Despite the current enthusiasm for GAMs, their susceptibility to concurvity - i.e., (possibly nonlinear) dependencies between the features - has hitherto been largely overlooked.

Digital pathology has revolutionized the field by enabling the digitization of tissue samples into whole slide images (WSIs). However, the high resolution and large size of WSIs present significant challenges when it comes to applying Deep Learning models. As a solution, WSIs are often divided into smaller patches with a global label (\textit{i.e., diagnostic}) per slide, instead of a (too) costly pixel-wise annotation. By treating each slide as a bag of patches, Multiple Instance Learning (MIL) methods have emerged as a suitable solution for WSI classification. A major drawback of MIL methods is their high variability in performance across different runs, which can reach up to 10-15 AUC points on the test set, making it difficult to compare different MIL methods reliably. This variability mainly comes from three factors: i) weight initialization, ii) batch (shuffling) ordering, iii) and learning rate. To address that, we introduce a Multi-Fidelity, Model Fusion strategy for MIL methods. We first train multiple models for a few epochs and average the most stable and promising ones based on validation scores. This approach can be applied to any existing MIL model to reduce performance variability. It also simplifies hyperparameter tuning and improves reproducibility while maintaining computational efficiency. We extensively validate our approach on WSI classification tasks using 2 different datasets, 3 initialization strategies and 5 MIL methods, for a total of more than 2000 experiments.

Pruning offers a promising solution to mitigate the associated costs and environmental impact of deploying large deep neural networks (DNNs). Traditional approaches rely on computationally expensive trained models or time-consuming iterative prune-retrain cycles, undermining their utility in resource-constrained settings. To address this issue, we build upon the established principles of saliency (LeCun et al., 1989) and connection sensitivity (Lee et al., 2018) to tackle the challenging problem of one-shot pruning neural networks (NNs) before training (PBT) at initialization. We introduce Fisher-Taylor Sensitivity (FTS), a computationally cheap and efficient pruning criterion based on the empirical Fisher Information Matrix (FIM) diagonal, offering a viable alternative for integrating first- and second-order information to identify a model's structurally important parameters. Although the FIM-Hessian equivalency only holds for convergent models that maximize the likelihood, recent studies (Karakida et al., 2019) suggest that, even at initialization, the FIM captures essential geometric information of parameters in overparameterized NNs, providing the basis for our method. Finally, we demonstrate empirically that layer collapse, a critical limitation of data-dependent pruning methodologies, is easily overcome by pruning within a single training epoch after initialization. We perform experiments on ResNet18 and VGG19 with CIFAR-10 and CIFAR-100, widely used benchmarks in pruning research. Our method achieves competitive performance against state-of-the-art techniques for one-shot PBT, even under extreme sparsity conditions. Our code is made available to the public.

Deep learning has become a particularly important set of machine learning techniques and is widely applied to solve real-world tasks. At the same time, many open questions remain with regard to the ability of these deep neural networks (DNNs) to generalize so well, that is, their ability to perform well on unseen data. Although there is not yet a theoretical framework to assist us in reasoning about these models [2], the generalization ability of DNNs has been studied from many perspectives, such as the geometry of the loss landscape [3], statistical measures of stability and robustness [4], size of margins (distance to the decision boundary between classes) [5], and information-theoretic techniques [6], among others. A promising research direction is to study the characteristics of the internal data representations formed by DNNs, where each representation is the vector of activation values from a specific layer for a given sample. Aspects of these representations that have been studied include the size of margins in the representation space [7, 8, 9]; the'quality' of representations, evaluated using the consistency of class-specific representations and their robustness when combined [9]; and representation sparsity, that is, the number of non-zero elements in a data representation [10]. In this work, we also study the characteristics of the internal representations of DNNs, but focus on the effect that a very specific technique - Batch Normalization (BatchNorm) - has on internal representation quality. BatchNorm [11] is a popular technique used to normalize hidden activations when training DNNs. Networks trained with BatchNorm show desirable properties such as faster convergence and better generalization ability [12, 13]. Despite the success and widespread adoption of BatchNorm, the exact mechanisms by which BatchNorm achieves its performance remain unclear.

Generalized Additive Models (GAMs) have recently experienced a resurgence in popularity due to their interpretability, which arises from expressing the target value as a sum of non-linear transformations of the features. Despite the current enthusiasm for GAMs, their susceptibility to concurvity - i.e., (possibly non-linear) dependencies between the features - has hitherto been largely overlooked. Here, we demonstrate how concurvity can severly impair the interpretability of GAMs and propose a remedy: a conceptually simple, yet effective regularizer which penalizes pairwise correlations of the non-linearly transformed feature variables. This procedure is applicable to any differentiable additive model, such as Neural Additive Models or NeuralProphet, and enhances interpretability by eliminating ambiguities due to self-canceling feature contributions. We validate the effectiveness of our regularizer in experiments on synthetic as well as real-world datasets for time-series and tabular data. Our experiments show that concurvity in GAMs can be reduced without significantly compromising prediction quality, improving interpretability and reducing variance in the feature importances.

Language is constantly changing and evolving, leaving language models to become quickly outdated. Consequently, we should continuously update our models with new data to expose them to new events and facts. However, that requires additional computing, which means new carbon emissions. Do any measurable benefits justify this cost? This paper looks for empirical evidence to support continuous training. We reproduce existing benchmarks and extend them to include additional time periods, models, and tasks. Our results show that the downstream task performance of temporally adapted English models for social media data do not improve over time. Pretrained models without temporal adaptation are actually significantly more effective and efficient. However, we also note a lack of suitable temporal benchmarks. Our findings invite a critical reflection on when and how to temporally adapt language models, accounting for sustainability.

Sequence optimization, where the items in a list are ordered to maximize some reward has many applications such as web advertisement placement, search, and control libraries in robotics. Previous work in sequence optimization produces a static ordering that does not take any features of the item or context of the problem into account. In this work, we propose a general approach to order the items within the sequence based on the context (e.g., perceptual information, environment description, and goals). We take a simple, efficient, reduction-based approach where the choice and order of the items is established by repeatedly learning simple classifiers or regressors for each "slot" in the sequence. Our approach leverages recent work on submodular function maximization to provide a formal regret reduction from submodular sequence optimization to simple cost-sensitive prediction. We apply our contextual sequence prediction algorithm to optimize control libraries and demonstrate results on two robotics problems: manipulator trajectory prediction and mobile robot path planning.