arXiv subjects

Sylvain Peyronnet

Publications and source records attributed to Sylvain Peyronnet.

6 recordsLinked to original sources

arXiv2026

When Search Eats the Web: A Model of Corpus Erosion under Generative Extraction

Generative search engines (GSEs) answer user queries directly from crawled web content. The capture of value from the corpus without a visit returned to the source (we call this capture extraction) diverts the traffic that finances content production. In response, publishers may restrict crawler access to their websites. In this paper, we model the crawlable corpus as a common-pool resource: the crawlable commons. It is described by three quantities: volume, average quality, and lifetime. Under two types of responses of publishers we prove that extraction degrades all three at once: publishers opt out, renewal loses its funding, and content becomes more perishable. After a given erosion threshold, the corpus goes extinct. A myopic GSE can cross this threshold, a long-run oriented GSE stays below it. We extend our model to several competing engines and prove, under a concavity condition on the steady-state value of the commons, that the symmetric equilibrium extraction rate is nondecreasing in their number and converges to the threshold. Adding users who strictly prefer direct answers, the assumption most favorable to extraction, we prove that the socially optimal extraction rate lies strictly below the erosion threshold, and no higher than the single engine's sustainable optimum. Finally, we discuss seven survival mechanisms.

cs.GT↗

arXiv2020

Graphcore C2 Card performance for image-based deep learning application: A Report

Recently, Graphcore has introduced an IPU Processor for accelerating machine learning applications. The architecture of the processor has been designed to achieve state of the art performance on current machine intelligence models for both training and inference. In this paper, we report on a benchmark in which we have evaluated the performance of IPU processors on deep neural networks for inference. We focus on deep vision models such as ResNeXt. We report the observed latency, throughput and energy efficiency.

cs.CV↗

arXiv2019

Image search using multilingual texts: a cross-modal learning approach between image and text

Multilingual (or cross-lingual) embeddings represent several languages in a unique vector space. Using a common embedding space enables for a shared semantic between words from different languages. In this paper, we propose to embed images and texts into a unique distributional vector space, enabling to search images by using text queries expressing information needs related to the (visual) content of images, as well as using image similarity. Our framework forces the representation of an image to be similar to the representation of the text that describes it. Moreover, by using multilingual embeddings we ensure that words from two different languages have close descriptors and thus are attached to similar images. We provide experimental evidence of the efficiency of our approach by experimenting it on two datasets: Common Objects in COntext (COCO) [19] and Multi30K [7].

cs.CV↗

arXiv2016

On the evaluation potential of quality functions in community detection for different contexts

Due to nowadays networks' sizes, the evaluation of a community detection algorithm can only be done using quality functions. These functions measure different networks/graphs structural properties, each of them corresponding to a different definition of a community. Since there exists many definitions for a community, choosing a quality function may be a difficult task, even if the networks' statistics/origins can give some clues about which one to choose. In this paper, we apply a general methodology to identify different contexts, i.e. groups of graphs where the quality functions behave similarly. In these contexts we identify the best quality functions, i.e. quality functions whose results are consistent with expectations from real life applications.

cs.SI↗

arXiv2015

Efficient filtering of adult content using textual information

Nowadays adult content represents a non negligible proportion of the Web content. It is of the utmost importance to protect children from this content. Search engines, as an entry point for Web navigation are ideally placed to deal with this issue. In this paper, we propose a method that builds a safe index i.e. adult-content free for search engines. This method is based on a filter that uses only textual information from the web page and the associated URL.

cs.IR↗

arXiv2006

Uniform Random Sampling of Traces in Very Large Models

This paper presents some first results on how to perform uniform random walks (where every trace has the same probability to occur) in very large models. The models considered here are described in a succinct way as a set of communicating reactive modules. The method relies upon techniques for counting and drawing uniformly at random words in regular languages. Each module is considered as an automaton defining such a language. It is shown how it is possible to combine local uniform drawings of traces, and to obtain some global uniform random sampling, without construction of the global model.

cs.LO↗