Autor(es)

Willinger, Walter ; Ferreira, Ronaldo A. ; Gupta, Arpit ; Beltiukov, Roman (Sylee) ; Guthula, Satyandra ; Granville, Lisandro Z. ; Jacobs, Arthur S.

• Tipo de documento: Artigo Científico
• Fonte: ACM SIGCOMM COMPUTER COMMUNICATION REVIEW; v. 55, n. 1, p. 6-pg., 2025-01-01.

Resumo

The purpose of this editorial note is to raise awareness about a deeply concerning and yet much-overlooked development in the use of Artificial Intelligence (AI) and Machine Learning (ML) for solving problems in science in general and in networking in particular. To put it simply, in today's age of AI/ML, the much-publicized and well-documented "reproducibility crisis" in science is further compounded by an inconspicuous and rarely mentioned "credibility crisis." More to the point, by focusing on the area of networking research, we provide evidence that among the already small number of reproducible scientific publications that describe AI/ML-based solutions, even fewer, and often none, describe trained AI/ML models that are "credible;" that is, can be trusted to not only perform well in their original training domain but also in new and untested environments. We elaborate on the root cause of this credibility crisis, discuss why the credibility of AI/ML models is of paramount importance for their successful use in practice, and put forward an aggressive but imminently practical proposal for addressing this crisis head-on so as to pave the way for a future where networking research can reap the full benefits of AI/ML. (AU)

• Processo FAPESP: 20/05183-0 - SkyNet: rumo à era dos planos de dados inteligentes
• Beneficiário: Luciano Paschoal Gaspary
• Modalidade de apoio: Auxílio à Pesquisa - Temático
• Processo FAPESP: 23/00811-0 - EcoSustain - Ciência de Dados e Computação para o Meio Ambiente
• Beneficiário: Antonio Jorge Gomes Abelém
• Modalidade de apoio: Auxílio à Pesquisa - Temático