The generalization and learning speed of a multi-class neural network can often be significantly improved by using soft targets that are a weighted average of the hard targets and the uniform distribution over labels. Smoothing the labels in this way prevents the network from becoming over-confident and label smoothing has been used in many state-of-the-art models, including image classification, language translation and speech recognition. Despite its widespread use, label smoothing is still poorly understood. Here we show empirically that in addition to improving generalization, label smoothing improves model calibration which can significantly improve beam-search. However, we also observe that if a teacher network is trained with label smoothing, knowledge distillation into a student network is much less effective. To explain these observations, we visualize how label smoothing changes the representations learned by the penultimate layer of the network. We show that label smoothing encourages the representations of training examples from the same class to group in tight clusters. This results in loss of information in the logits about resemblances between instances of different classes, which is necessary for distillation, but does not hurt generalization or calibration of the model's predictions.

Explainable AI transforms opaque decision strategies of ML models into explanations that are interpretable by the user, for example, identifying the contribution of each input feature to the prediction at hand. Such explanations, however, entangle the potentially multiple factors that enter into the overall complex decision strategy. We propose to disentangle explanations by finding relevant subspaces in activation space that can be mapped to more abstract human-understandable concepts and enable a joint attribution on concepts and input features. To automatically extract the desired representation, we propose new subspace analysis formulations that extend the principle of PCA and subspace analysis to explanations. These novel analyses, which we call principal relevant component analysis (PRCA) and disentangled relevant subspace analysis (DRSA), optimize relevance of projected activations rather than the more traditional variance or kurtosis. This enables a much stronger focus on subspaces that are truly relevant for the prediction and the explanation, in particular, ignoring activations or concepts to which the prediction model is invariant. Our approach is general enough to work alongside common attribution techniques such as Shapley Value, Integrated Gradients, or LRP. Our proposed methods show to be practically useful and compare favorably to the state of the art as demonstrated on benchmarks and three use cases.

Kernel machines have sustained continuous progress in the field of quantum chemistry. In particular, they have proven to be successful in the low-data regime of force field reconstruction. This is because many physical invariances and symmetries can be incorporated into the kernel function to compensate for much larger datasets. So far, the scalability of this approach has however been hindered by its cubical runtime in the number of training points. While it is known, that iterative Krylov subspace solvers can overcome these burdens, they crucially rely on effective preconditioners, which are elusive in practice. Practical preconditioners need to be computationally efficient and numerically robust at the same time. Here, we consider the broad class of Nystr\"om-type methods to construct preconditioners based on successively more sophisticated low-rank approximations of the original kernel matrix, each of which provides a different set of computational trade-offs. All considered methods estimate the relevant subspace spanned by the kernel matrix columns using different strategies to identify a representative set of inducing points. Our comprehensive study covers the full spectrum of approaches, starting from naive random sampling to leverage score estimates and incomplete Cholesky factorizations, up to exact SVD decompositions.

Applying deep learning concepts from image detection and graph theory has greatly advanced protein-ligand binding affinity prediction, a challenge with enormous ramifications for both drug discovery and protein engineering. We build upon these advances by designing a novel deep learning architecture consisting of a 3-dimensional convolutional neural network utilizing channel-wise attention and two graph convolutional networks utilizing attention-based aggregation of node features. HAC-Net (Hybrid Attention-Based Convolutional Neural Network) obtains state-of-the-art results on the PDBbind v.2016 core set, the most widely recognized benchmark in the field. We extensively assess the generalizability of our model using multiple train-test splits, each of which maximizes differences between either protein structures, protein sequences, or ligand extended-connectivity fingerprints. Furthermore, we perform 10-fold cross-validation with a similarity cutoff between SMILES strings of ligands in the training and test sets, and also evaluate the performance of HAC-Net on lower-quality data. We envision that this model can be extended to a broad range of supervised learning problems related to structure-based biomolecular property prediction. All of our software is available as open source at https://github.com/gregory-kyro/HAC-Net/.

Dataset scaling, also known as normalization, is an essential preprocessing step in a machine learning pipeline. It is aimed at adjusting attributes scales in a way that they all vary within the same range. This transformation is known to improve the performance of classification models, but there are several scaling techniques to choose from, and this choice is not generally done carefully. In this paper, we execute a broad experiment comparing the impact of 5 scaling techniques on the performances of 20 classification algorithms among monolithic and ensemble models, applying them to 82 publicly available datasets with varying imbalance ratios. Results show that the choice of scaling technique matters for classification performance, and the performance difference between the best and the worst scaling technique is relevant and statistically significant in most cases. They also indicate that choosing an inadequate technique can be more detrimental to classification performance than not scaling the data at all. We also show how the performance variation of an ensemble model, considering different scaling techniques, tends to be dictated by that of its base model. Finally, we discuss the relationship between a model's sensitivity to the choice of scaling technique and its performance and provide insights into its applicability on different model deployment scenarios. Full results and source code for the experiments in this paper are available in a GitHub repository.\footnote{https://github.com/amorimlb/scaling\_matters}

