-
Function-space Parameterization of Neural Networks for Sequential Learning
Authors:
Aidan Scannell,
Riccardo Mereu,
Paul Chang,
Ella Tamir,
Joni Pajarinen,
Arno Solin
Abstract:
Sequential learning paradigms pose challenges for gradient-based deep learning due to difficulties incorporating new data and retaining prior knowledge. While Gaussian processes elegantly tackle these problems, they struggle with scalability and handling rich inputs, such as images. To address these issues, we introduce a technique that converts neural networks from weight space to function space,…
▽ More
Sequential learning paradigms pose challenges for gradient-based deep learning due to difficulties incorporating new data and retaining prior knowledge. While Gaussian processes elegantly tackle these problems, they struggle with scalability and handling rich inputs, such as images. To address these issues, we introduce a technique that converts neural networks from weight space to function space, through a dual parameterization. Our parameterization offers: (i) a way to scale function-space methods to large data sets via sparsification, (ii) retention of prior knowledge when access to past data is limited, and (iii) a mechanism to incorporate new data without retraining. Our experiments demonstrate that we can retain knowledge in continual learning and incorporate new data efficiently. We further show its strengths in uncertainty quantification and guiding exploration in model-based RL. Further information and code is available on the project website.
△ Less
Submitted 16 March, 2024;
originally announced March 2024.
-
Sparse Function-space Representation of Neural Networks
Authors:
Aidan Scannell,
Riccardo Mereu,
Paul Chang,
Ella Tamir,
Joni Pajarinen,
Arno Solin
Abstract:
Deep neural networks (NNs) are known to lack uncertainty estimates and struggle to incorporate new data. We present a method that mitigates these issues by converting NNs from weight space to function space, via a dual parameterization. Importantly, the dual parameterization enables us to formulate a sparse representation that captures information from the entire data set. This offers a compact an…
▽ More
Deep neural networks (NNs) are known to lack uncertainty estimates and struggle to incorporate new data. We present a method that mitigates these issues by converting NNs from weight space to function space, via a dual parameterization. Importantly, the dual parameterization enables us to formulate a sparse representation that captures information from the entire data set. This offers a compact and principled way of capturing uncertainty and enables us to incorporate new data without retraining whilst retaining predictive performance. We provide proof-of-concept demonstrations with the proposed approach for quantifying uncertainty in supervised learning on UCI benchmark tasks.
△ Less
Submitted 5 September, 2023;
originally announced September 2023.
-
Learning Sequential Descriptors for Sequence-based Visual Place Recognition
Authors:
Riccardo Mereu,
Gabriele Trivigno,
Gabriele Berton,
Carlo Masone,
Barbara Caputo
Abstract:
In robotics, Visual Place Recognition is a continuous process that receives as input a video stream to produce a hypothesis of the robot's current position within a map of known places. This task requires robust, scalable, and efficient techniques for real applications. This work proposes a detailed taxonomy of techniques using sequential descriptors, highlighting different mechanism to fuse the i…
▽ More
In robotics, Visual Place Recognition is a continuous process that receives as input a video stream to produce a hypothesis of the robot's current position within a map of known places. This task requires robust, scalable, and efficient techniques for real applications. This work proposes a detailed taxonomy of techniques using sequential descriptors, highlighting different mechanism to fuse the information from the individual images. This categorization is supported by a complete benchmark of experimental results that provides evidence on the strengths and weaknesses of these different architectural choices. In comparison to existing sequential descriptors methods, we further investigate the viability of Transformers instead of CNN backbones, and we propose a new ad-hoc sequence-level aggregator called SeqVLAD, which outperforms prior state of the art on different datasets. The code is available at https://github.com/vandal-vpr/vg-transformers.
△ Less
Submitted 8 July, 2022;
originally announced July 2022.
-
Deep Visual Geo-localization Benchmark
Authors:
Gabriele Berton,
Riccardo Mereu,
Gabriele Trivigno,
Carlo Masone,
Gabriela Csurka,
Torsten Sattler,
Barbara Caputo
Abstract:
In this paper, we propose a new open-source benchmarking framework for Visual Geo-localization (VG) that allows to build, train, and test a wide range of commonly used architectures, with the flexibility to change individual components of a geo-localization pipeline. The purpose of this framework is twofold: i) gaining insights into how different components and design choices in a VG pipeline impa…
▽ More
In this paper, we propose a new open-source benchmarking framework for Visual Geo-localization (VG) that allows to build, train, and test a wide range of commonly used architectures, with the flexibility to change individual components of a geo-localization pipeline. The purpose of this framework is twofold: i) gaining insights into how different components and design choices in a VG pipeline impact the final results, both in terms of performance (recall@N metric) and system requirements (such as execution time and memory consumption); ii) establish a systematic evaluation protocol for comparing different methods. Using the proposed framework, we perform a large suite of experiments which provide criteria for choosing backbone, aggregation and negative mining depending on the use-case and requirements. We also assess the impact of engineering techniques like pre/post-processing, data augmentation and image resizing, showing that better performance can be obtained through somewhat simple procedures: for example, downscaling the images' resolution to 80% can lead to similar results with a 36% savings in extraction time and dataset storage requirement. Code and trained models are available at https://deep-vg-bench.herokuapp.com/.
