Transformers Pretrained on Procedural Data Contain Modular Structures for Algorithmic Reasoning

Pretraining on large, semantically rich datasets is key for developing language models. Surprisingly, recent studies have shown that even synthetic data, generated procedurally through simple semantic-free algorithms, can yield some of the same benefits as natural language pretraining. It is unclear what specific capabilities such simple synthetic data instils in a model, where these capabilities reside in the architecture, and how they manifest within its weights. In this short paper, we identify several beneficial forms of procedural data, together with specific algorithmic reasoning skills that improve in small transformers. Our core finding is that different procedural rules instil distinct but complementary inductive structures in the model. With extensive ablations and partial-transfer experiments, we discover that these structures reside in different parts of the model. Attention layers often carry the most transferable information, but some pretraining rules impart useful structure to MLP blocks instead. Most interestingly, the structures induced by multiple rules can be composed to jointly reinforce multiple capabilities. These results suggest an exciting possibility of disentangling the acquisition of knowledge from reasoning in language models, with the goal of improving their robustness and data efficiency.

Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild

Neural architectures tend to fit their data with relatively simple functions. This "simplicity bias" is widely regarded as key to their success. This paper explores the limits of this principle. Building on recent findings that the simplicity bias stems from ReLU activations [96], we introduce a method to meta-learn new activation functions and inductive biases better suited to specific tasks. Findings: We identify multiple tasks where the simplicity bias is inadequate and ReLUs suboptimal. In these cases, we learn new activation functions that perform better by inducing a prior of higher complexity. Interestingly, these cases correspond to domains where neural networks have historically struggled: tabular data, regression tasks, cases of shortcut learning, and algorithmic grokking tasks. In comparison, the simplicity bias induced by ReLUs proves adequate on image tasks where the best learned activations are nearly identical to ReLUs and GeLUs. Implications: Contrary to popular belief, the simplicity bias of ReLU networks is not universally useful. It is near-optimal for image classification, but other inductive biases are sometimes preferable. We showed that activation functions can control these inductive biases, but future tailored architectures might provide further benefits. Advances are still needed to characterize a model's inductive biases beyond "complexity", and their adequacy with the data.

OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable?

Out-of-distribution (OOD) generalization is challenging because distribution shifts come in many forms. A multitude of learning algorithms exist and each can improve performance in specific OOD situations. We posit that much of the challenge of OOD generalization lies in choosing the right algorithm for the right dataset. However, such algorithm selection is often elusive under complex real-world shifts. In this work, we formalize the task of algorithm selection for OOD generalization and investigate whether it could be approached by learning. We propose a solution, dubbed OOD-Chameleon that treats the task as a supervised classification over candidate algorithms. We construct a dataset of datasets to learn from, which represents diverse types, magnitudes and combinations of shifts (covariate shift, label shift, spurious correlations). We train the model to predict the relative performance of algorithms given a dataset's characteristics. This enables a priori selection of the best learning strategy, i.e. without training various models as needed with traditional model selection. Our experiments show that the adaptive selection outperforms any individual algorithm and simple selection heuristics, on unseen datasets of controllable and realistic image data. Inspecting the model shows that it learns non-trivial data/algorithms interactions, and reveals the conditions for any one algorithm to surpass another. This opens new avenues for (1) enhancing OOD generalization with existing algorithms instead of designing new ones, and (2) gaining insights into the applicability of existing algorithms with respect to datasets' properties.

Unraveling the Key Components of OOD Generalization via Diversification

Real-world datasets may contain multiple features that explain the training data equally well, i.e., learning any of them would lead to correct predictions on the training data. However, many of them can be spurious, i.e., lose their predictive power under a distribution shift and fail to generalize to out-of-distribution (OOD) data. Recently developed ``diversification'' methods approach this problem by finding multiple diverse hypotheses that rely on different features. This paper aims to study this class of methods and identify the key components contributing to their OOD generalization abilities. We show that (1) diversification methods are highly sensitive to the distribution of the unlabeled data used for diversification and can underperform significantly when away from a method-specific sweet spot. (2) Diversification alone is insufficient for OOD generalization. The choice of the used learning algorithm, e.g., the model's architecture and pretraining, is crucial, and using the second-best choice leads to an up to 20% absolute drop in accuracy.(3) The optimal choice of learning algorithm depends on the unlabeled data, and vice versa.Finally, we show that the above pitfalls cannot be alleviated by increasing the number of diverse hypotheses, allegedly the major feature of diversification methods. These findings provide a clearer understanding of the critical design factors influencing the OOD generalization of diversification methods. They can guide practitioners in how to use the existing methods best and guide researchers in developing new, better ones.

Test-Time Robust Personalization for Federated Learning

Federated Learning (FL) is a machine learning paradigm where many clients collaboratively learn a shared global model with decentralized training data. Personalization on FL model additionally adapts the global model to different clients, achieving promising results on consistent local training & test distributions. However, for real-world personalized FL applications, it is crucial to go one step further: robustifying FL models under evolving local test set during deployment, where various types of distribution shifts can arise. In this work, we identify the pitfalls of existing works under test-time distribution shifts and propose a novel test-time robust personalization method, namely Federated Test-time Head Ensemble plus tuning (FedTHE+). We illustrate the advancement of FedTHE+ (and its degraded computationally efficient variant FedTHE) over strong competitors, for training various neural architectures (CNN, ResNet, and Transformer) on CIFAR10 and ImageNet and evaluating on diverse test distributions. Along with this, we build a benchmark for assessing performance and robustness of personalized FL methods during deployment.