Publications

Computer systems have grown in complexity to the point where manual inspection of system behaviour for purposes of malfunction detection have become unfeasible. As these systems output voluminous logs of their activity, machine led analysis of them is a growing need with already several existing solutions. These largely depend on having hand-crafted features, require raw log preprocessing and feature extraction or use supervised learning necessitating having a labeled log dataset not always easily procurable. We propose a two part deep autoencoder model with LSTM units that requires no hand-crafted features, no preprocessing of data as it works on raw text and outputs an anomaly score for each log entry. This anomaly score represents the rarity of a log event both in terms of its content and temporal context. The model was trained and tested on a dataset of HDFS logs containing 2 million raw lines of which half was used for training and half for testing. While this model cannot match the performance of a supervised binary classifier, it could be a useful tool as a coarse filter for manual inspection of log files where a labeled dataset is unavailable.

Motivation: High-Throughput Next-Generation-Sequencing can generate huge sequence files, whose analysis requires alignment algorithms that are typically very demanding in terms of memory and computational resources. This is a significant issue, especially for machines with limited hardware capabilities. As the redundancy of the sequences typically increases with coverage, collapsing such files into compact sets of non-redundant reads has the two-fold advantage of reducing file size and speeding-up the alignment, avoiding to map the same sequence multiple times. Method: BioSeqZip generates compact and sorted lists of alignment-ready non-redundant sequences, keeping track of their occurrences in the raw files as well as of their quality score information. By exploiting a memory-constrained external sorting algorithm, it can be executed on either single or multi-sample data-sets even on computers with medium computational capabilities. On request, it can even re-expand the compacted files to their original state. Results: Our extensive experiments on RNA-seq data show that BioSeqZip considerably brings down the computational costs of a standard sequence analysis pipeline, with particular benefits for the alignment procedures that typically have the highest requirements in terms of memory and execution time. In our tests, BioSeqZip was able to compact 2.7 billions of reads into 963 millions of unique tags reducing the size of sequence files up to 70% and speeding-up the alignment by 50% at least. Availability: BioSeqZip is available at https://github.com/bioinformatics-polito/BioSeqZip Supplementary information: Supplementary data are available at Bioinformatics online.

Replicating the human ability to connect Vision and Language has recently been gaining a lot of attention in the Computer Vision and the Natural Language Processing communities. This research effort has resulted in algorithms that can retrieve images from textual descriptions and vice versa, when realistic images and sentences with simple semantics are employed and when paired training data is provided. In this paper, we go beyond these limitations and tackle the design of visual-semantic algorithms in the domain of the Digital Humanities. This setting not only advertises more complex visual and semantic structures but also features a significant lack of training data which makes the use of fully-supervised approaches infeasible. With this aim, we propose a joint visual-semantic embedding that can automatically align illustrations and textual elements without paired supervision. This is achieved by transferring the knowledge learned on ordinary visual-semantic datasets to the artistic domain. Experiments, performed on two datasets specifically designed for this domain, validate the proposed strategies and quantify the domain shift between natural images and artworks.

La procedura che rende un algoritmo più efficiente in termini di requisiti di memoria o tempo di esecuzione si chiama ottimizzazione e rappresenta un passaggio cruciale nell'elaborazione di immagini e video. È raro che il processo di ottimizzazione produca un algoritmo ottimo in senso assoluto, ma spesso occorre raggiungere un compromesso tra i requisiti di tempo e quelli di memoria. Ad ogni modo, esistono molti scenari in cui il tempo di esecuzione totale richiesto per completare un'attività è il vincolo più restrittivo. Gli algoritmi di elaborazione di immagini binarie, ad esempio, rappresentano un'operazione fondamentale nella maggior parte dei sistemi di analisi di immagini e video all'avanguardia, anche quando questi sono basati su tecniche di deep learning. Avere un'implementazione efficiente è quindi essenziale, specialmente quando questi sistemi devono essere impiegati in scenari con vincoli temporali, dove compromettere la qualità del risultato, o fare affidamento su hardware più performante, non è una strada percorribile. Questa tesi introduce ed esplora diversi approcci per l'ottimizzazione degli algoritmi di elaborazione di immagini binarie modellabili con tabelle decisionali. Esistono diversi problemi che possono essere definiti in questo modo: l’etichettatura delle componenti connesse, il thinning, il chain code e gli operatori morfologici sono alcuni di questi. In generale, tutti gli algoritmi in cui il valore di output per ciascun pixel dell'immagine è ottenuto dal valore del pixel stesso e di alcuni dei suoi vicini possono essere definiti utilizzando tabelle decisionali. Concentrandosi sull'etichettatura delle componenti connesse, vengono analizzati gli approcci all'avanguardia sia per ambienti sequenziali basati su CPU che per ambienti paralleli basati su CPU e GPU, focalizzandosi su come misurare in modo equo le prestazioni. Vengono quindi introdotti nuovi approcci per migliorare ulteriormente le prestazioni in termini di tempo totale di esecuzione, mostrando come queste tecniche possano essere generalizzate per migliorare qualsiasi algoritmo modellabile con tabelle decisionali. Infine, viene presentato un framework che consente di applicare automaticamente molte delle strategie di ottimizzazione precedentemente descritte ed analizzate ad un determinato algoritmo. Il framework, chiamato GRAPHGEN, prende come input una definizione del problema in termini di condizioni da verificare e azioni da eseguire ed è in grado di produrre come output il codice C/C++ che include tutte le ottimizzazioni necessarie. Rispetto agli approcci esistenti, gli algoritmi generati con GRAPHGEN hanno prestazioni significativamente migliori, sia su set di dati reali che su quelli sintetici.

