Representational Capacity Research Articles

Deep learning-enabled exploration of global spectral features for photosynthetic capacity estimation

Spectral analysis is a widely used method for monitoring photosynthetic capacity. However, vegetation indices-based linear regression exhibits insufficient utilization of spectral information, while full spectra-based traditional machine learning has limited representational capacity (partial least squares regression) or uninterpretable (convolution). In this study, we proposed a deep learning model with enhanced interpretability based on attention and vegetation indices calculation for global spectral feature mining to accurately estimate photosynthetic capacity. We explored the ability of the model to uncover the optimal vegetation indices form and illustrated its advantage over traditional methods. Furthermore, we verified that power compression was an effective method for spectral processing. Our results demonstrated that the new model outperformed traditional models, with an increase in the coefficient of determination (R2) of 0.01-0.43 and a decrease in root mean square error (RMSE) of 1.58-12.48 μmol m-2 s-1. The best performance of our model in R2 was 0.86 and 0.81 for maximum carboxylation rate (Vcmax) and maximum electron transport rate (Jmax), respectively. The photosynthesis-sensitive spectral bands identified by our model were predominantly in the visible range. The most sensitive vegetation indices form discovered by our model was Reflectancenear−infrared+Reflectancegreen/blueReflectancenear−infrared×Reflectancered. Our model provides a new framework for interpreting spectral information and accurately estimating photosynthetic capacity.

Semi-supervised learning based on distance measurement and augmentations comparison for SAR image recognition

ABSTRACT In recent years, deep learning has significantly improved the performance of Synthetic Aperture Radar Automatic Target Recognition (SAR-ATR), but the lack of large-scale labelled training data limits the further development of deep neural networks in this field. Traditionally, annotating SAR data requires manual labour by experts with specialized domain knowledge, which is costly and makes it difficult to obtain a large amount of labelled data. Semi-supervised learning (SSL) algorithms can effectively leverage unlabelled data for training, improving model performance. However, improper use of pseudo-labels can lead to error accumulation and result in performance degradation. This paper proposes an SSL algorithm for SAR image recognition that employs the following strategies to mitigate error accumulation and enhance performance: First, by combining pseudo-labelling and consistency-regularization methods, unlabelled data with different augmentations are compared in terms of classification predictions and feature vectors, improving the model representation capacity. Second, a weighted distance metric combining labelled and unlabelled data is designed to estimate model loss, guiding the model towards intra-class aggregation and inter-class dispersion in terms of data distribution. Experiments conducted on the MSTAR and OpenSARShip datasets demonstrate that the algorithm achieves superior performance in limited labelled sample scenarios, with an accuracy of over 99% when only 30 labelled data points (approximately 10%) per class are available, and an accuracy of over 85% with extremely scarce samples (only 5 labelled samples per class) on the MSTAR dataset.

MMSG-DTA: A Multimodal, Multiscale Model Based on Sequence and Graph Modalities for Drug-Target Affinity Prediction.

Drug-Target Affinity (DTA) prediction is a cornerstone of drug discovery and development, providing critical insights into the intricate interactions between candidate drugs and their biological targets. Despite its importance, existing methodologies often face significant limitations in capturing comprehensive global features from molecular graphs, which are essential for accurately characterizing drug properties. Furthermore, protein feature extraction is predominantly restricted to 1D amino acid sequences, which fail to adequately represent the spatial structures and complex functional regions of proteins. These shortcomings impede the development of models capable of fully elucidating the mechanisms underlying drug-target interactions. To overcome these challenges, we propose a multimodal, multiscale model based on Sequence and Graph Modalities for Drug-Target Affinity (MMSG-DTA) Prediction. The model combines graph neural networks with Transformers to effectively capture both local node-level features and global structural features of molecular graphs. Additionally, a graph-based modality is employed to improve the extraction of protein features from amino acid sequences. To further enhance the model's performance, an attention-based feature fusion module is incorporated to integrate diverse feature types, thereby strengthening its representation capacity and robustness. We evaluated MMSG-DTA on three public benchmark data sets─Davis, KIBA, and Metz─and the experimental results demonstrate that the proposed model outperforms several state-of-the-art methods in DTA prediction. These findings highlight the effectiveness of MMSG-DTA in advancing the accuracy and robustness of drug-target interaction modeling.

