Giovani in un'ora - Ciclo di seminari - Prima parte
- Day - Time: 29 January 2020, h.11:00
- Place: Area della Ricerca CNR di Pisa - Room: C-29
Fabio Valerio Massoli - "MiniMax Training for AdversarialRobustness"Abstract: The use of Deep Learning methods is nowadays widespreadin several research and industrial fields such as: vision, speechrecognition, self driving vehicles, etc. Despite their greatsuccess, it has been pointed out that these learning models areexposed to adversarial samples --- natural images augmented withan additive noise, imperceptible to humans, that maliciously foola neural network --- which set a serious limitation to the usageof these techniques in real-world applications. Several trainingalgorithms have been proposed to make the learning models morerobust against this type of threat, but unfortunately none ofthem has offered a final solution to this problem yet. Instead oftraining a neural network to make it robust to adversarialsamples, a different approach is to try to detect these maliciousdata. Indeed, it has been recently shown that it is possible todetect the adversarial samples in real world scenarios such asface recognition. In this context we propose our approach inwhich we tackle the adversarial attacks by means of a minimaxtwo-player game, in which we set a classifier and a detector towork in competition between each other. Specifically, for a giveninput, we extract the features maps from specific layers of theclassifier and we use them to build embeddings utilized to feedthe detector. Based on the embeddings properties, the detectorperforms the task of discerning adversarial inputs from naturalimages, while the classifier fulfills the task of adjusting itsweights so that the produced embeddings become robust toadversarial attacks by maximally confusing the detector. Thus,while training, the classifier inherently learns to be robust toadversarial attacks. The obtained system is then double-shielded:on one side, the classifier becomes more robust againstadversarial attacks, and on the other side, in the case of asuccessful attack fooling the classifier, the detector can easilyspot it.
Gaia Pavoni - "TagLab, a semi-automatic annotation tool forfast and accurate analysis of benthic species"Abstract: The use of autonomous data-driven robotics foracquiring underwater data is making large-scale underwaterimaging more and more popular, and tools to efficiently processand understand demographic changes and spatial dynamics of coralreef communities are strongly needed. Traditional acquisitiontechniques have resulted in the creation of thousands oforthorectified imagery, each capturing hundreds to thousands ofcoral colonies. Handling such streams of acquired data is hard tobe sustained. Current manual workflows that generate highlyaccurate and precise segmentation for fine-scale colony mappingnecessary for building demographic models are time-intensive (~1hour per m2), creating substantial bottlenecks in processingimage-based. While fully automated semantic segmentation cansignificantly reduce the amount of processing time, currentsolutions lack the human expert provided accuracy. We presentTagLab, an AI-powered configurable annotation tool designed tospeed up the human labeling and the automatic analysis of largemaps. Following the human-computer interaction paradigm, thissoftware integrates a hybrid approach based on multiple degreesof automation. Assisted labeling is supported by CNN-basednetworks specially trained for agnostic, (relative only to objectpartition) or semantic (also related to species) segmentation ofcorals. An intuitive Graphical User Interface (GUI) speeds upboth the human editing and the semi-automatic refinement ofuncertain predictions increasing the overall accuracy. TagLaboutputs annotated maps, statistics, or new training datasets,supporting the meantime the multi-temporal comparison of labels,the dataset preparation, the data analysis, and the validationsof predictions in an integrated way.
Leonardo Robol - "Detecting rank-structures in matrices"Abstract: Low-rank perturbations of symmetric and orthogonalmatrices appear frequently in applications, in particular inproblems related to rootfinding and/or eigenvalue problems. Forthis reason, very efficient eigensolvers are available formatrices in these classes. However, such structure is far frombeing apparent; for instance, given any matrix, how can I detectif it is a low-rank perturbation of a symmetric one? In contrast,symmetriy and, up to a certain extent, orthogonality are ratherimmediate to check. In this talk, we give elegantcharacterizations for these structures in terms of singularvalues, and we show that these can be used to solve the relateddistance problem as well (i.e., what is the minimum distance inEuclidean or Frobenius norms of a given matrix to the class ofsymmetric/orthogonal-plus-rank-k). Then, we show that thesecharacterizations can be used to recover the structure with acomplexity within the coplexity bound for the fast eigensolvers,and therefore allow to use them even when the structure is notapparent or known a priori. The analysis only require elementarytools of matrix analysis.