Ryan Lee, Jacob Biloki, Edward J. Hu, Jonathan May
原文摘要
Looped language models repeat a set of transformer layers through depth, reducing memory costs and providing natural early-exit points at loop boundaries. However, looped models do not scale as favorably as standard transformers with unique layers. We compare standard and Mixture-of-Experts (MoE) transformers, with and without looping, and find two main results. First, we find Looped-MoE models scale better than the standard baseline while dense looped models do not. We trace this to routing divergence between loops: in Looped-MoE models, different experts are activated on each pass through the same shared layers, recovering expressivity without additional parameters. Our second finding is that looped models have better compute-quality trade-offs with early exits than standard models. Because each loop ends with the same layers that produce the final output, loop boundaries are superior exit points, as confirmed by earlier output convergence at these points. In sum, we provide a clear direction for scaling looped models: a Looped-MoE model with early exits can not only beat standard transformers at scale, but also enable significant memory and inference savings with minimal degradation in quality.
We examine the main claims of Zada, Benou, and Irani (2021), which states that in an image classification task with imbalanced training data, (i) using pure noise to augment minority-class images encourages generalization by improving minority-class accuracy. This method is paired with (ii) a new batch normalization layer that normalizes noise images using affine parameters learned from natural images, which improves the model's performance. Moreover, (iii) this improvement is robust to varying levels of data augmentation. Finally, the authors propose that (iv) adding pure noise images can improve classification even on balanced training data. We implemented the training pipeline from the description of the paper using PyTorch and integrated authors' code snippets for sampling pure noise images and batch normalizing noise and natural images separately. We reproduced the main claims that (i) oversampling with pure noise improves generalization by improving the minority-class accuracy, (ii) the proposed batch normalization (BN) method outperforms baselines, (iii) and this improvement is robust across data augmentations. Our results also support that (iv) adding pure noise images can improve classification on balanced training data. However, additional experiments suggest that the performance improvement from OPeN may be more orthogonal to the improvement caused by a bigger network or more complex data augmentation.