Towards Realistic Error Models for Tabular Data

Procedural and reporting guidelines are crucial in framing scientific practices and communications among researchers and the broader community. These guidelines aim to ensure transparency, reproducibility, and reliability in scientific research. Despite several methodological frameworks proposed by various initiatives to foster reproducibility, challenges such as data leakage and reproducibility remain prevalent. Recent studies have highlighted the transformative potential of incorporating the FAIR (Findable, Accessible, Interoperable, and Reusable) principles into workflows, particularly in contexts like software and machine learning model development, to promote open science. This study aims to introduce a comprehensive framework, designed to calibrate existing reporting guidelines against the FAIR principles. The goal is to enhance reproducibility and promote open science by integrating these principles into the scientific reporting process. We employed the "Best fit" framework synthesis approach which involves systematically reviewing and synthesizing existing frameworks and guidelines to identify best practices and gaps. We then proposed a series of defined workflows to align reporting guidelines with FAIR principles. A use case was developed to demonstrate the practical application of the framework. The integration of FAIR principles with established reporting guidelines through the framework effectively bridges the gap between FAIR metrics and traditional reporting standards. The framework provides a structured approach to enhance the findability, accessibility, interoperability, and reusability of scientific data and outputs. The use case demonstrated the practical benefits of the framework, showing improved data management and reporting practices. The framework addresses critical challenges in scientific research, such as data leakage and reproducibility issues. By embedding FAIR principles into reporting guidelines, the framework ensures that scientific outputs are more transparent, reliable, and reusable. This integration not only benefits researchers by improving data management practices but also enhances the overall scientific process by promoting open science and collaboration. The proposed framework successfully combines FAIR principles with reporting guidelines, offering a robust solution to enhance reproducibility and open science. This framework can be applied across various contexts, including software and machine learning model development stages, to foster a more transparent and collaborative scientific environment.

The unmanned aerial vehicles (UAVs) have been extensively used in providing intelligence such as target tracking. In our field experiments, a pre-trained deep neural network (DNN) is deployed on the UAV to identify a target from the captured video frames and enable UAV to keep tracking. However, tracking in real time by the DNN requires a lot of computational resources. This motivates us to consider offloading this type of machine learning (ML) tasks to a mobile edge computing (MEC) server. Specifically, we propose a novel hierarchical ML tasks distribution framework for the UAV tracking system, where the UAV is embedded with lower layers of the pre-trained convolutional neural network (CNN) model due to its limited computing capability, while the MEC server with rich computing resources will handle the higher layers of the CNN model. An optimization problem is formulated to minimize the CNN inference delay while taking into account the communications delay, computing time, and ML error. Insights are provided to understand the tradeoff between communications and ML computing in offloading decisions. Numerical results demonstrate the effectiveness of the proposed ML tasks distribution framework with the optimized offloading strategy.

Studies of Earth’s magnetic field and its sources rely on accurate geomagnetic field models derived from ground and satellite-based magnetic data. During the field model estimation, data errors are usually assumed to be uncorrelated in time and independent of position. However, limitations in the field model parameterization, especially regarding ionospheric and magnetospheric fields, lead to data errors that are not only larger than the expected measurement noise but are also correlated in time and vary with position. As a result, the obtained model uncertainties are often underestimated, making it more challenging to evaluate the reliability of recovered signals in the field models. This study investigates the effect of including correlated data errors in field modeling. The approach involves building a stochastic data error model to treat correlated errors due to unmodeled magnetospheric fields within the CHAOS geomagnetic field modeling framework. The error model parameters are estimated using empirical covariances computed from vector residuals between the satellite magnetic observations made by the Swarm satellites and the CHAOS geomagnetic field model. Field modeling experiments are performed with and without including the data error covariances described in the stochastic error model. The inclusion of data error covariances due to unmodeled magnetospheric fields leads to only small changes in the estimated internal field, but also a noticeable increase in model uncertainty for the sectoral coefficients. This highlights the significant impact of unmodeled magnetospheric fields and the importance of accurately defining data errors, including the covariances between observations, for interpreting the retrieved magnetic signals in geomagnetic field modeling. • A method is presented to build a stochastic data error model to treat unmodeled magnetospheric fields in geomagnetic field modeling. • Empirical means and covariances in space and time are computed from Swarm magnetic observations and are then used to estimate the parameters of the stochastic data error model. • Application of the resulting data error model in field modeling experiments shows only minor changes in the estimated internal field model, but also a noticeable increase in model uncertainty for its sectoral coefficients.

