Timezone: »

Data Distributional Properties Drive Emergent In-Context Learning in Transformers
Stephanie Chan · Adam Santoro · Andrew Lampinen · Jane Wang · Aaditya Singh · Pierre Richemond · James McClelland · Felix Hill

Wed Nov 30 02:00 PM -- 04:00 PM (PST) @ Hall J #524

Large transformer-based models are able to perform in-context few-shot learning, without being explicitly trained for it. This observation raises the question: what aspects of the training regime lead to this emergent behavior? Here, we show that this behavior is driven by the distributions of the training data itself. In-context learning emerges when the training data exhibits particular distributional properties such as burstiness (items appear in clusters rather than being uniformly distributed over time) and having a large number of rarely occurring classes. In-context learning also emerges more strongly when item meanings or interpretations are dynamic rather than fixed. These properties are exemplified by natural language, but are also inherent to naturalistic data in a wide range of other domains. They also depart significantly from the uniform, i.i.d. training distributions typically used for standard supervised learning. In our initial experiments, we found that in-context learning traded off against more conventional weight-based learning, and models were unable to achieve both simultaneously. However, our later experiments uncovered that the two modes of learning could co-exist in a single model when it was trained on data following a skewed Zipfian distribution -- another common property of naturalistic data, including language. In further experiments, we found that naturalistic data distributions were only able to elicit in-context learning in transformers, and not in recurrent models. Our findings indicate how the transformer architecture works together with particular properties of the training data to drive the intriguing emergent in-context learning behaviour of large language models, and indicate how future work might encourage both in-context and in-weights learning in domains beyond language.

Author Information

Stephanie Chan (DeepMind)
Adam Santoro (DeepMind)
Andrew Lampinen (DeepMind)
Jane Wang (DeepMind)

Jane Wang is a research scientist at DeepMind on the neuroscience team, working on meta-reinforcement learning and neuroscience-inspired artificial agents. Her background is in physics, complex systems, and computational and cognitive neuroscience.

Aaditya Singh (University College London, University of London)
Pierre Richemond (DeepMind, Imperial College)
James McClelland (Stanford University)
Felix Hill (Deepmind)

More from the Same Authors