At its core, this thesis aims to enhance the practicality of deep learning by improving the label and training efficiency of deep learning models. To this end, we investigate data subset selection techniques, specifically active learning and active sampling, grounded in information-theoretic principles. Active learning improves label efficiency, while active sampling enhances training efficiency. Supervised deep learning models often require extensive training with labeled data. Label acquisition can be expensive and time-consuming, and training large models is resource-intensive, hindering the adoption outside academic research and ``big tech.'' Existing methods for data subset selection in deep learning often rely on heuristics or lack a principled information-theoretic foundation. In contrast, this thesis examines several objectives for data subset selection and their applications within deep learning, striving for a more principled approach inspired by information theory. We begin by disentangling epistemic and aleatoric uncertainty in single forward-pass deep neural networks, which provides helpful intuitions and insights into different forms of uncertainty and their relevance for data subset selection. We then propose and investigate various approaches for active learning and data subset selection in (Bayesian) deep learning. Finally, we relate various existing and proposed approaches to approximations of information quantities in weight or prediction space. Underpinning this work is a principled and practical notation for information-theoretic quantities that includes both random variables and observed outcomes. This thesis demonstrates the benefits of working from a unified perspective and highlights the potential impact of our contributions to the practical application of deep learning.
Information-theoretic approaches to active learning have traditionally focused on maximising the information gathered about the model parameters, most commonly by optimising the BALD score. We highlight that this can be suboptimal from the perspective of predictive performance. For example, BALD lacks a notion of an input distribution and so is prone to prioritise data of limited relevance. To address this we propose the expected predictive information gain (EPIG), an acquisition function that measures information gain in the space of predictions rather than parameters. We find that using EPIG leads to stronger predictive performance compared with BALD across a range of datasets and models, and thus provides an appealing drop-in replacement.
The paper 'Deep Learning on a Data Diet' by Paul et al. (2021) introduces two innovative metrics for pruning datasets during the training of neural networks. While we are able to replicate the results for the EL2N score at epoch 20, the same cannot be said for the GraNd score at initialization. The GraNd scores later in training provide useful pruning signals, however. The GraNd score at initialization calculates the average gradient norm of an input sample across multiple randomly initialized models before any training has taken place. Our analysis reveals a strong correlation between the GraNd score at initialization and the input norm of a sample, suggesting that the latter could have been a cheap new baseline for data pruning. Unfortunately, neither the GraNd score at initialization nor the input norm surpasses random pruning in performance. This contradicts one of the findings in Paul et al. (2021). We were unable to reproduce their CIFAR-10 results using both an updated version of the original JAX repository and in a newly implemented PyTorch codebase. An investigation of the underlying JAX/FLAX code from 2021 surfaced a bug in the checkpoint restoring code that was fixed in April 2021 (https://github.com/google/flax/commit/28fbd95500f4bf2f9924d2560062fa50e919b1a5).
Batch active learning is a popular approach for efficiently training machine learning models on large, initially unlabelled datasets, which repeatedly acquires labels for a batch of data points. However, many recent batch active learning methods are white-box approaches limited to differentiable parametric models: they score unlabeled points using acquisition functions based on model embeddings or first- and second-order derivatives. In this paper, we propose black-box batch active learning for regression tasks as an extension of white-box approaches. This approach is compatible with a wide range of machine learning models including regular and Bayesian deep learning models and non-differentiable models such as random forests. It is rooted in Bayesian principles and utilizes recent kernel-based approaches. Importantly, our method only relies on model predictions. This allows us to extend a wide range of existing state-of-the-art white-box batch active learning methods (BADGE, BAIT, LCMD) to black-box models. We demonstrate the effectiveness of our approach through extensive experimental evaluations on regression datasets, achieving surprisingly strong performance compared to white-box approaches for deep learning models.
Active learning is a powerful method for training machine learning models with limited labeled data. One commonly used technique for active learning is BatchBALD, which uses Bayesian neural networks to find the most informative points to label in a pool set. However, BatchBALD can be very slow to compute, especially for larger datasets. In this paper, we propose a new approximation, k-BALD, which uses k-wise mutual information terms to approximate BatchBALD, making it much less expensive to compute. Results on the MNIST dataset show that k-BALD is significantly faster than BatchBALD while maintaining similar performance. Additionally, we also propose a dynamic approach for choosing k based on the quality of the approximation, making it more efficient for larger datasets.
