Rural roads are vital for the transportation of persons and goods, but are vulnerable to animal-vehicle collisions, which cause substantial environmental, economic and public health impacts every year. WiFi sensing has recently emerged as a cost-effective approach to detect hazardous situations on roads, overcoming some limitations of traditional monitoring systems. However, scaling WiFi sensing to cover long road sections introduces challenges, including interference between WiFi nodes and with other protocols operating in the same Industrial, Scientific, and Medical (ISM) bands, such as IEEE 802.15.4. In this work, we present a scalable, low-cost architecture for WiFi-based monitoring of rural roads. We develop and optimize machine learning models for real-time detection and classification of pedestrians, vehicles, and animals using low-cost IoT devices, achieving over 95% accuracy and a miss rate of only 0.7%. We evaluate the models under varied conditions, including different sensor distances, light rain, and previously unseen rural locations, demonstrating robustness across environmental variations. Additionally, we conduct network simulations to assess the coexistence of WiFi sensing with IEEE 802.15.4 communications and perform LoRaWAN experiments in rural environments to validate long-range data transmission. We also release an open dataset containing all evaluated scenarios to support reproducibility and further research. Our results confirm the feasibility of the proposed system in terms of sensing accuracy, network performance, and deployment cost.

Deep learning drives a new wave in computing systems and triggers the automation of increasingly complex problems. In particular, Large Language Models (LLMs) have significantly advanced cognitive tasks, often matching or even surpassing human-level performance. However, their extensive parameters result in high computational costs and slow inference, posing challenges for deployment in resource-limited settings. Among the strategies to overcome the aforementioned challenges, pruning emerges as a successful mechanism since it reduces model size while maintaining predictive ability. In this paper, we introduce AMP: Attention Heads and MLP Pruning, a novel structured pruning method that efficiently compresses LLMs by removing less critical structures within Multi-Head Attention (MHA) and Multilayer Perceptron (MLP). By projecting the input data onto weights, AMP assesses structural importance and overcomes the limitations of existing techniques, which often fall short in flexibility or efficiency. In particular, AMP surpasses the current state-of-the-art on commonsense reasoning tasks by up to 1.49 percentage points, achieving a 30% pruning ratio with minimal impact on zero-shot task performance. Moreover, AMP also improves inference speeds, making it well-suited for deployment in resource-constrained environments. We confirm the flexibility of AMP on different families of LLMs, including LLaMA and Phi.

Deep learning stands as the modern paradigm for solving cognitive tasks. However, as the problem complexity increases, models grow deeper and computationally prohibitive, hindering advancements in real-world and resource-constrained applications. Extensive studies reveal that pruning structures in these models efficiently reduces model complexity and improves computational efficiency. Successful strategies in this sphere include removing neurons (i.e., filters, heads) or layers, but not both together. Therefore, simultaneously pruning different structures remains an open problem. To fill this gap and leverage the benefits of eliminating neurons and layers at once, we propose a new method capable of pruning different structures within a model as follows. Given two candidate subnetworks (pruned models), one from layer pruning and the other from neuron pruning, our method decides which to choose by selecting the one with the highest representation similarity to its parent (the network that generates the subnetworks) using the Centered Kernel Alignment (CKA) metric. Iteratively repeating this process provides highly sparse models that preserve the original predictive ability. Throughout extensive experiments on standard architectures and benchmarks, we confirm the effectiveness of our approach and show that it outperforms state-of-the-art layer and filter pruning techniques. At high levels of Floating Point Operations (FLOPs) reduction, most state-of-the-art methods degrade accuracy, whereas our approach either improves it or experiences only a minimal drop. Notably, on the popular ResNet56 and ResNet110, we achieve a milestone of 86.37% and 95.82% FLOPs reduction. Besides, our pruned models obtain robustness to adversarial and out-of-distribution samples and take an important step towards GreenAI, reducing carbon emissions by up to 83.31%. Overall, we believe our work opens a new chapter in pruning. Code is available at: https://github.com/NascimentoG/PruningEverything.

Sea-level rise is a well-known consequence of climate change. Several studies have estimated the social and economic impact of the increase in extreme flooding. An efficient way to mitigate its consequences is the development of a flood alert and prediction system, based on high-resolution numerical models and robust sensing networks. However, current models use various simplifying assumptions that compromise accuracy to ensure solvability within a reasonable timeframe, hindering more regular and cost-effective forecasts for various locations along the shoreline. To address these issues, this work proposes a hybrid model for multimodal data processing that combines physics-based numerical simulations, data obtained from a network of sensors, and satellite images to provide refined wave and sea-surface height forecasts, with real results obtained in a critical location within the Port of Santos (the largest port in Latin America). Our approach exhibits faster convergence than data-driven models while achieving more accurate predictions. Moreover, the model handles irregularly sampled time series and missing data without the need for complex preprocessing mechanisms or data imputation while keeping low computational costs through a combination of time encoding, recurrent and graph neural networks. Enabling raw sensor data to be easily combined with existing physics-based models opens up new possibilities for accurate extreme storm tide events forecast systems that enhance community safety and aid policymakers in their decision-making processes.

