitous in today's technological landscape, finding their way in a vast array of applications. Deep supervised learning, which relies on large labeled datasets, has been particularly successful in areas such as image cla ...

EPFL2023

Predicting nonlinear optical scattering with physics-driven neural networks

Demetri Psaltis, Ahmed Ayoub, Carlo Gigli

Deep neural networks trained on physical losses are emerging as promising surrogates for nonlinear numerical solvers. These tools can predict solutions to Maxwell's equations and compute gradients of output fields with respect to the material and geometric ...

AIP Publishing2023

RMAML: Riemannian meta-learning with orthogonality constraints

Soumava Kumar Roy

Meta-learning is the core capability that enables intelligent systems to rapidly generalize their prior ex-perience to learn new tasks. In general, the optimization-based methods formalize the meta-learning as a bi-level optimization problem, that is a nes ...

ELSEVIER SCI LTD2023

3D reconstruction of curvilinear one-dimensional objects viewed in transmission electron microscopy

Gulnaz Ganeeva

To obtain a more complete understanding of material microstructure at the nanoscale and to gain profound insights into their properties, there is a growing need for more efficient and precise methods that can streamline the process of 3D imaging using a tr ...

EPFL2023

A new age in protein design empowered by deep learning

Bruno Emanuel Ferreira De Sousa Correia, Michael Bronstein, Casper Alexander Goverde, Ilia Igashov, Arne Schneuing, Hamed Khakzad

The rapid progress in the field of deep learning has had a significant impact on protein design. Deep learning methods have recently produced a breakthrough in protein structure prediction, leading to the availability of high-quality models for millions of ...

Cell Press2023

The inductive bias of deep learning: Connecting weights and functions

Guillermo Ortiz Jimenez

Years of a fierce competition have naturally selected the fittest deep learning algorithms. Yet, although these models work well in practice, we still lack a proper characterization of why they do so. This poses serious questions about the robustness, trus ...

EPFL2023

Robustness and invariance properties of image classifiers

Apostolos Modas

Deep neural networks have achieved impressive results in many image classification tasks. However, since their performance is usually measured in controlled settings, it is important to ensure that their decisions remain correct when deployed in noisy envi ...

EPFL2022

Understanding Deep Neural Function Approximation in Reinforcement Learning via ϵ-Greedy Exploration

Volkan Cevher, Luca Viano, Fanghui Liu

This paper provides a theoretical study of deep neural function approximation in reinforcement learning (RL) with the ϵ-greedy exploration under the online setting. This problem setting is motivated by the successful deep Q-networks (DQN) framework that fa ...

2022

Image Denoising with Control over Deep Network Hallucination

Sabine Süsstrunk, Majed El Helou, Qiyuan Liang

Deep image denoisers achieve state-of-the-art results but with a hidden cost. As witnessed in recent literature, these deep networks are capable of overfitting their training distributions, causing inaccurate hallucinations to be added to the output and ge ...

Society for Imaging Science and Technology (IS&T)2022

Leveraging Self-Supervision for Cross-Domain Crowd Counting

Pascal Fua, Weizhe Liu, Nikita Durasov

State-of-the-art methods for counting people in crowded scenes rely on deep networks to estimate crowd density. While effective, these data-driven approaches rely on large amount of data annotation to achieve good performance, which stops these models from ...

IEEE COMPUTER SOC2022

BIGPrior: Towards Decoupling Learned Prior Hallucination and Data Fidelity in Image Restoration

Sabine Süsstrunk, Majed El Helou

Classic image-restoration algorithms use a variety of priors, either implicitly or explicitly. Their priors are hand-designed and their corresponding weights are heuristically assigned. Hence, deep learning methods often produce superior image restoration ...

2022