Perishable material choice indicates symbolic and representational capacities.

The absence of symbolic material cultural objects in the archaeological record does not prove absence of symbolic cognition. Sometimes perishable materials are selected for symbolic roles, for practical concerns or to indicate a temporary condition. Also some symbolic functions may predate the use of durable materials. Finally, child play and artisan experimentations usually involve cheap and perishable materials. These are symbolic and representational activities that do not leave a material trace.

Peningkatan kemampuan representasi matematis siswa melalui model pembelajaran berbasis inkuiri berbantuan software geogebra

The students' inadequate mathematical representation abilities, as seen by their trigonometry learning outcomes, are the driving force for this investigation. This study aims to determine whether using Geogebra software in combination with an inquiry-based learning approach may improve students' mathematical representation skills. Of the seventy students enrolled at SMA Negeri 1 Ketambe, twenty-five students from class X IPA 1 and twenty-five students from class X IPA 2 were assigned to the experimental and control courses, respectively. The second class used a quasi-experimental technique with a pre- and post-test design. Tests used as research instruments. The data analysis methods that were employed were the N-Gain Test, Normality Test, Homogeneity Test, and T-Test. The results of the study on hypothesis testing indicate that an inquiry-based learning model can enhance students' capacity for mathematical representation, and that students who employed this strategy saw higher gains in this capacity than those who received conventional instruction.

Harmonized system code classification using supervised contrastive learning with sentence BERT and multiple negative ranking loss

PurposeInternational trade transactions, extracted from customs declarations, include several fields, among which the product description and the product category are the most important. The product category, also referred to as the Harmonised System Code (HS code), serves as a pivotal component for determining tax rates and administrative purposes. A predictive tool designed for product categories or HS codes becomes an important resource aiding traders in their decision to choose a suitable code. This tool is instrumental in preventing misclassification arising from the ambiguities present in product nomenclature, thus mitigating the challenges associated with code interpretation. Moreover, deploying this tool would streamline the validation process for government officers dealing with extensive transactions, optimising their workload and enhancing tax revenue collection within this domain.Design/methodology/approachThis study introduces a methodology focused on the generation of sentence embeddings for trade transactions, employing Sentence BERT (SBERT) framework in conjunction with the Multiple Negative Ranking (MNR) Loss function following a contrastive learning paradigm. The procedure involves the construction of pairwise samples, including anchors and positive transactions. The proposed method is evaluated using two publicly available real-world datasets, specifically the India Import 2016 and United States Import 2018 datasets, to fine-tune the SBERT model. Several configurations involving pooling strategies, loss functions, and training parameters are explored within the experimental setup. The acquired representations serve as inputs for traditional machine learning algorithms employed in predicting the product categories within trade transactions.FindingsEncoding trade transactions utilising SBERT with MNR loss facilitates the creation of enhanced embeddings that exhibit improved representational capacity. These fixed-length embeddings serve as adaptable inputs for training machine learning models, including support vector machine (SVM) and random forest, intended for downstream tasks of HS code classification. Empirical evidence supports the superior performance of our proposed approach compared to fine-tuning transformer-based models in the domain of trade transaction classification.Originality/valueOur approach generates more representative sentence embeddings by creating the network architectures from scratch with the SBERT framework. Instead of exploiting a data augmentation method generally used in contrastive learning for measuring the similarity between the samples, we arranged positive samples following a supervised paradigm and determined loss through distance learning metrics. This process involves continuous updating of the Siamese or bi-encoder network to produce embeddings derived from commodity transactions. This strategy aims to ensure that similar concepts of transactions within the same class converge closer within the feature embedding space, thereby improving the performance of downstream tasks.