Dimensionality reduction plays an important role in computer vision problems since it reduces computational cost and is often capable of yielding more discriminative data representation. In this context, Partial Least Squares (PLS) has presented notable results in tasks such as image classification and neural network optimization. However, PLS is infeasible on large datasets, such as ImageNet, because it requires all the data to be in memory in advance, which is often impractical due to hardware limitations. Additionally, this requirement prevents us from employing PLS on streaming applications where the data are being continuously generated. Motivated by this, we propose a novel incremental PLS, named Covariance-free Incremental Partial Least Squares (CIPLS), which learns a low-dimensional representation of the data using a single sample at a time. In contrast to other state-of-the-art approaches, instead of adopting a partially-discriminative or SGD-based model, we extend Nonlinear Iterative Partial Least Squares (NI-PALS) - the standard algorithm used to compute PLS - for incremental processing. Among the advantages of this approach are the preservation of discriminative information across all components, the possibility of employing its score matrices for feature selection, and its computational efficiency. We validate CIPLS on face verification and image classification tasks, where it outperforms several other incremental dimensionality reduction techniques. In the context of feature selection, CIPLS achieves comparable results when compared to state-of-the-art techniques.

Modern convolutional networks such as ResNet and NASNet have achieved state-of-the-art results in many computer vision applications. These architectures consist of stages, which are sets of layers that operate on representations in the same resolution. It has been demonstrated that increasing the number of layers in each stage improves the prediction ability of the network. However, the resulting architecture becomes computationally expensive in terms of floating point operations, memory requirements and inference time. Thus, significant human effort is necessary to evaluate different trade-offs between depth and performance. To handle this problem, recent works have proposed to automatically design high-performance architectures, mainly by means of neural architecture search (NAS). Current NAS strategies analyze a large set of possible candidate architectures and, hence, require vast computational resources and take many GPUs days. Motivated by this, we propose a NAS approach to efficiently design accurate and low-cost convolutional architectures and demonstrate that an efficient strategy for designing these architectures is to learn the depth stage-by-stage. For this purpose, our approach increases depth incrementally in each stage taking into account its importance, such that stages with low importance are kept shallow while stages with high importance become deeper. We conduct experiments on the CIFAR and different versions of ImageNet datasets, where we show that architectures discovered by our approach achieve better accuracy and efficiency than human-designed architectures. Additionally, we show that architectures discovered on CIFAR-10 can be successfully transferred to large datasets. Compared to previous NAS approaches, our method is substantially more efficient, as it evaluates one order of magnitude fewer models and yields architectures on par with the state-of-the-art.

Features extracted with deep learning have now achieved state-of-the-art results in many tasks. However, to reuse a learned deep model, transfer learning with fine-tuning needs to be employed, which requires to re-train the whole model or part of it to extract useful features in the new domain. This step is burdensome and requires heavy computing power. Therefore, this work investigates alternatives in transfer-learning that do not involve performing fine-tuning for a model with the new domain. Namely, we explore the correlation of depth and scale in deep models, and look for the layer/scale that yields the best results for the new domain, we also explore metrics for the verification task, using locally connected convolutions to learn distance metrics. Our experiments use a model pre-trained in face identification and adapt it to the face verification task with different data, but still on the face domain. We achieve 96.65% mean accuracy on the Labeled Faces in the Wild dataset and 93.12% mean accuracy on the Youtube Faces dataset which are in the state-of-the-art.

Since Convolutional Neural Networks (ConvNets) are able to simultaneously learn features and classifiers to discriminate different categories of activities, recent works have employed ConvNets approaches to perform human activity recognition (HAR) based on wearable sensors, allowing the removal of expensive human work and expert knowledge. However, these approaches have their power of discrimination limited mainly by the large number of parameters that compose the network and the reduced number of samples available for training. Inspired by this, we propose an accurate and robust approach, referred to as Latent HyperNet (LHN). The LHN uses feature maps from early layers (hyper) and projects them, individually, onto a low dimensionality (latent) space. Then, these latent features are concatenated and presented to a classifier. To demonstrate the robustness and accuracy of the LHN, we evaluate it using four different network architectures in five publicly available HAR datasets based on wearable sensors, which vary in the sampling rate and number of activities. We experimentally demonstrate that the proposed LHN is able to capture rich information, improving the results regarding the original ConvNets. Furthermore, the method outperforms existing state-of-the-art methods, on average, by 5.1 percentage points.

The increasing popularity of approaches based on random forest in computer vision tasks is due to its simplicity and flexibility with complex data. Random forest is a set of decision trees that can be divided in two subsets according to the view of the feature descriptors provided as input: orthogonal and oblique. In the former, the feature space is separated orthogonally (axis-aligned) by a single feature at a time. In the latter, it separates the space by oriented hyperplanes, which usually provides better data modeling. This work proposes a novel oblique random forest associated with Partial Least Squares to perform the oblique split. We validate the proposed approach, referred to as oRF-PLS, on the challenge INRIA Person dataset. Experimental results demonstrate that the proposed method outperforms traditional state-of-the-art detectors. In addition, we demonstrate that PLS is a more suitable choice to build oblique random forest than SVM, being faster and producing more accurate forests.

Pedestrian detection is a well-known problem in Computer Vision. To improve detection, several feature descriptors have been proposed and combined. However, there are cases where the most powerful features fail to discriminate between false positives similar to the human body structure and actual true positives, which is a critical problem for applications such as surveillance, driving assistance and robotics. To address this issue, we propose a novel approach to combine results of distinct pedestrian detectors by reinforcing the human hypothesis. The method is able to reduce the confidence of the false positives due to the lack of spatial consensus when multiple detectors are considered. Our experimental validation, performed on three pedestrian detection benchmarks, INRIA person, ETH and Caltech pedestrian dataset, demonstrates that the proposed approach, referred to as Spatial Consensus (SC), outperforms the state-of-the-art on INRIA and ETH datasets and achieves comparable results on the Caltech dataset.