Intelligence artificielle et données

This document provides an in-depth critique of MapReduce's limitations, emphasizing its inefficiencies in complex, multi-step computations. It introduces Apache Spark as a more advanced, memory-optimized solution with higher performance, scalability, and interactive APIs for diverse programming languages. Key Spark components like RDDs, DataFrames, Datasets, and its ecosystem libraries (e.g., MLlib, GraphX) are detailed alongside their integration capabilities. The document further outlines Spark's architecture, including its execution model and resource management, while contrasting its pe...

This paper investigates the properties of deep neural networks that contribute to their state-of-the-art performance on speech and visual recognition tasks. It highlights two counter-intuitive findings: the indistinguishability of individual high-level units and the effects of small perturbations on network predictions. The authors argue that it is the space of activations that retains semantic information rather than individual units.

This document explores the challenges and implications of providing explanations for automated decisions in the context of the GDPR. It discusses counterfactual explanations as a means to understand, contest, and alter decisions made by algorithms. The paper also highlights legal and technical barriers to implementing the right to explanation.

This paper discusses the challenges of evaluating defenses against adversarial examples in machine learning. The authors highlight common pitfalls in existing evaluations and propose a checklist to guide researchers in conducting more robust evaluations. The work aims to improve the understanding and completeness of defense evaluations among both developers and reviewers.

This paper discusses the vulnerability of neural networks to adversarial examples, which are inputs modified to mislead models. It asserts that this vulnerability stems from the linear nature of neural networks rather than previously suggested factors. The authors propose a method for generating adversarial examples that aids in enhancing model training.

The document explores various techniques for performing incremental ETL (Extract, Transform, Load) processes. It outlines three architectural approaches: single-stage ETL for small-scale BI solutions with minimal data sources, two-stage ETL for moderately complex or larger data volumes that a single stage cannot accommodate, and three-stage ETL for even more complex architectures that optimize data flow and reduce load on source systems. The content is designed to guide data engineers in building efficient ETL pipelines tailored to specific use cases.

The document explores incremental ETL techniques with three specific architectures: one-stage, two-stage, and three-stage ETL processes. A one-stage ETL is suitable for small-scale BI systems with limited data sources and simple requirements. Two-stage ETL addresses complexities and larger data volumes, optimizing performance. Lastly, a three-stage ETL architecture reduces extraction overhead and minimizes source system load, enhancing efficiency in data flow management.

This article discusses the limitations of deep learning in AI, particularly its lack of interpretability and ability to answer logical questions. It introduces the concept of Neurosymbolic AI, which combines aspects of symbolic AI with deep learning to provide more explainable and interpretable AI systems. The text highlights the historical context of AI development and contrasts traditional symbolic AI with modern deep learning approaches.

This document focuses on implementing incremental ETL processes to extract modified data from source systems. Techniques such as using primary keys with checksums, datetime columns, Change Data Capture (CDC), and Change Tracking in SQL Server are discussed in detail. It includes step-by-step processes for configuration and extraction using CDC and Change Tracking. Additionally, it provides guidance on methods to identify record changes, whether inserted, updated, or deleted, and how to handle logical and physical deletions efficiently.

The document discusses designing an effective incremental ETL process for refreshing data warehouses with new or modified data from source systems. It outlines key considerations for planning such processes, including managing data modifications, understanding dimension keys, and handling updates to dimension and fact records. The document also details common ETL data flow architectures—single-stage, two-stage, and three-stage—and provides guidelines for planning data extraction windows and transformations to enhance performance and data consistency.

This paper introduces the CLEVRER dataset designed to evaluate computational models on temporal and causal reasoning tasks in videos. It emphasizes the importance of understanding the underlying causal structures in video data, contrasting with the predominant focus on visual pattern recognition. The study identifies the limitations of current models in handling causal reasoning tasks, proposing a new benchmark for future research.