New Modes of Materiality in the Angoysses douloureuses qui procedent d’amours (1538) by Hélisenne de Crenne

This article moves outside the conventional methodologies of visual and literary studies and material bibliography to produce a reading of the Angoysses douloureuses that demonstrates how the text and the material object of the book can be read together to recreate inscribed reading practices, including those of women readers in addition to the reading communities (of women) evoked by the text alone. The narrative of the Angoysses douloureuses displays a preoccupation with representing the materiality of the text throughout the work by means of letters exchanged between the lovers, a single, continuous narrative written by the heroine Helisenne, a manuscript book recuperated from beside her body, and a printed book whose circulation in Paris was sanctioned by Jupiter. This interest in the text’s material form goes beyond representational issues to include woodcuts and chapter headings, showing the capacity of representation to collaborate with the materiality of the printed book in producing meaning. These material elements variously show us how the text seeks to conform to the convention of publishing women’s writing posthumously while also indicating how materiality might both anticipate and shape reading practices. They also allow us to read the material book as one that delineates the problematics of a female-authored work circulating in print at this early point in the sixteenth century.

SiamRhic: Improved Cross-Correlation and Ranking Head-Based Siamese Network for Object Tracking in Remote Sensing Videos

Object tracking in remote sensing videos is a challenging task in computer vision. Recent advances in deep learning have sparked significant interest in tracking algorithms based on Siamese neural networks. However, many existing algorithms fail to deliver satisfactory performance in complex scenarios due to challenging conditions and limited computational resources. Thus, enhancing tracking efficiency and improving algorithm responsiveness in complex scenarios are crucial. To address tracking drift caused by similar objects and background interference in remote sensing image tracking, we propose an enhanced Siamese network based on the SiamRhic architecture, incorporating a cross-correlation and ranking head for improved object tracking. We first use convolutional neural networks for feature extraction and integrate the CBAM (Convolutional Block Attention Module) to enhance the tracker’s representational capacity, allowing it to focus more effectively on the objects. Additionally, we replace the original depth-wise cross-correlation operation with asymmetric convolution, enhancing both speed and performance. We also introduce a ranking loss to reduce the classification confidence of interference objects, addressing the mismatch between classification and regression. We validate the proposed algorithm through experiments on the OTB100, UAV123, and OOTB remote sensing datasets. Specifically, SiamRhic achieves success, normalized precision, and precision rates of 0.533, 0.786, and 0.812, respectively, on the OOTB benchmark. The OTB100 benchmark achieves a success rate of 0.670 and a precision rate of 0.892. Similarly, in the UAV123 benchmark, SiamRhic achieves a success rate of 0.621 and a precision rate of 0.823. These results demonstrate the algorithm’s high precision and success rates, highlighting its practical value.

Pruning networks at once via nuclear norm-based regularization and bi-level optimization

Most network pruning methods focus on identifying redundant channels from pre-trained models, which is inefficient due to its three-step process: pre-training, pruning and fine-tuning, and reconfiguration. In this paper, we propose a pruning-from-scratch framework that unifies these processes into a single approach. We introduce nuclear norm-based regularization to maintain the representational capacity of large networks during pruning. Combining this with MACs-based regularization enhances the performance of the pruned network at the target compression rate. Our bi-level optimization approach simultaneously improves pruning efficiency and representation capacity. Experimental results show that our method achieves 75.4% accuracy on ImageNet without a pre-trained network, using only 41% of the original model’s computational cost. It also attains 0.5% higher performance in compressing the SSD network for object detection. Furthermore, we analyze the effects of nuclear norm-based regularization.

