Benchmarks such as GLUE have helped drive advances in NLP by incentivizing the creation of more accurate models. While this leaderboard paradigm has been remarkably successful, a historical focus on performance-based evaluation has been at the expense of other qualities that the NLP community values in models, such as compactness, fairness, and energy efficiency. In this opinion paper, we study the divergence between what is incentivized by leaderboards and what is useful in practice through the lens of microeconomic theory. We frame both the leaderboard and NLP practitioners as consumers and the benefit they get from a model as its utility to them. With this framing, we formalize how leaderboards -- in their current form -- can be poor proxies for the NLP community at large. For example, a highly inefficient model would provide less utility to practitioners but not to a leaderboard, since it is a cost that only the former must bear. To allow practitioners to better estimate a model's utility to them, we advocate for more transparency on leaderboards, such as the reporting of statistics that are of practical concern (e.g., model size, energy efficiency, and inference latency).

A growing body of recent evidence has highlighted the limitations of natural language processing (NLP) datasets and classifiers. These include the presence of annotation artifacts in datasets, classifiers relying on shallow features like a single word (e.g., if a movie review has the word "romantic", the review tends to be positive), or unnecessary words (e.g., learning a proper noun to classify a movie as positive or negative). The presence of such artifacts has subsequently led to the development of challenging datasets to force the model to generalize better. While a variety of heuristic strategies, such as counterfactual examples and contrast sets, have been proposed, the theoretical justification about what makes these examples difficult is often lacking or unclear. In this paper, using tools from information geometry, we propose a theoretical way to quantify the difficulty of an example in NLP. Using our approach, we explore difficult examples for two popular NLP architectures. We discover that both BERT and CNN are susceptible to single word substitutions in high difficulty examples. Consequently, examples with low difficulty scores tend to be robust to multiple word substitutions. Our analysis shows that perturbations like contrast sets and counterfactual examples are not necessarily difficult for the model, and they may not be accomplishing the intended goal. Our approach is simple, architecture agnostic, and easily extendable to other datasets. All the code used will be made publicly available, including a tool to explore the difficult examples for other datasets.

Few-shot-learning seeks to find models that are capable of fast-adaptation to novel tasks. Unlike typical few-shot learning algorithms, we propose a contrastive learning method which is not trained to solve a set of tasks, but rather attempts to find a good representation of the underlying data-generating processes (\emph{functions}). This allows for finding representations which are useful for an entire series of tasks sharing the same function. In particular, our training scheme is driven by the self-supervision signal indicating whether two sets of samples stem from the same underlying function. Our experiments on a number of synthetic and real-world datasets show that the representations we obtain can outperform strong baselines in terms of downstream performance and noise robustness, even when these baselines are trained in an end-to-end manner.

Many types of neural network layers rely on matrix properties such as invertibility or orthogonality. Retaining such properties during optimization with gradient-based stochastic optimizers is a challenging task, which is usually addressed by either reparameterization of the affected parameters or by directly optimizing on the manifold. In contrast, this work presents a novel, general approach of preserving matrix properties by using parameterized perturbations. In lieu of directly optimizing the network parameters, the introduced P$^{4}$ update optimizes perturbations and merges them into the actual parameters infrequently such that the desired property is preserved. As a demonstration, we use this concept to preserve invertibility of linear layers during training. This P$^{4}$Inv update allows keeping track of inverses and determinants via rank-one updates and without ever explicitly computing them. We show how such invertible blocks improve the mixing of coupling layers and thus the mode separation of the resulting normalizing flows.

In this work, we present a novel approach for simultaneous knowledge transfer and model compression called Weight Squeezing. With this method, we perform knowledge transfer from a pre-trained teacher model by learning the mapping from its weights to smaller student model weights, without significant loss of model accuracy. We applied Weight Squeezing combined with Knowledge Distillation to a pre-trained text classification model, and compared it to various knowledge transfer and model compression methods on several downstream text classification tasks. We observed that our approach produces better results than Knowledge Distillation methods without any loss in inference speed. We also compared Weight Squeezing with Low Rank Factorization methods and observed that our method is significantly faster at inference while being competitive in terms of accuracy.

