Abstract: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.
| Comments: | Code, models, and data at this https URL |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2406.11741 [cs.LG] |
| (or arXiv:2406.11741v4 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2406.11741 arXiv-issued DOI via DataCite |
Submission history
From: Eddie Zhang [view email]
[v1]
Mon, 17 Jun 2024 17:00:52 UTC (15,728 KB)
[v2]
Fri, 21 Jun 2024 23:45:55 UTC (15,105 KB)
[v3]
Fri, 28 Jun 2024 05:28:27 UTC (15,097 KB)
[v4]
Sat, 12 Oct 2024 18:46:20 UTC (17,333 KB)