The brain is much better than current artificial models at unsupervised learning.

Judge as of 2015-02-18. True: research and expert assessments from that time show that the brain learns useful representations from unlabelled sensory data far more effectively than the artificial unsupervised learning methods of 2015, and that those methods were broadly considered to lag well behind what biological learning does. Mostly true: the brain was clearly better, but 'much better' overstates the gap on some specific measures. Mixed: the brain had the advantage on some aspects of unsupervised learning, but artificial models matched or exceeded it on roughly as many others. Mostly false: artificial models were comparable on most measures and the brain led only narrowly. False: the evidence shows that artificial models of the time matched or exceeded the brain at unsupervised learning. If the evidence is consensus without direct measurement, that can support true or mostly true; only evidence contradicting the claim supports false.

Set before the research began.

1. E1 Perception Science in the Age of Deep Neural Networks PubMed Central (article from Frontiers in Perception Science) · 2 February 2017

   “In contrast, humans can often learn without explicit supervision or “labels.” Unsupervised learning methods do exist for artificial neural networks, but they often give rise to a feature space that is insufficiently powerful and needs to be complemented by supervised fine-tuning in order to allow, for example, for accurate object recognition”

2. E2 Perception Science in the Age of Deep Neural Networks PubMed Central (article from Frontiers in Perception Science) · 2 February 2017

   “Some of these methods can also learn features and representations from one or just a few exemplars, a form of “one-shot learning” on par with human capabilities (Anselmi et al., 2013; Rezende et al., 2016; Santoro et al., 2016).”

E1 is an independent perspective from 2017, after the judge-as-of date. It says humans can often learn without labels, while artificial unsupervised methods of that period often produced feature spaces too weak to use without supervised fine-tuning. That is consensus-style support for the brain having a clear advantage, which the criteria accept for true or mostly true. E2 notes that some methods reached human-level one-shot learning on certain tasks, but most of the work it cites dates from 2016, and one-shot learning is not strictly unsupervised learning. Neither source gives a direct measurement, so 'much better' is supported only by expert assessment, not quantified; the verdict is mostly true rather than true.