1,721,065 research outputs found
B flavour tagging with leptons and measurement of the CP violation phase phi_s in the B_s^0 -> J/psi phi decay at the CMS experiment
Full text linkThis thesis presents the development and optimization of an algorithm used to determine the flavour at production time of neutral B^0 and B_s^0 mesons. The flavour tagging algorithm developed in this thesis exploits muons and electrons produced in the semileptonic decay of the additional b-hadron produced in pp -> bbX collisions at the LHC. The charge of the lepton is used to infer the flavour of the neutral b-meson. Three simulated samples of B_s^0 -> J/psi phi, B^+ -> J/psi K^+ and B^0 -> J/psi K* decays are exploited to develop and test the algorithm. Two independent neural networks are defined for muons and electrons, trained on 24 000 and 20 400 simulated B_s^0 -> J/psi phi events respectively, to parametrize the probability of wrong flavour tag omega of the algorithm. The tagging performances are further measured and calibrated on a sample of self-tagging B^+ -> J/psi K^+ decays collected by the CMS experiment during 2012, corresponding to 20 fb^-1 . The charge of the kaon univocally determines the flavour of the B^+ at production time and allows the direct measure of the mis-identification probability. A tagging power P_tag = epsilon_tag (1 - 2 omega )^2 of 0.833 +/- 0.024 (stat.) +/- 0.012 (syst.) % is measured using muons and 0.483 +/- 0.020 (stat.) +/- 0.003 (syst.) % using electrons. Combining the two algorithms results in the overall tagging power of 1.307 ± 0.031 (stat.) ± 0.007 (syst.) %.
The combined lepton flavour tagging algorithm is used in the measurement of the charge-parity (CP) violation parameters phi_s and DeltaGamma_s , sensitive to potential new physics processes not included in the standard model description. A time-dependent and flavour-tagged full angular analysis of the mu^+ mu^- K^+ K^- final state of the B_s^0 -> J/psi phi decay is performed based on the 2012 CMS dataset. A total of 49 000 reconstructed B_s^0 decays are used to extract the weak phase phi_s and decay with difference DeltaGamma_s values:
phi_s = -0.075 +/- 0.097 (stat.) +/- 0.031 (syst.) rad
DeltaGamma_s = 0.095 +/- 0.013 (stat.) +/- 0.007 (syst.) ps^-
Beyond Transformers: fault type detection in maintenance tickets with Kernel Methods, Boost Decision Trees and Neural Networks
Full text linkThe proper handling of customer tickets and maintenance requests is pivotal for enterprises as it directly impacts customer satisfaction.
The ability to rapidly and efficiently react and solve reported issues is in fact a key factor from the customers' perspective, resulting in positive feedback for the company, leading to higher economic and brand-image revenues.
The automatic detection of failures from maintenance tickets and support requests can grant faster and more efficient reactions to customers' equipment failures as well as reduced maintenance costs.
The analysis of support and maintenance requests is a well-known problem in Natural Language Processing (NLP).
State-of-the-art solutions in this field rely on Transformers models, pre-trained on large text corpora, and then fine-tuned on the specific downstream task.
However, due to their intrinsic nature, support requests are highly domain-specific and usually similar to short telegraph messages, where the focus is typically encapsulated in short sequences rather than in long dependencies.
Hence, ad-hoc methods for pattern recognition might provide comparable performances with respect to Transformers.
In this work, two alternative approaches are proposed, based on: Kernel methods in conjunction with Boost Decision Trees (SpectrumBoost), and Neural Networks for Multiple Representation Learning (DeepMRL).