VNN-DM: a vector neural network-based detection model for time synchronization attacks in park-level energy internet

Micro phasor measurement units (µPMUs) provide high-precision voltage and current phasor data, allowing real-time state estimation and fault detection, which are critical for the stability and reliability of modern power systems. However, their reliance on accurate time synchronization makes them vulnerable to time synchronization attacks (TSAs), which can disrupt grid monitoring and control by corrupting µPMU data. Addressing these vulnerabilities is essential to ensure the secure and resilient operation of smart grids and energy internet technologies. To address these challenges, intelligent detection methods are essential. Therefore, this paper proposes a µPMU measurement data TSA detection model based on vector neural networks (VNNs). This model initially employs a vector neural network to process raw data, effectively extracting and analyzing temporal features. During the same time, a capsule network is employed to classify these temporal features. On this basis, a reconstruction network is used to verify the representational capacity of the model. Simulations based on µPMU measurement data demonstrate that the model exhibits excellent detection capacity in various performance metrics, underscoring its precision and robustness.

Noise-Aware Network with Shared Channel-Attention Encoder and Joint Constraint for Noisy Speech Separation

Recently, significant progress has been made in the end-to-end single-channel speech separation in clean environments. For noisy speech separation, existing research mainly uses deep neural networks to implicitly process the noise in speech signals, which does not fully utilize the impact of noise reconstruction errors on network training. We propose a lightweight noise-aware network with shared channel-attention encoder and joint constraint, named NSCJnet, which aims to improve the speech separation system performance in noisy environments. Firstly, to reduce network parameters, the model uses a parameter sharing channel attention encoder to convert noisy speech signals into a feature space. In addition, the channel attention layer (CAlayer) in encoder enhances the network's representational capacity and separation performance in noisy environments by calculating different weights of the filters in the convolution. Secondly, to make the network converge quickly, we regard noise as an estimation target of equal significance to speech, which compel the network to separate residual noise from the estimated speech, effectively suppressing lingering noise within the speech signal. Furthermore, by integrating a multi-resolution frequency constraint into the time domain loss, we introduce a weighted time-frequency joint loss constraint, empowering the network to acquire information across both dimensions to conducive to separating mixed speech with noise.It automatically strengthens important features for separation and suppresses unimportant ones during the learning process. The results on the noisy WHAM! dataset and the noisy Libri2Mix dataset show that our method has less computational complexity, and outperforms some advanced methods in various speech quality and intelligibility metrics.

Multi‐View Self‐Supervised Auxiliary Task for Few‐Shot Remote Sensing Classification

ABSTRACTIn the past few years, the swift advancement of remote sensing technology has greatly promoted its widespread application in the agricultural field. For example, remote sensing technology is used to monitor the planting area and growth status of crops, classify crops, and detect agricultural disasters. In these applications, the accuracy of image classification is of great significance in improving the efficiency and sustainability of agricultural production. However, many of the existing studies primarily rely on contrastive self‐supervised learning methods, which come with certain limitations such as complex data construction and a bias towards invariant features. To address these issues, additional techniques like knowledge distillation are often employed to optimize the learned features. In this article, we propose a novel approach to enhance feature acquisition specific to remote sensing images by introducing a classification‐based self‐supervised auxiliary task. This auxiliary task involves performing image transformation self‐supervised learning tasks directly on the remote sensing images, thereby improving the overall capacity for feature representation. In this work, we design a texture fading reinforcement auxiliary task to reinforce texture features and color features that are useful for distinguishing similar classes of remote sensing. Different auxiliary tasks are fused to form a multi‐view self‐supervised auxiliary task and integrated with the main task to optimize the model training in an end‐to‐end manner. The experimental results on several popular few‐shot remote sensing image datasets validate the effectiveness of the proposed method. The performance better than many advanced algorithms is achieved with a more concise structure.