State-of-the-art machine learning models require access to significant amount of annotated data in order to achieve the desired level of performance. While unlabelled data can be largely available and even abundant, annotation process can be quite expensive and limiting. Under the assumption that some samples are more important for a given task than others, active learning targets the problem of identifying the most informative samples that one should acquire annotations for. Instead of the conventional reliance on model uncertainty as a proxy to leverage new unknown labels, in this work we propose a simple sample selection criterion that moves beyond uncertainty. By first accepting the model prediction and then judging its effect on the generalization error, we can better identify wrongly predicted samples. We further present an approximation to our criterion that is very efficient and provides a similarity based interpretation. In addition to evaluating our method on the standard benchmarks of active learning, we consider the challenging yet realistic scenario of imbalanced data where categories are not equally represented. We show state-of-the-art results and better rates at identifying wrongly predicted samples. Our method is simple, model agnostic and relies on the current model status without the need for re-training from scratch.

We present neural mixture distributional regression (NMDR), a holistic framework to estimate complex finite mixtures of distributional regressions defined by flexible additive predictors. Our framework is able to handle a large number of mixtures of potentially different distributions in high-dimensional settings, allows for efficient and scalable optimization and can be applied to recent concepts that combine structured regression models with deep neural networks. While many existing approaches for mixture models address challenges in optimization of such and provide results for convergence under specific model assumptions, our approach is assumption-free and instead makes use of optimizers well-established in deep learning. Through extensive numerical experiments and a high-dimensional deep learning application we provide evidence that the proposed approach is competitive to existing approaches and works well in more complex scenarios.

Anomaly detection in time series has been widely researched and has important practical applications. In recent years, anomaly detection algorithms are mostly based on deep-learning generative models and use the reconstruction error to detect anomalies. They try to capture the distribution of normal data by reconstructing normal data in the training phase, then calculate the reconstruction error of test data to do anomaly detection. However, most of them only use the normal data in the training phase and can not ensure the reconstruction process of anomaly data. So, anomaly data can also be well reconstructed sometimes and gets low reconstruction error, which leads to the omission of anomalies. What's more, the neighbor information of data points in time series data has not been fully utilized in these algorithms. In this paper, we propose RAN based on the idea of Reconstruct Anomalies to Normal and apply it for unsupervised time series anomaly detection. To minimize the reconstruction error of normal data and maximize this of anomaly data, we do not just ensure normal data to reconstruct well, but also try to make the reconstruction of anomaly data consistent with the distribution of normal data, then anomalies will get higher reconstruction errors. We implement this idea by introducing the "imitated anomaly data" and combining a specially designed latent vector-constrained Autoencoder with the discriminator to construct an adversary network. Extensive experiments on time-series datasets from different scenes such as ECG diagnosis also show that RAN can detect meaningful anomalies, and it outperforms other algorithms in terms of AUC-ROC.

Cortical networks are strongly recurrent, and neurons have intrinsic temporal dynamics. This sets them apart from deep feed-forward networks. Despite the tremendous progress in the application of feed-forward networks and their theoretical understanding, it remains unclear how the interplay of recurrence and non-linearities in recurrent cortical networks contributes to their function. The purpose of this work is to present a solvable recurrent network model that links to feed forward networks. By perturbative methods we transform the time-continuous, recurrent dynamics into an effective feed-forward structure of linear and non-linear temporal kernels. The resulting analytical expressions allow us to build optimal time-series classifiers from random reservoir networks. Firstly, this allows us to optimize not only the readout vectors, but also the input projection, demonstrating a strong potential performance gain. Secondly, the analysis exposes how the second order stimulus statistics is a crucial element that interacts with the non-linearity of the dynamics and boosts performance.

Built upon GPflow and RDKit, FlowMO enables the user to make predictions with well-calibrated uncertainty estimates, an output central to active learning and molecular design applications. Gaussian Processes are particularly attractive for modelling small molecular datasets, a characteristic of many real-world virtual screening campaigns where high-quality experimental data is scarce. Computational experiments across three small datasets demonstrate comparable predictive performance to deep learning methods but with superior uncertainty calibration.