Image formation algorithms in a variety of applications have explicit or implicit dependence on a mathematical model of the observation process. Inaccuracies in the observation model may cause various degradations and artifacts in the reconstructed images. The application of interest in this paper is synthetic aperture radar (SAR) imaging, which particularly suffers from motion-induced model errors. These types of errors result in phase errors in SAR data, which cause defocusing of the reconstructed images. Particularly focusing on imaging of fields that admit a sparse representation, we propose a sparsity-driven method for joint SAR imaging and phase error correction. Phase error correction is performed during the image formation process. The problem is set up as an optimization problem in a nonquadratic regularization-based framework. The method involves an iterative algorithm, where each iteration of which consists of consecutive steps of image formation and model error correction. Experimental results show the effectiveness of the approach for various types of phase errors, as well as the improvements that it provides over existing techniques for model error compensation in SAR.

Recently there has been a rapidly growing demand for faster machine learning (ML) processing in data centers and migration of ML inference applications to edge devices. These developments have prompted both industry and academia to explore custom accelerators to optimize ML executions for performance and power. However, identifying which accelerator is best equipped for performing a particular ML task is challenging, especially given the growing range of ML tasks, the number of target environments, and the limited number of integrated modeling tools. To tackle this issue, it is of paramount importance to provide the computer architecture research community with a common framework capable of performing a comprehensive, uniform, and fair comparison across different accelerator designs targeting a particular ML task. To this aim, we propose a new framework named TFLITE-SOC (System On Chip) that integrates a lightweight system modeling library (SystemC) for fast design space exploration of custom ML accelerators into the build/execution environment of Tensorflow Lite (TFLite), a highly popular ML framework for ML inference. Using this approach, we are able to model and evaluate new accelerators developed in SystemC by leveraging the language's hierarchical design capabilities, resulting in faster design prototyping. Furthermore, any accelerator designed using TFLITE-SOC can be benchmarked for inference with any DNN model compatible with TFLite, which enables end-to-end DNN processing and detailed (i.e., per DNN layer) performance analysis. In addition to providing rapid prototyping, integrated benchmarking, and a range of platform configurations, TFLITE-SOC offers comprehensive performance analysis of accelerator occupancy and execution time breakdown as well as a rich set of modules that can be used by new accelerators to implement scaling up studies and optimized memory transfer protocols. We present our framework and demonstrate its utility by considering the design space of a TPU-like systolic array and describing possible directions for optimization. Using a compression technique, we implement an optimization targeting reducing the memory traffic between DRAM and on-device buffers. Compared to the baseline accelerator, our optimized design shows up to 1.26x speedup on accelerated operations and up to 1.19x speedup on end-to-end DNN execution.

Datasets of labeled network traces are essential for a multitude of machine learning (ML) tasks in networking, yet their availability is hindered by privacy and maintenance concerns, such as data staleness. To overcome this limitation, synthetic network traces can often augment existing datasets. Unfortunately, current synthetic trace generation methods, which typically produce only aggregated flow statistics or a few selected packet attributes, do not always suffice, especially when model training relies on having features that are only available from packet traces. This shortfall manifests in both insufficient statistical resemblance to real traces and suboptimal performance on ML tasks when employed for data augmentation. In this paper, we apply diffusion models to generate high-resolution synthetic network traffic traces. We present NetDiffusion, a tool that uses a finely-tuned, controlled variant of a Stable Diffusion model to generate synthetic network traffic that is high fidelity and conforms to protocol specifications. Our evaluation demonstrates that packet captures generated from NetDiffusion can achieve higher statistical similarity to real data and improved ML model performance than current state-of-the-art approaches (e.g., GAN-based approaches). Furthermore, our synthetic traces are compatible with common network analysis tools and support a myriad of network tasks, suggesting that NetDiffusion can serve a broader spectrum of network analysis and testing tasks, extending beyond ML-centric applications.

