In this paper, we study the generalization properties of Model-Agnostic MetaLearning (MAML) algorithms for supervised learning problems. We focus on the setting in which we train the MAML model over mtasks, each with ndata points, and characterize its generalization error from two points of view: First, we assume the new task at test time is one of the training tasks, and we show that, for strongly convex objective functions, the expected excess population loss is bounded by O(1/mn). Second, we consider the MAML algorithm's generalization to an unseen task and show that the resulting generalization error depends on the total variation distance between the underlying distributions of the new task and the tasks observed during the training process. Our proof techniques rely on the connections between algorithmic stability and generalization bounds of algorithms. In particular, we propose a new definition of stability for meta-learning algorithms, which allows us to capture the role of both the number of tasks mand number of samples per task non the generalization error of MAML.

The ViT model we use in this work is based on a standard Vision Transformer [7] model scaled to577 nearly 3 billion parameters, using a patch size of 14, 16 heads, 64 blocks, an MLP dimension of 8192578 and a hidden dimension of 2048. The model is defined and trained in Lingvo [32]; we additionally579 employ GSPMD [41] for training. The model is pre-trained on JFT-3B [35] using training settings580 that optimize for performance on JFT-3B rather than for fine-tuning on ImageNet; notably, we do not581 use the training recipe that helps few-shot transfer performance [44]. BReview tools586 We include screenshots of the reviewing tools we built to analyze model mistakes. Figure 3 shows587 the UI for reviewing model predictions and Figure 4 shows the UI that displays the labeling guide588 and slide bar to browse images for a particular class.

Image classification accuracy on the ImageNet dataset has been a barometer for progress in computer vision over the last decade. Several recent papers have questioned the degree to which the benchmark remains useful to the community [33, 3, 31, 42, 36], yet innovations continue to contribute gains to performance, with today's largest models achieving 90%+ top-1 accuracy. To help contextualize progress on ImageNet and provide a more meaningful evaluation for today's stateof-the-art models, we manually review and categorize every remaining mistake that a few top models make and provide insights into the long-tail of errors on one of the most benchmarked datasets in computer vision. We focus on the multi-label subset evaluation of ImageNet, where today's best models achieve upwards of 97% top-1 accuracy. Our analysis reveals that nearly half of the supposed mistakes are not mistakes at all, and we uncover new valid multi-labels, demonstrating that, without careful review, we are significantly underestimating the performance of these models. On the other hand, we also find that today's best models still make a significant number of mistakes (40%) that are obviously wrong to human reviewers. To calibrate future progress on ImageNet, we provide an updated multilabel evaluation set, and we curate ImageNet-Major1: a 68-example "major error" slice of the obvious mistakes made by today's top models--a slice where models should achieve near perfection, but today are far from doing so.

In addition to CYCLIP described in 2, we train two more instantiations of it by keeping either of the two consistency regularizers active in the loss objective (Eq. The instantiation trained by setting λ1 = 0and λ2 = 0.5is termed as C-CYCLIP as only cross-modal consistency regularizer term is added to the loss objective. Similarly, we get I-CYCLIP where only in-modal consistency regularizer is added to the loss by setting λ1 = 0.5 and λ2 = 0. We evaluate C-CYCLIP and I-CYCLIP on most of the experiments discussed in the main text to understand their zero-shot transfer ability on standard datasets and robustness to natural distribution shifts. A.1 Zero-shot Transfer Table 7 presents our results of the zero-shot transfer experiment described in 3.1. We find that CYCLIP outperforms its sub-variants and the CLIP model on the ImageNet1K dataset.

Stochastic gradient descent (SGD) exhibits strong algorithmic regularization effects in practice, which has been hypothesized to play an important role in the generalization of modern machine learning approaches. In this work, we seek to understand these issues in the simpler setting of linear regression (including both underparameterized and overparameterized regimes), where our goal is to make sharp instance-based comparisons of the implicit regularization afforded by (unregularized) average SGD with the explicit regularization of ridge regression. For a broad class of least squares problem instances (that are natural in high-dimensional settings), we show: (1) for every problem instance and for every ridge parameter, (unregularized) SGD, when provided with logarithmically more samples than that provided to the ridge algorithm, generalizes no worse than the ridge solution (provided SGD uses a tuned constant stepsize); (2) conversely, there exist instances (in this wide problem class) where optimally-tuned ridge regression requires quadratically more samples than SGD in order to have the same generalization performance. Taken together, our results show that, up to the logarithmic factors, the generalization performance of SGD is always no worse than that of ridge regression in a wide range of overparameterized problems, and, in fact, could be much better for some problem instances. More generally, our results show how algorithmic regularization has important consequences even in simpler (overparameterized) convex settings.