Marcello Pelillo

Marcello Pelillo

Ca' Foscari University of Venice · Drexel University · Frontiers in Computer Vision · Bertinoro International Center for Informatics
6 previous affiliations
286Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Marcello Pelillo focuses on computer vision, pattern recognition, and the application of evolutionary game theory to machine learning. His foundational work established the use of continuous optimization and game-theoretic dynamics for complex clustering and matching problems. More recently, his research addresses the security of machine learning systems, adversarial optimization, and energy-efficient neural network training.

Research Trajectory

In the 1990s, Pelillo's research explored the theoretical foundations of relaxation labeling and graph-based pattern recognition. He investigated how to systematically derive parameters for iterative contextual processes in Learning Compatibility Coefficients for Relaxation Labeling Processes (1994), aiming to reduce local ambiguities through global consistency. This interest in mapping discrete combinatorial problems to continuous domains led to Replicator Equations, Maximal Cliques, and Graph Isomorphism (1999). In this work, he introduced an energy-minimization framework based on the Motzkin-Straus theorem, demonstrating that the graph isomorphism problem could be effectively solved using replicator equations from evolutionary game theory.

This game-theoretic and continuous optimization approach naturally extended to structural matching in computer vision. In Matching Free Trees, Maximal Cliques, and Monotone Game Dynamics (2002), he mapped the problem of matching unrooted trees to finding maximal cliques in an association graph, solving it with payoff-monotonic dynamics to match articulated and deformed shapes. Around this time, Pelillo and his collaborators developed the concept of "dominant sets," a framework that generalizes the notion of a maximal complete subgraph to edge-weighted graphs. This concept was formalized for hierarchical structures in Dominant Sets and Hierarchical Clustering (2003), which introduced a parameterized regularization term to control the energy landscape and addressed the research direction of Overcoming the Chaining Effect in Hierarchical Clustering. The theoretical underpinnings and practical applications of this approach to image segmentation were further detailed in Dominant Sets and Pairwise Clustering (2007), establishing a correspondence between dominant sets and the extrema of a quadratic form over the standard simplex.

Moving into the late 2000s and 2010s, Pelillo continued to apply non-cooperative game theory to classic vision problems. In Matching as a non-cooperative game (2009), his group framed potential feature matches as "strategies" and their mutual compatibility as "payoffs," advancing the research direction of Evolutionary Selection for Robust Correspondence and Clustering by finding stable sets of correspondences that offered a robust alternative to traditional consensus methods.

During the 2010s, his work expanded to encompass the vulnerabilities of machine learning algorithms in adversarial environments. He participated in early investigations into the security of unsupervised learning, such as Is data clustering in adversarial settings secure? (2013). This work proposed a general framework to evaluate the impact of deliberate data manipulation on clustering processes, aligning with the broader area of Data Poisoning under Realistic Information Constraints. Concurrently, he contributed to large-scale remote sensing challenges, participating in the development of DOTA: A Large-Scale Dataset for Object Detection in Aerial Images (2018). This highly annotated dataset was designed to handle the massive variations in scale, orientation, and shape found in earth observation data, supporting research directions like Learning Oriented Bounding Boxes for Arbitrary Object Detection and Multi-Scale Architectures for Extreme Scale Variance in Remote Sensing.

In the 2020s, Pelillo has continued to refine game-theoretic models while increasingly focusing on the efficiency and security of deep learning. He applied graph transduction games to specialized semi-supervised learning tasks, such as the classification of historical artifacts in Two sides of the same coin: Improved ancient coin classification using Graph Transduction Games (2022), contributing to the area of Evolutionary Game Dynamics for Data Labeling. His group has also continued to advance core pattern recognition techniques, developing methods for Game-theoretic hypergraph matching with density enhancement (2023), exploring Towards Parameter-Free Clustering for Real-World Data (2023), and revisiting foundational contextual models in Exploiting Context in Handwriting Recognition Using Trainable Relaxation Labeling (2023).

A significant portion of his recent trajectory addresses the computational footprint and adversarial robustness of neural networks. In Minimizing Energy Consumption of Deep Learning Models by Energy-Aware Training (2023), his team introduced an algorithm that inverts known security attacks to force activation sparsity during training, allowing specialized hardware to skip unnecessary calculations and save energy. On the security front, he surveyed the landscape of adversarial threats in Machine Learning Security Against Data Poisoning: Are We There Yet? (2024), analyzing the gap between theoretical understanding and scalable real-world defenses. Most recently, he proposed $\sigma$-zero: Gradient-based Optimization of $\ell_0$-norm Adversarial Examples (2025), a highly efficient attack method that uses a smooth, differentiable approximation of the $\ell_0$ norm to find sparse adversarial perturbations without requiring complex hyperparameter tuning.

Current Directions

The most recent papers in the evidence indicate a strong focus on the intersection of optimization, security, and efficiency in deep learning systems. Current work investigates adversarial machine learning, particularly the generation of sparse adversarial examples using differentiable approximations of discrete norms. This research seeks to understand the vulnerabilities of neural networks to minimal, targeted perturbations. Alongside adversarial optimization, recent work addresses the broader landscape of data poisoning attacks, evaluating the readiness of current defenses for real-world deployment. Additionally, there is an ongoing focus on the computational efficiency of neural networks, specifically exploring energy-aware training methods that leverage activation sparsity to reduce inference costs on resource-constrained hardware.

Profile updated

See something wrong? Report a mistake