The mutual information between predictions and model parameters -- also referred to as expected information gain or BALD in machine learning -- measures informativeness. It is a popular acquisition function in Bayesian active learning and Bayesian optimal experiment design. In data subset selection, i.e. active learning and active sampling, several recent works use Fisher information, Hessians, similarity matrices based on the gradients, or simply the gradient lengths to compute the acquisition scores that guide sample selection. Are these different approaches connected, and if so how? In this paper, we revisit the Fisher information and use it to show how several otherwise disparate methods are connected as approximations of information-theoretic quantities.
A recent trend in artificial intelligence is the use of pretrained models for language and vision tasks, which have achieved extraordinary performance but also puzzling failures. Probing these models' abilities in diverse ways is therefore critical to the field. In this paper, we explore the reliability of models, where we define a reliable model as one that not only achieves strong predictive performance but also performs well consistently over many decision-making tasks involving uncertainty (e.g., selective prediction, open set recognition), robust generalization (e.g., accuracy and proper scoring rules such as log-likelihood on in- and out-of-distribution datasets), and adaptation (e.g., active learning, few-shot uncertainty). We devise 10 types of tasks over 40 datasets in order to evaluate different aspects of reliability on both vision and language domains. To improve reliability, we developed ViT-Plex and T5-Plex, pretrained large model extensions for vision and language modalities, respectively. Plex greatly improves the state-of-the-art across reliability tasks, and simplifies the traditional protocol as it improves the out-of-the-box performance and does not require designing scores or tuning the model for each task. We demonstrate scaling effects over model sizes up to 1B parameters and pretraining dataset sizes up to 4B examples. We also demonstrate Plex's capabilities on challenging tasks including zero-shot open set recognition, active learning, and uncertainty in conversational language understanding.
Training on web-scale data can take months. But most computation and time is wasted on redundant and noisy points that are already learnt or not learnable. To accelerate training, we introduce Reducible Holdout Loss Selection (RHO-LOSS), a simple but principled technique which selects approximately those points for training that most reduce the model's generalization loss. As a result, RHO-LOSS mitigates the weaknesses of existing data selection methods: techniques from the optimization literature typically select 'hard' (e.g. high loss) points, but such points are often noisy (not learnable) or less task-relevant. Conversely, curriculum learning prioritizes 'easy' points, but such points need not be trained on once learned. In contrast, RHO-LOSS selects points that are learnable, worth learning, and not yet learnt. RHO-LOSS trains in far fewer steps than prior art, improves accuracy, and speeds up training on a wide range of datasets, hyperparameters, and architectures (MLPs, CNNs, and BERT). On the large web-scraped image dataset Clothing-1M, RHO-LOSS trains in 18x fewer steps and reaches 2% higher final accuracy than uniform data shuffling.
Principled Bayesian deep learning (BDL) does not live up to its potential when we only focus on marginal predictive distributions (marginal predictives). Recent works have highlighted the importance of joint predictives for (Bayesian) sequential decision making from a theoretical and synthetic perspective. We provide additional practical arguments grounded in real-world applications for focusing on joint predictives: we discuss online Bayesian inference, which would allow us to make predictions while taking into account additional data without retraining, and we propose new challenging evaluation settings using active learning and active sampling. These settings are motivated by an examination of marginal and joint predictives, their respective cross-entropies, and their place in offline and online learning. They are more realistic than previously suggested ones, building on work by Wen et al. (2021) and Osband et al. (2022), and focus on evaluating the performance of approximate BNNs in an online supervised setting. Initial experiments, however, raise questions on the feasibility of these ideas in high-dimensional parameter spaces with current BDL inference techniques, and we suggest experiments that might help shed further light on the practicality of current research for these problems. Importantly, our work highlights previously unidentified gaps in current research and the need for better approximate joint predictives.
Jiang et al. (2021) give empirical evidence that the average test error of deep neural networks can be estimated via the prediction disagreement of two separately trained networks. They also provide a theoretical explanation that this 'Generalization Disagreement Equality' follows from the well-calibrated nature of deep ensembles under the notion of a proposed 'class-aggregated calibration'. In this paper we show that the approach suggested might be impractical because a deep ensemble's calibration deteriorates under distribution shift, which is exactly when the coupling of test error and disagreement would be of practical value. We present both theoretical and experimental evidence, re-deriving the theoretical statements using a simple Bayesian perspective and show them to be straightforward and more generic: they apply to any discriminative model -- not only ensembles whose members output one-hot class predictions. The proposed calibration metrics are also equivalent to two metrics introduced by Nixon et al. (2019): 'ACE' and 'SCE'.