linear classifiers

We study the feasibility of identifying epistemic uncertainty (reflecting a lack of knowledge), as opposed to aleatoric uncertainty (reflecting entropy in the underlying distribution), in the outputs of large language models (LLMs) over free-form text. In the absence of ground-truth probabilities, we explore a setting where, in order to (approximately) disentangle a given LLM’s uncertainty, a significantly larger model stands in as a proxy for the ground truth. We show that small linear probes trained on the embeddings of frozen, pretrained models accurately predict when larger models will be more confident at the token level and that probes trained on one text domain generalize to others. Going further, we propose a fully unsupervised method that achieves non-trivial accuracy on the same task. Taken together, we interpret these results as evidence that LLMs naturally contain internal representations of different types of uncertainty that could potentially be leveraged to devise more informative indicators of model confidence in diverse practical settings. Code can be found at: https://github.com/KempnerInstitute/llm_uncertainty

The study of strategic or adversarial manipulation of testing data to fool a classifier has attracted much recent attention. Most previous works have focused on two extreme situations where any testing data point either is completely adversarial or always equally prefers the positive label. In this paper, we generalize both of these through a unified framework by considering strategic agents with heterogenous preferences, and introduce the notion of strategic VC-dimension (SVC) to capture the PAC-learnability in our general strategic setup. SVC provably generalizes the recent concept of adversarial VC-dimension (AVC) introduced by Cullina et al. (2018). We instantiate our framework for the fundamental strategic linear classification problem. We fully characterize: (1) the statistical learnability of linear classifiers by pinning down its SVC; (2) its computational tractability by pinning down the complexity of the empirical risk minimization problem. Interestingly, the SVC of linear classifiers is always upper bounded by its standard VC-dimension. This characterization also strictly generalizes the AVC bound for linear classifiers in (Cullina et al., 2018). Finally, we briefly investigate the power of randomization in our strategic classification setup. We show that randomization may strictly increase the accuracy in general, but will not help in the special case of adversarial classification with zero-manipulation-cost.

Investigating active learning, we focus on the relation between the number of labeled examples (budget size), and suitable querying strategies. Our theoretical analysis shows a behavior reminiscent of phase transition: typical examples are best queried when the budget is low, while unrepresentative examples are best queried when the budget is large. Combined evidence shows that a similar phenomenon occurs in common classification models. Accordingly, we propose TypiClust – a deep active learning strategy suited for low budgets. In a comparative empirical investigation of supervised learning, using a variety of architectures and image datasets, TypiClust outperforms all other active learning strategies in the low-budget regime. Using TypiClust in the semi-supervised framework, performance gets an even more significant boost. In particular, state-of-the-art semi-supervised methods trained on CIFAR-10 with 10 labeled examples selected by TypiClust, reach 93.2% accuracy – an improvement of 39.4% over random selection. Code is available at https://github.com/avihu111/TypiClust.

We consider the problem of producing compact architectures for text classification, such that the full model fits in a limited amount of memory. After considering different solutions inspired by the hashing literature, we propose a method built upon product quantization to store word embeddings. While the original technique leads to a loss in accuracy, we adapt this method to circumvent quantization artefacts. Our experiments carried out on several benchmarks show that our approach typically requires two orders of magnitude less memory than fastText while being only slightly inferior with respect to accuracy. As a result, it outperforms the state of the art by a good margin in terms of the compromise between memory usage and accuracy.

Many meta-learning approaches for few-shot learning rely on simple base learners such as nearest-neighbor classifiers. However, even in the few-shot regime, discriminatively trained linear predictors can offer better generalization. We propose to use these predictors as base learners to learn representations for few-shot learning and show they offer better tradeoffs between feature size and performance across a range of few-shot recognition benchmarks. Our objective is to learn feature embeddings that generalize well under a linear classification rule for novel categories. To efficiently solve the objective, we exploit two properties of linear classifiers: implicit differentiation of the optimality conditions of the convex problem and the dual formulation of the optimization problem. This allows us to use high-dimensional embeddings with improved generalization at a modest increase in computational overhead. Our approach, named MetaOptNet, achieves state-of-the-art performance on miniImageNet, tieredImageNet, CIFAR-FS, and FC100 few-shot learning benchmarks. Our code is available at this https URL.

During the last two decades a considerable amount of effort has been devoted to the analysis of the influence of both training and testing sample size on the design and performance of pattern recognition systems. These questions are interesting to practitioners as well as theoreticians, because the small-sample effects can easily contaminate the design and evaluation of a proposed system. For applications with a large number of features and a complex classification rule, the training sample size must be quite large. A large test sample is required to accurately evaluate a classifier with a low error rate. The design of a pattern recognition system consists of several stages: data collection, formation of the pattern classes, feature selection, specification of the classification algorithm, and estimation of the classification error. In this paper, we will discuss the effects of sample size on feature selection and error estimation for several types of classifier. In addition to surveying prior work in this area, our emphasis is on giving practical advice to today's designers and users of statistical pattern recognition systems.

We compare discriminative and generative learning as typified by logistic regression and naive Bayes. We show, contrary to a widely-held belief that discriminative classifiers are almost always to be preferred, that there can often be two distinct regimes of performance as the training set size is increased, one in which each algorithm does better. This stems from the observation—which is borne out in repeated experiments—that while discriminative learning has lower asymptotic error, a generative classifier may also approach its (higher) asymptotic error much faster.

The Fisher kernel (FK) is a generic framework which combines the benefits of generative and discriminative approaches. In the context of image classification the FK was shown to extend the popular bag-of-visual-words (BOV) by going beyond count statistics. However, in practice, this enriched representation has not yet shown its superiority over the BOV. In the first part we show that with several well-motivated modifications over the original framework we can boost the accuracy of the FK. On PASCAL VOC 2007 we increase the Average Precision (AP) from 47.9% to 58.3%. Similarly, we demonstrate state-of-the-art accuracy on CalTech 256. A major advantage is that these results are obtained using only SIFT descriptors and costless linear classifiers. Equipped with this representation, we can now explore image classification on a larger scale. In the second part, as an application, we compare two abundant resources of labeled images to learn classifiers: ImageNet and Flickr groups. In an evaluation involving hundreds of thousands of training images we show that classifiers learned on Flickr groups perform surprisingly well (although they were not intended for this purpose) and that they can complement classifiers learned on more carefully annotated datasets.

The automated categorization (or classification) of texts into predefined categories has witnessed a booming interest in the last ten years, due to the increased availability of documents in digital form and the ensuing need to organize them. In the research community the dominant approach to this problem is based on machine learning techniques: a general inductive process automatically builds a classifier by learning, from a set of preclassified documents, the characteristics of the categories. The advantages of this approach over the knowledge engineering approach (consisting in the manual definition of a classifier by domain experts) are a very good effectiveness, considerable savings in terms of expert manpower, and straightforward portability to different domains. This survey discusses the main approaches to text categorization that fall within the machine learning paradigm. We will discuss in detail issues pertaining to three different problems, namely document representation, classifier construction, and classifier evaluation.