Transcendence: Generative Models Can Outperform The Experts That Train Them

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Zhang, Edwin, Zhu, Vincent, Saphra, Naomi, Kleiman, Anat, Edelman, Benjamin L., Tambe, Milind, Kakade, Sham M., Malach, Eran
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909346481307648
author Zhang, Edwin
Zhu, Vincent
Saphra, Naomi
Kleiman, Anat
Edelman, Benjamin L.
Tambe, Milind
Kakade, Sham M.
Malach, Eran
author_facet Zhang, Edwin
Zhu, Vincent
Saphra, Naomi
Kleiman, Anat
Edelman, Benjamin L.
Tambe, Milind
Kakade, Sham M.
Malach, Eran
contents Generative models are trained with the simple objective of imitating the conditional probability distribution induced by the data they are trained on. Therefore, when trained on data generated by humans, we may not expect the artificial model to outperform the humans on their original objectives. In this work, we study the phenomenon of transcendence: when a generative model achieves capabilities that surpass the abilities of the experts generating its data. We demonstrate transcendence by training an autoregressive transformer to play chess from game transcripts, and show that the trained model can sometimes achieve better performance than all players in the dataset. We theoretically prove that transcendence can be enabled by low-temperature sampling, and rigorously assess this claim experimentally. Finally, we discuss other sources of transcendence, laying the groundwork for future investigation of this phenomenon in a broader setting.
format Preprint
id arxiv_https___arxiv_org_abs_2406_11741
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Transcendence: Generative Models Can Outperform The Experts That Train Them
Zhang, Edwin
Zhu, Vincent
Saphra, Naomi
Kleiman, Anat
Edelman, Benjamin L.
Tambe, Milind
Kakade, Sham M.
Malach, Eran
Machine Learning
Artificial Intelligence
Generative models are trained with the simple objective of imitating the conditional probability distribution induced by the data they are trained on. Therefore, when trained on data generated by humans, we may not expect the artificial model to outperform the humans on their original objectives. In this work, we study the phenomenon of transcendence: when a generative model achieves capabilities that surpass the abilities of the experts generating its data. We demonstrate transcendence by training an autoregressive transformer to play chess from game transcripts, and show that the trained model can sometimes achieve better performance than all players in the dataset. We theoretically prove that transcendence can be enabled by low-temperature sampling, and rigorously assess this claim experimentally. Finally, we discuss other sources of transcendence, laying the groundwork for future investigation of this phenomenon in a broader setting.
title Transcendence: Generative Models Can Outperform The Experts That Train Them
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2406.11741