Neuromorphic systems require user-friendly software to support the design and optimization of experiments. In this work, we address this need by presenting our development of a machine learning-based modeling framework for the BrainScaleS-2 neuromorphic system. This work represents an improvement over previous efforts, which either focused on the matrix-multiplication mode of BrainScaleS-2 or lacked full automation. Our framework, called hxtorch.snn, enables the hardware-in-the-loop training of spiking neural networks within PyTorch, including support for auto differentiation in a fully-automated hardware experiment workflow. In addition, hxtorch.snn facilitates seamless transitions between emulating on hardware and simulating in software. We demonstrate the capabilities of hxtorch.snn on a classification task using the Yin-Yang dataset employing a gradient-based approach with surrogate gradients and densely sampled membrane observations from the BrainScaleS-2 hardware system.

Human motion prediction is a complex task as it involves forecasting variables over time on a graph of connected sensors. This is especially true in the case of few-shot learning, where we strive to forecast motion sequences for previously unseen actions based on only a few examples. Despite this, almost all related approaches for few-shot motion prediction do not incorporate the underlying graph, while it is a common component in classical motion prediction. Furthermore, state-of-the-art methods for few-shot motion prediction are restricted to motion tasks with a fixed output space meaning these tasks are all limited to the same sensor graph. In this work, we propose to extend recent works on few-shot time-series forecasting with heterogeneous attributes with graph neural networks to introduce the first few-shot motion approach that explicitly incorporates the spatial graph while also generalizing across motion tasks with heterogeneous sensors. In our experiments on motion tasks with heterogeneous sensors, we demonstrate significant performance improvements with lifts from 10.4% up to 39.3% compared to best state-of-the-art models. Moreover, we show that our model can perform on par with the best approach so far when evaluating on tasks with a fixed output space while maintaining two magnitudes fewer parameters.

Video segmentation consists of a frame-by-frame selection process of meaningful areas related to foreground moving objects. Some applications include traffic monitoring, human tracking, action recognition, efficient video surveillance, and anomaly detection. In these applications, it is not rare to face challenges such as abrupt changes in weather conditions, illumination issues, shadows, subtle dynamic background motions, and also camouflage effects. In this work, we address such shortcomings by proposing a novel deep learning video segmentation approach that incorporates residual information into the foreground detection learning process. The main goal is to provide a method capable of generating an accurate foreground detection given a grayscale video. Experiments conducted on the Change Detection 2014 and on the private dataset PetrobrasROUTES from Petrobras support the effectiveness of the proposed approach concerning some state-of-the-art video segmentation techniques, with overall F-measures of $\mathbf{0.9535}$ and $\mathbf{0.9636}$ in the Change Detection 2014 and PetrobrasROUTES datasets, respectively. Such a result places the proposed technique amongst the top 3 state-of-the-art video segmentation methods, besides comprising approximately seven times less parameters than its top one counterpart.

Scene change detection is an image processing problem related to partitioning pixels of a digital image into foreground and background regions. Mostly, visual knowledge-based computer intelligent systems, like traffic monitoring, video surveillance, and anomaly detection, need to use change detection techniques. Amongst the most prominent detection methods, there are the learning-based ones, which besides sharing similar training and testing protocols, differ from each other in terms of their architecture design strategies. Such architecture design directly impacts on the quality of the detection results, and also in the device resources capacity, like memory. In this work, we propose a novel Multiscale Cascade Residual Convolutional Neural Network that integrates multiscale processing strategy through a Residual Processing Module, with a Segmentation Convolutional Neural Network. Experiments conducted on two different datasets support the effectiveness of the proposed approach, achieving average overall $\boldsymbol{F\text{-}measure}$ results of $\boldsymbol{0.9622}$ and $\boldsymbol{0.9664}$ over Change Detection 2014 and PetrobrasROUTES datasets respectively, besides comprising approximately eight times fewer parameters. Such obtained results place the proposed technique amongst the top four state-of-the-art scene change detection methods.

Research on remote sensing image classification significantly impacts essential human routine tasks such as urban planning and agriculture. Nowadays, the rapid advance in technology and the availability of many high-quality remote sensing images create a demand for reliable automation methods. The current paper proposes two novel deep learning-based architectures for image classification purposes, i.e., the Discriminant Deep Image Prior Network and the Discriminant Deep Image Prior Network+, which combine Deep Image Prior and Triplet Networks learning strategies. Experiments conducted over three well-known public remote sensing image datasets achieved state-of-the-art results, evidencing the effectiveness of using deep image priors for remote sensing image classification.