Datasets of labeled network traces are essential for a multitude of machine learning (ML) tasks in networking, yet their availability is hindered by privacy and maintenance concerns, such as data staleness. To overcome this limitation, synthetic network traces can often augment existing datasets. Unfortunately, current synthetic trace generation methods, which typically produce only aggregated flow statistics or a few selected packet attributes, do not always suffice, especially when model training relies on having features that are only available from packet traces. This shortfall manifests in both insufficient statistical resemblance to real traces and suboptimal performance on ML tasks when employed for data augmentation. In this paper, we apply diffusion models to generate high-resolution synthetic network traffic traces. We present NetDiffusion, a tool that uses a finely-tuned, controlled variant of a Stable Diffusion model to generate synthetic network traffic that is high fidelity and conforms to protocol specifications. Our evaluation demonstrates that packet captures generated from NetDiffusion can achieve higher statistical similarity to real data and improved ML model performance than current state-of-the-art approaches (e.g., GAN-based approaches). Furthermore, our synthetic traces are compatible with common network analysis tools and support a myriad of network tasks, suggesting that NetDiffusion can serve a broader spectrum of network analysis and testing tasks, extending beyond ML-centric applications.

The emerging paradigm of edge computing has revolutionized network applications, delivering computational power closer to the end-user. Consequently, Machine Learning (ML) tasks, typically performed in a data centre (Centralized Learning - CL), can now be offloaded to the edge (Edge Learning - EL) or mobile devices (Federated Learning - FL). While the inherent flexibility of such distributed schemes has drawn considerable attention, a thorough investigation on their resource consumption footprint is still missing.In our work, we consider a FL scheme and two EL variants, representing varying proximity to the end users (data sources) and corresponding levels of workload distribution across the network; namely Access Edge Learning (AEL), where edge nodes are essentially co-located with the base stations and Regional Edge Learning (REL), where they lie towards the network core. Based on real systems' measurements and user mobility traces, we devise a realistic simulation model to evaluate and compare the performance of the considered ML schemes under an image classification task. Our results indicate that FL and EL can act as viable alternatives to CL. Edge learning effectiveness is shaped by the configuration of edge nodes in the network with REL achieving the prominent combination of accuracy and bandwidth needs. Energy-wise, edge learning is shown to offer an attractive choice (for involved stakeholders) to offload centralised ML tasks.

Outsourcing Machine Learning (ML) tasks to cloud servers is a cost-effective solution when dealing with distributed data. However, outsourcing these tasks to cloud servers could lead to data breaches. Secure computing methods, such as Homomorphic Encryption (HE) and Trusted Execution Environments (TEE), have been used to protect outsourced data. Nevertheless, HE remains inefficient in processing complicated functions (e.g., non-linear functions) and TEE (e.g., Intel SGX) is not ideal for directly processing ML tasks due to side-channel attacks and parallel-unfriendly computation.In this paper, we propose a hybrid framework integrating SGX and HE, called HT2ML, to protect user's data and models. In HT2ML, HE-friendly functions are protected with HE and performed outside the enclave, while the remaining operations are performed inside the enclave obliviously. HT2ML leverages optimised HE matrix multiplications to accelerate HE computations outside the enclave while using oblivious blocks inside the enclave to prevent access-pattern-based attacks. We evaluate HT2ML using Linear Regression (LR) training and Convolutional Neural Network (CNN) inference as two instantiations. The performance results show that HT2ML is up to ∼11× faster than HE only baseline with 6-dimensional data in LR training. For CNN inference, HT2ML is ∼196× faster than the most recent approach (Xiao et al., ICDCS'21).

In recent years, Machine Learning (ML) models have led to a substantial improvement in hydrological predictions. It appears these models can distill information from catchment properties that is relevant for the relationship between meteorological drivers and streamflow, which has so far eluded hydrologists.In the first part of this talk, I shall demonstrate some of our attempts towards understanding these improvements. Utilising Autoencoders and intrinsic dimension estimators, we have shown that the wealth of available catchment properties can effectively be summarised into merely three features, insofar as they are relevant for streamflow prediction. Hybrid models, which combine the flexibility of ML models with mechanistic mass-balance models, are equally adept at predicting as pure ML models but come with only a few interpretable interior states. Combining these findings will, hopefully, bring us closer to understanding what these ML models seem to have 'grasped'.In the second part of the talk, I will address the issue of uncertainty quantification. I contend that error modelling should not be attempted on the residuals. Rather, we should model the errors where they originate, i.e., on the inputs, model states, and/or parameters. Such stochastic models are more adept at expressing the intricate distributions exhibited by real data. However, they come at the cost of a very large number of unobserved latent variables and thus pose a high-dimensional inference problem. This is particularly pertinent when our models include ML components. Fortunately, advances in inference algorithms and parallel computing infrastructure continue to extend the limits on the number of variables that can be inferred within a reasonable timeframe. I will present a straightforward example of a stochastic hydrological model with input uncertainty, where Hamiltonian Monte Carlo enables a comprehensive Bayesian inference of model parameters and the actual rain time-series simultaneously.

