Patterns最新文献

Standard offline evaluations for language models fail to capture how these models actually behave in practice, where personalization fundamentally alters model behavior. In this work, we provide empirical evidence showcasing this phenomenon by comparing offline evaluations to field evaluations conducted by having 800 real users of ChatGPT and Gemini pose benchmark and other questions to their chat interfaces.

DNA barcode networks are the basis of sequencing-based microscopy, an emerging family of chemical imaging methods aiming to reconstruct spatial information, without optics, using sequencing technology. These methods capture microscopic spatial information by forming networks composed of many local chemical interactions, each marked by a unique, DNA-based barcode. However, the fundamental laws governing such networks are not yet understood, and spatial barcode networks are influenced by structural distortions such as false or shortcut edges. Current methods lack ground-truth-free tools to validate spatial quality, and we address this with a framework for topology-based quality control. We define a fundamental feature of spatial networks, spatial coherence, which quantifies geometric self-consistency in a network. By formalizing this relationship into quantitative metrics adapted from classical geometric rules, we could quantify spatial distortions by using only network data and show how these can be used as an optimization criterion to iteratively improve spatial reconstruction.

Despite moving our eyes from one location to another, our perception of the world is stable-an aspect thought to rely on predictive computations that use efference copies to predict the upcoming foveal input. Are these complex computations and required connectivity scaffolds genetically encoded, or could they emerge from simpler principles? Here, we consider the organism's limited energy budget as a potential origin. We expose a recurrent neural network to sequences of fixation patches and saccadic efference copies, training the model to minimize energy consumption (preactivation). We show that targeted inhibitory predictive remapping emerges from this energy-efficiency optimization alone. Furthermore, this computation relies on the model's learned ability to re-code egocentric eye coordinates into an allocentric (image-centric) reference frame. Together, our findings suggest that both allocentric coding and predictive remapping can emerge from energy-efficiency constraints, demonstrating how complex neural computations can arise from simple physical principles.

In this People of Data, we asked five researchers, including three members of the journal's advisory board, whether they feel AI technologies are currently overhyped. Their responses reveal both optimism about the future impact of these technologies and serious concerns about overblown expectations and uncritical applications.

Nelson et al. developed LA⁴SR, a large-scale language-model framework to classify translated open reading frames (ORFeomes) across ten algal phyla. LA⁴SR achieves near-complete protein classification coverage even for algal dark proteomes, where alignment tools often return "no hit." The framework provides a leap forward for high-throughput protein classification in algae and could be a promising tool for better classifying proteins from many organisms.

Visual information is central to human perception, and how we represent images critically shapes downstream analysis. While recent years have witnessed remarkable advances in deep learning for image processing, Valous et al. now introduce a non-data-driven framework grounded in hypercomplex algebras for natural and biomedical image processing. This plug-and-play approach operates without training data and shows its effect across tasks such as re-colorization, de-colorization, contrast enhancement, re-staining, and integration into machine-learning pipelines.

Rueda et al. argue that the concept of dignity is problematic for AI ethics due to its complexity, ambiguity, and biased usage. While agreeing on many points, we propose that adding the necessary precision to the use of the term is neither difficult nor onerous. Further, doing so may allow understanding of the factors that promote dignity affirmation, match the multifaceted nature of AI systems themselves, and promote pragmatically better design outcomes than will be likely if the idea is avoided in AI ethics discussions.

This study investigates the adoption of the FAIR (Findable, Accessible, Interoperable, and Reusable) principles by 14 environmental research infrastructures (RIs) operating at the Italian level. Through a three-step process (surveys, interviews, and a resource analysis), we explore the diverse FAIR practices adopted across four environmental subdomains, namely atmosphere, marine, biosphere, and geosphere. The findings reveal significant heterogeneity in the implemented practices, with ongoing efforts to converge on common strategies, particularly in the marine subdomain. Serving as a stepping stone toward more coordinated FAIR implementations, the analysis herein provides a solid foundation for monitoring future progress regarding the adoption of FAIR practices across environmental RIs within and beyond Italy and Europe.

Three-factor learning rules in spiking neural networks (SNNs) have emerged as a crucial extension of traditional Hebbian learning and spike-timing-dependent plasticity (STDP), incorporating neuromodulatory signals to improve adaptation and learning efficiency. These mechanisms enhance biological plausibility and facilitate improved credit assignment in artificial neural systems. This paper considers this topic from a machine learning perspective, providing an overview of recent advances in three-factor learning and discussing theoretical foundations, algorithmic implementations, and their relevance to reinforcement learning and neuromorphic computing. In addition, we explore interdisciplinary approaches, scalability challenges, and potential applications in robotics, cognitive modeling, and artificial intelligence (AI) systems. Finally, we highlight key research gaps and propose future directions for bridging the gap between neuroscience and AI.

Ranging from health to cybersecurity, real-world data are heavily imbalanced. Handling imbalance is among the formidable challenges of machine learning (ML), as it deteriorates ML's performance, yielding biased results toward majority classes. However, finding an adequate measure to assess the impact of data imbalance is a field of research by itself. Following a review of the available imbalance measures, we introduce the status of imbalance (SIMBA), which considers data distribution and overlap, both of which are crucial to assess the impact of imbalance. SIMBA is benchmarked against seven imbalance measures on five ML models, 428 synthetic and 70 non-synthetic datasets from various domains. Resulting correlation coefficients between imbalance measures and classification performance and an analysis with 20 complexity measures prove that SIMBA consistently outperforms other measures. Overall, SIMBA accurately quantifies multiclass data imbalance and may help alleviate ML data imbalance challenges in the future.