Nowadays, robots are heavily used in factories for different tasks, most of them including grasping and manipulation of generic objects in unstructured scenarios. In order to better mimic a human operator involved in a grasping action, where he/she needs to identify the object and detect an optimal grasp by means of visual information, a widely adopted sensing solution is Artificial Vision. Nonetheless, state-of-art applications need long training and fine-tuning for manually build the object's model that is used at run-time during the normal operations, which reduce the overall operational throughput of the robotic system. To overcome such limits, the paper presents a framework based on Deep Convolutional Neural Networks (DCNN) to predict both single and multiple grasp poses for multiple objects all at once, using a single RGB image as input. Thanks to a novel loss function, our framework is trained in an end-to-end fashion and matches state-of-art accuracy with a substantially smaller architecture, which gives unprecedented real-time performances during experimental tests, and makes the application reliable for working on real robots. The system has been implemented using the ROS framework and tested on a Baxter collaborative robot.

Text-image retrieval has been recently becoming a hot-spot research field, thanks to the development of deeply-learnable architectures which can retrieve visual items given textual queries and vice-versa. The key idea of many state-of-the-art approaches has been that of learning a joint multi-modal embedding space in which text and images could be projected and compared. Here we take a different approach and reformulate the problem of text-image retrieval as that of learning a translation between the textual and visual domain. Our proposal leverages an end-to-end trainable architecture that can translate text into image features and vice versa and regularizes this mapping with a cycle-consistency criterion. Experimental evaluations for text-to-image and image-to-text retrieval, conducted on small, medium and large-scale datasets show consistent improvements over the baselines, thus confirming the appropriateness of using a cycle-consistent constrain for the text-image matching task.

Summary: In the last decade, increasing attention has been paid to the study of gene fusions. However, the problem of determining whether a gene fusion is a cancer driver or just a passenger mutation is still an open issue. Here we present DEEPrior, an inherently flexible deep learning tool with two modes (Inference and Retraining). Inference mode predicts the probability of a gene fusion being involved in an oncogenic process, by directly exploiting the amino acid sequence of the fused protein. Retraining mode allows to obtain a custom prediction model including new data provided by the user. Availability and implementation: Both DEEPrior and the protein fusions dataset are freely available from GitHub at (https://github.com/bioinformatics-polito/DEEPrior). The tool was designed to operate in Python 3.7, with minimal additional libraries. Supplementary information: Supplementary data are available at Bioinformatics online.

Continual Learning has inspired a plethora of approaches and evaluation settings; however, the majority of them overlooks the properties of a practical scenario, where the data stream cannot be shaped as a sequence of tasks and offline training is not viable. We work towards General Continual Learning (GCL), where task boundaries blur and the domain and class distributions shift either gradually or suddenly. We address it through mixing rehearsal with knowledge distillation and regularization; our simple baseline, Dark Experience Replay, matches the network's logits sampled throughout the optimization trajectory, thus promoting consistency with its past. By conducting an extensive analysis on both standard benchmarks and a novel GCL evaluation setting (MNIST-360), we show that such a seemingly simple baseline outperforms consolidated approaches and leverages limited resources. We further explore the generalization capabilities of our objective, showing its regularization being beneficial beyond mere performance.

State of the art Computer Vision techniques exploit the availability of large-scale datasets, most of which consist of images captured from the world as it is. This brings to an incompatibility between such methods and digital data from the artistic domain, on which current techniques under-perform. A possible solution is to reduce the domain shift at the pixel level, thus translating artistic images to realistic copies. In this paper, we present a model capable of translating paintings to photo-realistic images, trained without paired examples. The idea is to enforce a patch level similarity between real and generated images, aiming to reproduce photo-realistic details from a memory bank of real images. This is subsequently adopted in the context of an unpaired image-to-image translation framework, mapping each image from one distribution to a new one belonging to the other distribution. Qualitative and quantitative results are presented on Monet, Cezanne and Van Gogh paintings translation tasks, showing that our approach increases the realism of generated images with respect to the CycleGAN approach.

As vision and language techniques are widely applied to realistic images, there is a growing interest in designing visual-semantic models suitable for more complex and challenging scenarios. In this paper, we address the problem of cross-modal retrieval of images and sentences coming from the artistic domain. To this aim, we collect and manually annotate the Artpedia dataset that contains paintings and textual sentences describing both the visual content of the paintings and other contextual information. Thus, the problem is not only to match images and sentences, but also to identify which sentences actually describe the visual content of a given image. To this end, we devise a visual-semantic model that jointly addresses these two challenges by exploiting the latent alignment between visual and textual chunks. Experimental evaluations, obtained by comparing our model to different baselines, demonstrate the effectiveness of our solution and highlight the challenges of the proposed dataset. The Artpedia dataset is publicly available at: http://aimagelab.ing.unimore.it/artpedia.