Research is needed to explore the limitations and potential for improvement of machine learning for building energy prediction. With this aim, the ASHRAE Great Energy Predictor III (GEPIII) Kaggle competition was launched in 2019. This effort was the largest building energy meter machine learning competition of its kind, with 4370 participants who submitted 39,403 predictions. The test dataset included two years of hourly whole building readings from 2380 meters in 1448 buildings at 16 locations. This paper analyzes the various sources and types of residual model error from an aggregation of the competition’s top 50 solutions. This analysis reveals the limitations for machine learning using the standard model inputs of historical meter, weather, and basic building metadata. The errors are classified according to timeframe, behavior, magnitude, and incidence in single buildings or across a campus. The results show machine learning models have errors within a range of acceptability (RMSLEscaled = < 0.1) on 79.1% of the test data. Lower magnitude (in-range) model errors (0.1 < RMSLEscaled = < 0.3) occur in 16.1% of the test data. These errors could be remedied using innovative training data from onsite and Web-based sources. Higher magnitude (out-of-range) errors (RMSLEscaled > 0.3) occur in 4.8% of the test data and are unlikely to be accurately predicted.

Machine learning (ML) teams often work on a project just to realize the performance of the model is not good enough. Indeed, the success of ML-enabled systems involves aligning data with business problems, translating them into ML tasks, experimenting with algorithms, evaluating models, capturing data from users, among others. Literature has shown that ML-enabled systems are rarely built based on precise specifications for such concerns, leading ML teams to become misaligned due to incorrect assumptions, which may affect the quality of such systems and overall project success. In order to help addressing this issue, this paper describes our work towards a perspective-based approach for specifying ML-enabled systems. The approach involves analyzing a set of 45 ML concerns grouped into five perspectives: objectives, user experience, infrastructure, model, and data. The main contribution of this paper is to provide two new artifacts that can be used to help specifying ML-enabled systems: (i) the perspective-based ML task and concern diagram and (ii) the perspective-based ML specification template.

Machine learning (ML) is being increasingly employed in dental research and application. We aimed to systematically compile studies using ML in dentistry and assess their methodological quality, including the risk of bias and reporting standards. We evaluated studies employing ML in dentistry published from 1 January 2015 to 31 May 2021 on MEDLINE, IEEE Xplore, and arXiv. We assessed publication trends and the distribution of ML tasks (classification, object detection, semantic segmentation, instance segmentation, and generation) in different clinical fields. We appraised the risk of bias and adherence to reporting standards, using the QUADAS-2 and TRIPOD checklists, respectively. Out of 183 identified studies, 168 were included, focusing on various ML tasks and employing a broad range of ML models, input data, data sources, strategies to generate reference tests, and performance metrics. Classification tasks were most common. Forty-two different metrics were used to evaluate model performances, with accuracy, sensitivity, precision, and intersection-over-union being the most common. We observed considerable risk of bias and moderate adherence to reporting standards which hampers replication of results. A minimum (core) set of outcome and outcome metrics is necessary to facilitate comparisons across studies.

In recent years, proficiency in data science and machine learning (ML) became one of the most requested skills for jobs in both industry and academy. Machine learning algorithms typically require large sets of data to train the models and extensive usage of computing resources, both for training and inference. Especially for deep learning algorithms, training performances can be dramatically improved by exploiting Graphical Processing Units (GPUs). The needed skill set for a data scientist is therefore extremely broad, and ranges from knowledge of ML models to distributed programming on heterogeneous resources. While most of the available training resources focus on ML algorithms and tools such as TensorFlow, we designed a course for doctoral students where model training is tightly coupled with underlying technologies that can be used to dynamically provision resources. Throughout the course, students have access to a dedicated cluster of computing nodes on local premises. A set of libraries and helper functions is provided to execute a parallelized ML task by automatically deploying a Spark driver and several Spark execution nodes as Docker containers. Task scheduling is managed by an orchestration layer (Kubernetes). This solution automates the delivery of the software stack required by a typical ML workflow and enables scalability by allowing the execution of ML tasks, including training, over commodity (i.e. CPUs) or high-performance (i.e. GPUs) resources distributed over different hosts across a network. The adaptation of the same model on OCCAM, the HPC facility at the University of Turin, is currently under development.

Patient safety research in human medicine has identified the causes and common types of medical error and subsequently informed the development of interventions which mitigate harm, such as the WHO's...