△ Less
Submitted 9 June, 2023; v1 submitted 7 April, 2022;
originally announced April 2022.
-
The Borexino Thermal Monitoring & Management System and simulations of the fluid-dynamics of the Borexino detector under asymmetrical, changing boundary conditions
Authors:
D. Bravo-Berguño,
R. Mereu,
P. Cavalcante,
M. Carlini,
A. Ianni,
A. Goretti,
F. Gabriele,
T. Wright,
Z. Yokley,
R. B. Vogelaar,
F. Calaprice,
F. Inzoli
Abstract:
A comprehensive monitoring system for the thermal environment inside the Borexino neutrino detector was developed and installed in order to reduce uncertainties in determining temperatures throughout the detector. A complementary thermal management system limits undesirable thermal couplings between the environment and Borexino's active sections. This strategy is bringing improved radioactive back…
▽ More
A comprehensive monitoring system for the thermal environment inside the Borexino neutrino detector was developed and installed in order to reduce uncertainties in determining temperatures throughout the detector. A complementary thermal management system limits undesirable thermal couplings between the environment and Borexino's active sections. This strategy is bringing improved radioactive background conditions to the region of interest for the physics signal thanks to reduced fluid mixing induced in the liquid scintillator. Although fluid-dynamical equilibrium has not yet been fully reached, and thermal fine-tuning is possible, the system has proven extremely effective at stabilizing the detector's thermal conditions while offering precise insights into its mechanisms of internal thermal transport. Furthermore, a Computational Fluid-Dynamics analysis has been performed, based on the empirical measurements provided by the thermal monitoring system, and providing information into present and future thermal trends. A two-dimensional modeling approach was implemented in order to achieve a proper understanding of the thermal and fluid-dynamics in Borexino. It was optimized for different regions and periods of interest, focusing on the most critical effects that were identified as influencing background concentrations. Literature experimental case studies were reproduced to benchmark the method and settings, and a Borexino-specific benchmark was implemented in order to validate the modeling approach for thermal transport. Finally, fully-convective models were applied to understand general and specific fluid motions impacting the detector's Active Volume.
△ Less
Submitted 15 December, 2017;
originally announced December 2017.
-
Fluid-dynamics in the Borexino Neutrino Detector: behavior of a pseudo-stably-stratified, near-equilibrium closed system under asymmetrical, changing boundary conditions
Authors:
David Bravo-Berguño,
Riccardo Mereu,
Robert Bruce Vogelaar,
Fabio Inzoli
Abstract:
The strategy to install Borexino's Thermal Monitoring and Management System (BTMMS) successfully stabilized the thermal environment inside the Borexino neutrino observatory, which is understood to be a necessary step to improve and minimize radioactive background contamination inside the active volume of the detector, allowing for it to achieve better sensitivity in the regions of interest. Two-di…
▽ More
The strategy to install Borexino's Thermal Monitoring and Management System (BTMMS) successfully stabilized the thermal environment inside the Borexino neutrino observatory, which is understood to be a necessary step to improve and minimize radioactive background contamination inside the active volume of the detector, allowing for it to achieve better sensitivity in the regions of interest. Two-dimensional numerical simulations to achieve a proper understanding of Borexino's fluid-dynamics were developed and optimized for different regions and periods of interest, focusing on the most critical effects that were identified as influencing background concentrations. Literature experimental case studies were reproduced to benchmark the method and settings, and a Borexino-specific benchmark was constructed in order to validate the model's thermal transport. Finally, fully-convective models were implemented to understand general and specific fluid motions impacting the active detector volume.
△ Less
Submitted 26 May, 2017;
originally announced May 2017.
-
The Borexino Thermal Monitoring and Management System
Authors:
David Bravo-Berguño,
Riccardo Mereu,
Paolo Cavalcante,
Marco Carlini,
Andrea Ianni,
Augusto Goretti,
Federico Gabriele,
Tristan Wright,
Zachary Yokley,
Robert Bruce Vogelaar,
Frank Calaprice
Abstract:
A monitoring and management system for the thermal environment inside the Borexino neutrino detector was developed and installed in order to reduce uncertainties in temperature determination and limit undesirable thermal couplings between the environment and Borexino's active sections. This strategy is bringing improved radioactive background conditions to the physics signal thanks to the reduced…
▽ More
A monitoring and management system for the thermal environment inside the Borexino neutrino detector was developed and installed in order to reduce uncertainties in temperature determination and limit undesirable thermal couplings between the environment and Borexino's active sections. This strategy is bringing improved radioactive background conditions to the physics signal thanks to the reduced mixing taking place in the liquid scintillator. While equilibrium has not yet been reached, and fine-tuning is possible, the system has proven extremely effective at greatly stabilizing the detector while offering precise insight into its internal thermal conditions and transport. Numerical simulations have also been used to feed this empirical data into a global model, providing information into present and future thermal trends.
△ Less
Submitted 25 October, 2017; v1 submitted 25 May, 2017;
originally announced May 2017.
