Alexis Huet

Alexis Huet, Jose Manuel Navarro, Dario Rossi

In recent years, specific evaluation metrics for time series anomaly detection algorithms have been developed to handle the limitations of the classical precision and recall. However, such metrics are heuristically built as an aggregate of multiple desirable aspects, introduce parameters and wipe out the interpretability of the output. In this article, we first highlight the limitations of the classical precision/recall, as well as the main issues of the recent event-based metrics -- for instance, we show that an adversary algorithm can reach high precision and recall on almost any dataset under weak assumption. To cope with the above problems, we propose a theoretically grounded, robust, parameter-free and interpretable extension to precision/recall metrics, based on the concept of ``affiliation'' between the ground truth and the prediction sets. Our metrics leverage measures of duration between ground truth and predictions, and have thus an intuitive interpretation. By further comparison against random sampling, we obtain a normalized precision/recall, quantifying how much a given set of results is better than a random baseline prediction. By construction, our approach keeps the evaluation local regarding ground truth events, enabling fine-grained visualization and interpretation of algorithmic results. We compare our proposal against various public time series anomaly detection datasets, algorithms and metrics. We further derive theoretical properties of the affiliation metrics that give explicit expectations about their behavior and ensure robustness against adversary strategies.

Jose Manuel Navarro, Alexis Huet, Dario Rossi

Anomaly detection research works generally propose algorithms or end-to-end systems that are designed to automatically discover outliers in a dataset or a stream. While literature abounds concerning algorithms or the definition of metrics for better evaluation, the quality of the ground truth against which they are evaluated is seldom questioned. In this paper, we present a systematic analysis of available public (and additionally our private) ground truth for anomaly detection in the context of network environments, where data is intrinsically temporal, multivariate and, in particular, exhibits spatial properties, which, to the best of our knowledge, we are the first to explore. Our analysis reveals that, while anomalies are, by definition, temporally rare events, their spatial characterization clearly shows some type of anomalies are significantly more popular than others. We find that simple clustering can reduce the need for human labeling by a factor of 2x-10x, that we are first to quantitatively analyze in the wild.

Alexis Huet, Zied Ben Houidi, Dario Rossi

Episodic memory -- the ability to recall specific events grounded in time and space -- is a cornerstone of human cognition, enabling not only coherent storytelling, but also planning and decision-making. Despite their remarkable capabilities, Large Language Models (LLMs) lack a robust mechanism for episodic memory: we argue that integrating episodic memory capabilities into LLM is essential for advancing AI towards human-like cognition, increasing their potential to reason consistently and ground their output in real-world episodic events, hence avoiding confabulations. To address this challenge, we introduce a comprehensive framework to model and evaluate LLM episodic memory capabilities. Drawing inspiration from cognitive science, we develop a structured approach to represent episodic events, encapsulating temporal and spatial contexts, involved entities, and detailed descriptions. We synthesize a unique episodic memory benchmark, free from contamination, and release open source code and datasets to assess LLM performance across various recall and episodic reasoning tasks. Our evaluation of state-of-the-art models, including GPT-4 and Claude variants, Llama 3.1, and o1-mini, reveals that even the most advanced LLMs struggle with episodic memory tasks, particularly when dealing with multiple related events or complex spatio-temporal relationships -- even in contexts as short as 10k-100k tokens.

Simone Clemente, Zied Ben Houidi, Alexis Huet et al.

Through systematic empirical investigation, we uncover a fundamental and concerning property of Large Language Models: while they can safely learn facts that don't contradict their knowledge, attempting to update facts with contradictory information triggers catastrophic corruption of unrelated knowledge. Unlike humans, who naturally resist contradictory information, these models indiscriminately accept contradictions, leading to devastating interference, destroying up to 80% of unrelated knowledge even when learning as few as 10-100 contradicting facts. To understand whether this interference could be mitigated through selective plasticity, we experiment with targeted network updates, distinguishing between previously used (stubborn) and rarely used (plastic) neurons. We uncover another asymmetry: while sparing frequently-used neurons significantly improves retention of existing knowledge for non-contradictory updates (98% vs 93% with standard updates), contradictory updates trigger catastrophic interference regardless of targeting strategy. This effect which persists across tested model scales (GPT-2 to GPT-J-6B), suggests a fundamental limitation in how neural networks handle contradictions. Finally, we demonstrate that contradictory information can be reliably detected (95%+ accuracy) using simple model features, offering a potential protective mechanism. These findings motivate new architectures that can, like humans, naturally resist contradictions rather than allowing destructive overwrites.

Jose M. Navarro, Alexis Huet, Dario Rossi

Network troubleshooting is still a heavily human-intensive process. To reduce the time spent by human operators in the diagnosis process, we present a system based on (i) unsupervised learning methods for detecting anomalies in the time domain, (ii) an attention mechanism to rank features in the feature space and finally (iii) an expert knowledge module able to seamlessly incorporate previously collected domain-knowledge. In this paper, we thoroughly evaluate the performance of the full system and of its individual building blocks: particularly, we consider (i) 10 anomaly detection algorithms as well as (ii) 10 attention mechanisms, that comprehensively represent the current state of the art in the respective fields. Leveraging a unique collection of expert-labeled datasets worth several months of real router telemetry data, we perform a thorough performance evaluation contrasting practical results in constrained stream-mode settings, with the results achievable by an ideal oracle in academic settings. Our experimental evaluation shows that (i) the proposed system is effective in achieving high levels of agreement with the expert, and (ii) that even a simple statistical approach is able to extract useful information from expert knowledge gained in past cases, significantly improving troubleshooting performance.