These models have been tested and compared against state-of-the-art models on a real-world set of 131305 maintenance tickets in the Italian language, suggesting that the proposed models outperform Transformers both in the prediction accuracy and in the time and computational resources required for their training
Differential hypersaline stress response in Zygosaccharomyces rouxii complex yeasts: A physiological and transcriptional study
Full text linkThe Zygosaccharomyces rouxii complex comprises three distinct lineages of halotolerant yeasts relevant in food processing and spoilage, such as Z. sapae, Z. rouxii and a mosaic group of allodiploid strains. They manifest plastic genome architecture (variation in karyotype, ploidy level and Na+/H+ antiporter-encoding gene copy number), and exhibit diverse tolerances to salt concentrations. Here, we investigated accumulation of compatible osmolytes and transcriptional regulation of Na+/H+ antiporter-encoding ZrSOD genes during salt exposure in strains representative for the lineages, namely Z. sapae ABT301T (low salt tolerant), Z. rouxii CBS 732T (middle salt tolerant) and allodiploid strain ATCC 42981 (high salt tolerant). Growth curve modelling in 2 M NaCl-containing media supplemented with or without yeast extract as nitrogen source indicates that moderate salt tolerance of CBS 732T mainly depends on nitrogen availability rather than intrinsic inhibitory effects of salt. All the strains produce glycerol and not mannitol under salt stress and use two different glycerol balance strategies. ATCC 42981 produces comparatively more glycerol than Z. sapae and Z. rouxii under standard growth conditions and better retains it intracellularly under salt injuries. Conversely, Z. sapae and Z. rouxii enhance glycerol production under salt stress and intracellularly retain glycerol less efficiently than ATCC 42981. Expression analysis shows that, in diploid Z. sapae and allodiploid ATCC 42981, transcription of gene variants ZrSOD2-22/ZrSOD2 and ZrSOD22 is constitutive and salt unresponsive
Tree Tensor Network implemented on FPGA as ultra-low latency binary classifiers
Full text linkTensor Networks (TNs) are a computational framework traditionally used to model quantum many-body systems. Recent research has demonstrated that TNs can also be effectively applied to Machine Learning (ML) tasks, producing results comparable to conventional supervised learning methods. In this work, we investigate the use of Tree Tensor Networks (TTNs) for high-frequency real-time applications by harnessing the low-latency capabilities of Field-Programmable Gate Arrays (FPGAs). We present the implementation of TTN classifiers on FPGA hardware, optimized for performing inference on classical ML benchmarking datasets. Various degrees of parallelization are explored to evaluate the trade-offs between resource utilization and algorithm latency. By deploying these TTNs on a hardware accelerator and utilizing an FPGA integrated into a server, we fully offload the TTN inference process, demonstrating the system’s viability for real-time ML applications
Going Beyond Counting First Authors in Author Co-citation Analysis
Full text linkThe present study examines one of the fundamental aspects of author co-citation analysis (ACA) - the way co-citation
counts are defined. Co-citation counting provides the data on which all subsequent statistical analyses and mappings
are based, and we compare ACA results based on two different types of co-citation counting - the traditional type that
only counts the first one among a cited work's authors on the one hand and a non-traditional type that takes into
account the first 5 authors of a cited work on the other hand. Results indicate that the picture produced through this non-traditional author co-citation counting contains more coherent author groups and is therefore considerably clearer. However, this picture represents fewer specialties in the research field being studied than that produced through the traditional first-author co-citation counting when the same number of top-ranked authors is selected and analyzed. Reasons for these effects are discussed
Deep Learning Models for real-time Fusion Device Data Compression Algorithms
No full textreservedThe recently enhanced RFX-mod2 experiment, located at Consorzio RFX in Padova, presents a set of distinctive prospects for the advancement and validation of cutting-edge ML and DL algorithms and techniques, for plasma control. RFX-mod2 operates as a multi-configuration device, generating plasmas across various magnetic configurations: tokamak, ultra-low-q, and reversed field pinch (RFP). Among the peculiar features of RFX-mod2, it will provide a very high spatial resolution magnetic diagnostic with more than 1700 sensors, along with more than 200 actuator coils independently controlled.
However, the overall throughput required for the complete transfer of information from the sensors to the central control system cannot be handled in real-time. The compromise applied so far is a dual-channel acquisition: one channel for low-latency, low-bandwidth data acquisition, specifically designed for the control system, and a second channel for full-resolution data. The second channel takes advantage of the transient nature of the experimental setup by buffering the data locally and storing all the acquired raw data on the central acquisition server after the pulse. However, the useful information within the signals acquired by both channels is rich only for very short periods, resulting in large amounts of data that are mostly noise for the rest of the pulse. Additionally, most of the non-zero information signals can actually be modeled by a composition of known response functions.
This thesis focuses on the application of time series compression algorithms, specifically trained with the historical information acquired by the full length row signals in the RFX database, at the edge of the sensor devices.The recently enhanced RFX-mod2 experiment, located at Consorzio RFX in Padova, presents a set of distinctive prospects for the advancement and validation of cutting-edge ML and DL algorithms and techniques, for plasma control. RFX-mod2 operates as a multi-configuration device, generating plasmas across various magnetic configurations: tokamak, ultra-low-q, and reversed field pinch (RFP). Among the peculiar features of RFX-mod2, it will provide a very high spatial resolution magnetic diagnostic with more than 1700 sensors, along with more than 200 actuator coils independently controlled.
