AI Insight
This study investigates how Mixture-of-Experts (MoE) language models respond to repeated training data compared to traditional dense models. Researchers found that MoE models, ranging from 80M to 1B active parameters, overfit significantly more when training data is repeated, with degradation beginning at just 4x repetition versus 8x for dense models. While regularization techniques like dropout can partially mitigate this overfitting, the analysis reveals that MoE models suffer from early routing stabilization and excessive expert specialization when exposed to repeated data.
Why it matters
As available human-written text for training AI models becomes exhausted, understanding how different architectures handle data repetition is critical for developing efficient language models. These findings suggest that the increasingly popular MoE architectures may require different training strategies or stronger regularization when working with limited or repeated datasets.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Abstract: As the supply of human-written text is exhausted, it has become standard practice to repeat language model training data. Prior work has studied data repetition for densely activated Transformers, but the effects of data repetition remains largely unexplored for recently dominant sparse architectures such as Mixture-of-Experts (MoE), despite their increased compute efficiency. We vary data repetition rates across single- and multi-domain data mixes, and across MoE settings, including expert count and granularity. We consistently find, for models ranging from 80M to 1B active (8.5B total) parameters, that MoEs degrade more rapidly under data repetition. This effect increases with sparsity, dictated by total rather than active parameters. While 80M dense models can repeat data over 8x with minimal degradation, MoEs instead begin to suffer at 4x, and deteriorate rapidly, ceding their performance benefits in all-unique data settings to underperform dense models after 32x. We experiment with existing regularization methods as a potential remedy. We find that some methods, such as dropout, can mitigate overfitting. In particular, with strong masking-based regularization, MoEs are able to outperform dense models even when data is repeated more than 64 times. However, no method fully matches the performance of all-unique training data. Finally, we analyze internal mechanisms correlated with MoE overfitting in high repetition regimes, and find that MoE routing universally stabilizes early in training, and that expert specialization correlates with overfitting to repeated data. In sum, our work addresses the adverse interactions between sparsity and data repetition: we present evidence for the core mechanisms of overfitting and its potential remediation, and suggest promising avenues for future methods to reduce over-specialization in model parameters by disrupting memorization patterns.
Source: Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data