Talks and presentations

An overview of Graph Neural Networks up to Lagrangian Propagation GNNs

December 04, 2020

Seminar, Maasai Nice cote D'Azur, Nice Cote D'Azur, France

Several real-world applications are characterized by data that exhibit a complex structure that can be represented using graphs. The popularity of deep learning techniques renewed the interest in neural architectures able to process these patterns, inspired by the Graph Neural Network(GNN) model, proposed by SAILab. In this talk, I will present the original GNN model, recent evolutions (GCN, GATs, …) and our implementation framework in Tensorflow and PyTorch. The original GNNs encode the state of the nodes of the graph by means of an iterative diffusion procedure that, during the learning stage, must be computed at every epoch, until the fixed point of a learnable state transition function is reached, propagating the information among the neighbouring nodes. We propose a novel approach to learning in GNNs, based on constrained optimization in the Lagrangian framework, named Lagrangian Propagation Graph Neural Networks. Learning both the transition function and the node states is the outcome of a joint process, in which the state convergence procedure is implicitly expressed by a constraint satisfaction mechanism, avoiding iterative epoch-wise procedures and the network unfolding. Our computational structure searches for saddle points of the Lagrangian in the adjoint space composed of weights, nodes state variables and Lagrange multipliers. This process is further enhanced by multiple layers of constraints that accelerate the diffusion process. An experimental analysis shows that the proposed approach compares favourably with popular models on several benchmarks.

LabMeeting: Knowledge graphs for image classification and semantic navigation – from reality to virtual worlds

September 25, 2019

LabMeeting, SAILab, Siena, Italy

One characteristic that sets humans apart from modern learning-based computer vision algorithms is the ability to acquire knowledge about the world and use that knowledge to reason about the visual world. Humans can learn about the characteristics of objects and the relationships that occur between them to learn a large variety of visual concepts, often with few examples.

ACDL Satellite Workshop on Graph Neural Networks - Lagrangian Propagation Graph Neural Networks – A constraint-based formulation

July 22, 2019

Workshops, SAILab, Siena, Italy

GNNs exploit a set of state variables, each assigned to a graph node, and a diffusion mechanism of the states among neighbor nodes, to implement an iterative procedure to compute the fixed point of the (learnable) state transition function. We propose a novel approach to the state computation and the learning algorithm for GNNs, based on a constraint optimization task solved in the Lagrangian framework. The state convergence procedure is implicitly expressed by the constraint satisfaction mechanism and does not require a separate iterative phase for each epoch of the learning procedure. In fact, the computational structure is based on the search for saddle points of the Lagrangian in the adjoint space composed of weights, neural outputs (node states), and Lagrange multipliers.

LabMeeting: Adversarial Reprogramming of Neural Networks

September 06, 2018

LabMeeting, SAILab, Siena, Italy

Adversarial examples are defined as “inputs to machine learning models that an attacker has intentionally designed to cause the model to make a mistake”. Indeed, in the computer vision scenario it has been shown that well-crafted perturbation to input images can induce classification errors, such as confusing a cat with a computer. In general, adversarial attacks are designed to degrade the performances of models or to prompt the prediction of specific output classes. In this recent work, the authors introduce a framework in which the goal of adversarial attacks is to reprogram the target model to perform a completely new task. This is accomplished by optimizing for a single adversarial perturbation that can be added to all test-time inputs. In such a way, the target model performs a task chosen by the adversary when processing these inputs—even if it was not trained on this task.

LabMeeting: Excursus in the State of the art of object detection - Deep learning models, performance, practical tests

February 01, 2018

LabMeeting, SAILab, Siena, Italy

In image classification an image with a single object is the focus and the task is to say what it contains. But when we look at the world around us, we carry out far more complex tasks. There are multiple overlapping objects, different backgrounds and we  not only classify these different objects but also identify their boundaries, differences, and relations among them. This task fall under the name of object detection and instance segmentation.

LabMeeting: Capsule Networks

February 01, 2018

LabMeeting, SAILab, Siena, Italy

Convolutional neural networks that are used to recognize shapes typically use one or more layers of learned feature detectors that produce scalar outputs, interleaved with subsampling.