Recently, there has been a surge in interest in developing optimization algorithms for overparameterized models as achieving generalization is believed to require algorithms with suitable biases. This interest centers on minimizing sharpness of the original loss function; the Sharpness-Aware Minimization (SAM) algorithm has proven effective. However, most literature only considers a few sharpness measures, such as the maximum eigenvalue or trace of the training loss Hessian, which may not yield meaningful insights for non-convex optimization scenarios like neural networks. Additionally, many sharpness measures are sensitive to parameter invariances in neural networks, magnifying significantly under rescaling parameters. Motivated by these challenges, we introduce a new class of sharpness measures in this paper, leading to new sharpness-aware objective functions. We prove that these measures are universally expressive, allowing any function of the training loss Hessian matrix to be represented by appropriate hyperparameters. Furthermore, we show that the proposed objective functions explicitly bias towards minimizing their corresponding sharpness measures, and how they allow meaningful applications to models with parameter invariances (such as scale-invariances). Finally, as instances of our proposed general framework, we present Frob-SAM and Det-SAM, which are specifically designed to minimize the Frobenius norm and the determinant of the Hessian of the training loss, respectively. We also demonstrate the advantages of our general framework through extensive experiments.

Sharpness-aware minimization (SAM) is a recently proposed method that minimizes the sharpness of the training loss of a neural network. While its generalization improvement is well-known and is the primary motivation, we uncover an additional intriguing effect of SAM: reduction of the feature rank which happens at different layers of a neural network. We show that this low-rank effect occurs very broadly: for different architectures such as fully-connected networks, convolutional networks, vision transformers and for different objectives such as regression, classification, language-image contrastive training. To better understand this phenomenon, we provide a mechanistic understanding of how low-rank features arise in a simple two-layer network. We observe that a significant number of activations gets entirely pruned by SAM which directly contributes to the rank reduction. We confirm this effect theoretically and check that it can also occur in deep networks, although the overall rank reduction mechanism can be more complex, especially for deep networks with pre-activation skip connections and self-attention layers.

Work in information retrieval has largely been centered around ranking and relevance: given a query, return some number of results ordered by relevance to the user. The problem of result list truncation, or where to truncate the ranked list of results, however, has received less attention despite being crucial in a variety of applications. Such truncation is a balancing act between the overall relevance, or usefulness of the results, with the user cost of processing more results. Result list truncation can be challenging because relevance scores are often not well-calibrated. This is particularly true in large-scale IR systems where documents and queries are embedded in the same metric space and a query's nearest document neighbors are returned during inference. Here, relevance is inversely proportional to the distance between the query and candidate document, but what distance constitutes relevance varies from query to query and changes dynamically as more documents are added to the index. In this work, we propose Surprise scoring, a statistical method that leverages the Generalized Pareto Distribution that arises in extreme value theory to produce interpretable and calibrated relevance scores at query time using nothing more than the ranked scores. We demonstrate its effectiveness on the result list truncation task across image, text, and IR datasets and compare it to both classical and recent baselines. We draw connections to hypothesis testing and p-values.

The allure of superhuman-level capabilities has led to considerable interest in language models like GPT-3 and T5, wherein the research has, by and large, revolved around new model architectures, training tasks, and loss objectives, along with substantial engineering efforts to scale up model capacity and dataset size.Comparatively little work has been done to improve the generalization of these models through better optimization.In this work, we show that Sharpness-Aware Minimization (SAM), a recently proposed optimization procedure that encourages convergence to flatter minima, can substantially improve the generalization of language models without much computational overhead.We show that SAM is able to boost performance on SuperGLUE, GLUE, Web Questions, Natural Questions, Trivia QA, and TyDiQA, with particularly large gains when training data for these tasks is limited.