However, the overall throughput required for the complete transfer of information from the sensors to the central control system cannot be handled in real-time. The compromise applied so far is a dual-channel acquisition: one channel for low-latency, low-bandwidth data acquisition, specifically designed for the control system, and a second channel for full-resolution data. The second channel takes advantage of the transient nature of the experimental setup by buffering the data locally and storing all the acquired raw data on the central acquisition server after the pulse. However, the useful information within the signals acquired by both channels is rich only for very short periods, resulting in large amounts of data that are mostly noise for the rest of the pulse. Additionally, most of the non-zero information signals can actually be modeled by a composition of known response functions.
This thesis focuses on the application of time series compression algorithms, specifically trained with the historical information acquired by the full length row signals in the RFX database, at the edge of the sensor devices
Lightweight Model for Spatial Pattern Classification in Electrical Wafer Maps with Interactive Human Labeling
No full textreservedIn semiconductor manufacturing, spatial pattern recognition is essential for identifying defects or obtaining other crucial information on electrical wafer maps. During the wafer testing stage, deep learning methods are widely used for their powerful feature extraction capabilities.
The aim of this thesis is to propose a lightweight CNN model that achieves comparable or superior results to more complex models, enabling faster training and greater flexibility for experimentation and futurue improvements.
Additionally, we introduce a human-in-the-loop process, where domain experts label recognized spatial patterns. By retraining the model with these human-labeled images, our objective is to enhance classification accuracy and optimize defect detection.In semiconductor manufacturing, spatial pattern recognition is essential for identifying defects or obtaining other crucial information on electrical wafer maps. During the wafer testing stage, deep learning methods are widely used for their powerful feature extraction capabilities.
The aim of this thesis is to propose a lightweight CNN model that achieves comparable or superior results to more complex models, enabling faster training and greater flexibility for experimentation and futurue improvements.
Additionally, we introduce a human-in-the-loop process, where domain experts label recognized spatial patterns. By retraining the model with these human-labeled images, our objective is to enhance classification accuracy and optimize defect detection
Domain-Specific Cross-Lingual RAG Applications: Information Retrieval Fine Tuning with Synthetic Data Generation
Full text linkopenIn the context of Retrieval-Augmented Generation (RAG) systems, the development of high-performance retrieval algorithms for effective query-document matching is essential to answer the user's queries, especially with complex queries spanning multiple documents. This thesis proposes a novel framework leveraging generative capabilities of large language models to generate domain-specific, synthetic query-document pairs in a cross-lingual setting, where queries are created in multiple languages (including high and low-resource languages) and documents are primarily in Italian. The synthetic data generated is used to fine-tune transformer based dense sentence encoders, enabling more effective information retrieval.
The proposed approach introduces a query generation pipeline to enhance retrieval efficiency while preserving semantic integrity. Through comprehensive experiments, the framework demonstrates that fine-tuning smaller, task-specific models using high-quality synthetic data can outperform state-of-the-art retrieval solutions in terms of accuracy and computational efficiency (up to +7.5\% in retrieval MAP@10 and +4.4% in question answering accuracy).
This work highlights the potential of cross-lingual synthetic data generation as a cost-effective and scalable solution for improving domain-specific information retrieval in RAG applications, especially in scenarios involving multilingual queries and limited annotated data. This approach not only improves efficiency but also addresses critical security concerns. By enabling the evaluation of sensitive data to be conducted locally on company machines, risks associated with data leaks are significantly mitigated, resulting in enhanced compliance with data protection regulations.In the context of Retrieval-Augmented Generation (RAG) systems, the development of high-performance retrieval algorithms for effective query-document matching is essential to answer the user's queries, especially with complex queries spanning multiple documents. This thesis proposes a novel framework leveraging generative capabilities of large language models to generate domain-specific, synthetic query-document pairs in a cross-lingual setting, where queries are created in multiple languages (including high and low-resource languages) and documents are primarily in Italian. The synthetic data generated is used to fine-tune transformer based dense sentence encoders, enabling more effective information retrieval.
The proposed approach introduces a query generation pipeline to enhance retrieval efficiency while preserving semantic integrity. Through comprehensive experiments, the framework demonstrates that fine-tuning smaller, task-specific models using high-quality synthetic data can outperform state-of-the-art retrieval solutions in terms of accuracy and computational efficiency (up to +7.5\% in retrieval MAP@10 and +4.4% in question answering accuracy).
This work highlights the potential of cross-lingual synthetic data generation as a cost-effective and scalable solution for improving domain-specific information retrieval in RAG applications, especially in scenarios involving multilingual queries and limited annotated data. This approach not only improves efficiency but also addresses critical security concerns. By enabling the evaluation of sensitive data to be conducted locally on company machines, risks associated with data leaks are significantly mitigated, resulting in enhanced compliance with data protection regulations
- …