Recent advances in Transformer-based large language models (LLMs) have led to significant performance improvements across many tasks. These gains come with a drastic increase in the models' size, potentially leading to slow and costly use at inference time. In practice, however, the series of generations made by LLMs is composed of varying levels of difficulty. While certain predictions truly benefit from the models' full capacity, other continuations are more trivial and can be solved with reduced compute. In this work, we introduce Confident Adaptive Language Modeling (CALM), a framework for dynamically allocating different amounts of compute per input and generation timestep. Early exit decoding involves several challenges that we address here, such as: (1) what confidence measure to use; (2) connecting sequence-level constraints to local per-token exit decisions; and (3) attending back to missing hidden representations due to early exits in previous tokens. Through theoretical analysis and empirical experiments on three diverse text generation tasks, we demonstrate the efficacy of our framework in reducing compute---potential speedup of up to $\times 3$---while provably maintaining high performance.

In this paper, we demonstrate that information retrieval can be accomplished with a single Transformer, in which all information about the corpus is encoded in the parameters of the model. To this end, we introduce the Differentiable Search Index (DSI), a new paradigm that learns a text-to-text model that maps string queries directly to relevant docids; in other words, a DSI model answers queries directly using only its parameters, dramatically simplifying the whole retrieval process. We study variations in how documents and their identifiers are represented, variations in training procedures, and the interplay between models and corpus sizes. Experiments demonstrate that given appropriate design choices, DSI significantly outperforms strong baselines such as dual encoder models. Moreover, DSI demonstrates strong generalization capabilities, outperforming a BM25 baseline in a zero-shot setup.

Yikang Shen, Yi Tay, Che Zheng, Dara Bahri, Donald Metzler, Aaron Courville. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Yi Tay, Mostafa Dehghani, Jai Prakash Gupta, Vamsi Aribandi, Dara Bahri, Zhen Qin, Donald Metzler. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Training modern neural networks is an inherently noisy process that can lead to high \emph{prediction churn}– disagreements between re-trainings of the same model due to factors such as randomization in the parameter initialization and mini-batches– even when the trained models all attain similar accuracies. Such prediction churn can be very undesirable in practice. In this paper, we present several baselines for reducing churn and show that training on soft labels obtained by adaptively smoothing each example’s label based on the example’s neighboring labels often outperforms the baselines on churn while improving accuracy on a variety of benchmark classification tasks and model architectures.

This paper proposes Omnidirectional Representations from Transformers (OMNINET). In OmniNet, instead of maintaining a strictly horizon-tal receptive field, each token is allowed to attend to all tokens in the entire network. This process can also be interpreted as a form of extreme or intensive attention mechanism that has the receptive field of the entire width and depth of the network. To this end, the omnidirectional attention is learned via a meta-learner, which is essentially another self-attention based model. In order to mitigate the computationally expensive costs of full receptive field attention, we leverage efficient self-attention models such as kernel-based, low-rank attention and/or Big Bird as the meta-learner. Extensive experiments are conducted on autoregressive language modeling(LM1B, C4), Machine Translation, Long Range Arena (LRA), and Image Recognition.The experiments show that OmniNet achieves considerable improvements across these tasks, including achieving state-of-the-art performance on LM1B,WMT’14 En-De/En-Fr, and Long Range Arena.Moreover, using omnidirectional representation in Vision Transformers leads to significant improvements on image recognition tasks on both few-shot learning and fine-tuning setups.

The dot product self-attention is known to be central and indispensable to state-of-the-art Transformer models. But is it really required? This paper investigates the true importance and contribution of the dot product-based self-attention mechanism on the performance of Transformer models. Via extensive experiments, we find that (1) random alignment matrices surprisingly perform quite competitively and (2) learning attention weights from token-token (query-key) interactions is useful but not that important after all. To this end, we propose \textsc{Synthesizer}, a model that learns synthetic attention weights without token-token interactions. In our experiments, we first show that simple Synthesizers achieve highly competitive performance when compared against vanilla Transformer models across a range of tasks, including machine translation, language modeling, text generation and GLUE/SuperGLUE benchmarks. When composed with dot product attention, we find that Synthesizers consistently outperform Transformers. Moreover, we conduct additional comparisons of Synthesizers against Dynamic Convolutions, showing that simple Random Synthesizer is not only $60%$ faster but also improves perplexity by a relative $3.5%$. Finally, we show that simple factorized Synthesizers can outperform Linformers on encoding only tasks.

Large generative language models such as GPT-2 are well-known for their ability to generate text as well as their utility in supervised downstream tasks via fine-tuning. Its prevalence on the web, however, is still not well understood - if we run GPT-2 detectors across the web, what will we find? Our work is twofold: firstly we demonstrate via human evaluation that classifiers trained to discriminate between human and machine-generated text emerge as unsupervised predictors of "page quality", able to detect low quality content without any training. This enables fast bootstrapping of quality indicators in a low-resource setting. Secondly, curious to understand the prevalence and nature of low quality pages in the wild, we conduct extensive qualitative and quantitative analysis over 500 million web articles, making this the largest-scale study ever conducted on the topic.

This paper seeks to develop a deeper understanding of the fundamental properties of neural text generations models.The study of artifacts that emerge in machine generated text as a result of modeling choices is a nascent research area.Previously, the extent and degree to which these artifacts surface in generated text has not been well studied.In the spirit of better understanding generative text models and their artifacts, we propose the new task of distinguishing which of several variants of a given model generated a piece of text, and we conduct an extensive suite of diagnostic tests to observe whether modeling choices (e.g., sampling methods, top-k probabilities, model architectures, etc.) leave detectable artifacts in the text they generate.Our key finding, which is backed by a rigorous set of experiments, is that such artifacts are present and that different modeling choices can be inferred by observing the generated text alone.This suggests that neural text generators may be more sensitive to various modeling choices than previously thought.

Modern machine learning models are often trained on examples with noisy labels that hurt performance and are hard to identify. In this paper, we provide an empirical study showing that a simple $k$-nearest neighbor-based filtering approach on the logit layer of a preliminary model can remove mislabeled training data and produce more accurate models than many recently proposed methods. We also provide new statistical guarantees into its efficacy.

We propose Sparse Sinkhorn Attention, a new efficient and sparse method for learning to attend. Our method is based on differentiable sorting of internal representations. Concretely, we introduce a meta sorting network that learns to generate latent permutations over sequences. Given sorted sequences, we are then able to compute quasi-global attention with only local windows, improving the memory efficiency of the attention module. To this end, we propose new algorithmic innovations such as Causal Sinkhorn Balancing and SortCut, a dynamic sequence truncation method for tailoring Sinkhorn Attention for encoding and/or decoding purposes. Via extensive experiments on algorithmic seq2seq sorting, language modeling, pixel-wise image generation, document classification and natural language inference, we demonstrate that our memory efficient Sinkhorn Attention method is competitive with vanilla attention and consistently outperforms recently proposed efficient Transformer models such as Sparse Transformers.

Work in information retrieval has traditionally focused on ranking and relevance: given a query, return some number of results ordered by relevance to the user. However, the problem of determining how many results to return, i.e. how to optimally truncate the ranked result list, has received less attention despite being of critical importance in a range of applications. Such truncation is a balancing act between the overall relevance, or usefulness of the results, with the user cost of processing more results. In this work, we propose Choppy, an assumption-free model based on the widely successful Transformer architecture, to the ranked list truncation problem. Needing nothing more than the relevance scores of the results, the model uses a powerful multi-head attention mechanism to directly optimize any user-defined IR metric. We show Choppy improves upon recent state-of-the-art methods.

We investigate machine learning models that can provide diminishing returns and accelerating returns guarantees to capture prior knowledge or policies about how outputs should depend on inputs. We show that one can build flexible, nonlinear, multi-dimensional models using lattice functions with any combination of concavity/convexity and monotonicity constraints on any subsets of features, and compare to new shape-constrained neural networks. We demonstrate on real-world examples that these shape constrained models can provide tuning-free regularization and improve model understandability.