Lightweight UAV Small Target Detection and Perception Based on Improved YOLOv8-E

Traditional unmanned aerial vehicle (UAV) detection methods struggle with multi-scale variations during flight, complex backgrounds, and low accuracy, whereas existing deep learning detection methods have high accuracy but high dependence on equipment, making it difficult to detect small UAV targets efficiently. To address the above challenges, this paper proposes an improved lightweight high-precision model, YOLOv8-E (Enhanced YOLOv8), for the fast and accurate detection and identification of small UAVs in complex environments. First, a Sobel filter is introduced to enhance the C2f module to form the C2f-ESCFFM (Edge-Sensitive Cross-Stage Feature Fusion Module) module, which achieves higher computational efficiency and feature representation capacity while preserving detection accuracy as much as possible by fusing the SobelConv branch for edge extraction and the convolution branch to extract spatial information. Second, the neck network is based on the HSFPN (High-level Screening-feature Pyramid Network) architecture, and the CAA (Context Anchor Attention) mechanism is introduced to enhance the semantic parsing of low-level features to form a new CAHS-FPN (Context-Augmented Hierarchical Scale Feature Pyramid Network) network, enabling the fusion of deep and shallow features. This improves the feature representation capability of the model, allowing it to detect targets of different sizes efficiently. Finally, the optimized detail-enhanced convolution (DEConv) technique is introduced into the head network, forming the LSCOD (Lightweight Shared Convolutional Object Detector Head) module, enhancing the generalization ability of the model by integrating a priori information and adopting the strategy of shared convolution. This ensures that the model enhances its localization and classification performance without increasing parameters or computational costs, thus effectively improving the detection performance of small UAV targets. The experimental results show that compared with the baseline model, the YOLOv8-E model achieved (mean average precision at IoU = 0.5) an mAP@0.5 improvement of 6.3%, reaching 98.4%, whereas the model parameter scale was reduced by more than 50%. Overall, YOLOv8-E significantly reduces the demand for computational resources while ensuring high-precision detection.

Neural Methods for Amortized Inference

Simulation-based methods for statistical inference have evolved dramatically over the past 50 years, keeping pace with technological advancements. The field is undergoing a new revolution as it embraces the representational capacity of neural networks, optimization libraries, and graphics processing units for learning complex mappings between data and inferential targets. The resulting tools are amortized, in the sense that, after an initial setup cost, they allow rapid inference through fast feed-forward operations. In this article we review recent progress in the context of point estimation, approximate Bayesian inference, summary-statistic construction, and likelihood approximation. We also cover software and include a simple illustration to showcase the wide array of tools available for amortized inference and the benefits they offer over Markov chain Monte Carlo methods. The article concludes with an overview of relevant topics and an outlook on future research directions.

SIFusion: Lightweight infrared and visible image fusion based on semantic injection.

The objective of image fusion is to integrate complementary features from source images to better cater to the needs of human and machine vision. However, existing image fusion algorithms predominantly focus on enhancing the visual appeal of the fused image for human perception, often neglecting their impact on subsequent high-level visual tasks, particularly the processing of semantic information. Moreover, these fusion methods that incorporate downstream tasks tend to be overly complex and computationally intensive, which is not conducive to practical applications. To address these issues, a lightweight infrared and visible light image fusion method known as SIFusion, which is based on semantic injection, is proposed in this paper. This method employs a semantic-aware branch to extract semantic feature information, and then integrates these features into the fused features through a Semantic Injection Module (SIM) to meet the semantic requirements of high-level visual tasks. Furthermore, to simplify the complexity of the fusion network, this method introduces an Edge Convolution Module (ECB) based on structural reparameterization technology to enhance the representational capacity of the encoder and decoder. Extensive experimental comparisons demonstrate that the proposed method performs excellently in terms of visual appeal and advanced semantics, providing satisfactory fusion results for subsequent high-level visual tasks even in challenging scenarios.

Self-Supervised Time Series Representation Learning via Cross Reconstruction Transformer.

Since labeled samples are typically scarce in real-world scenarios, self-supervised representation learning in time series is critical. Existing approaches mainly employ the contrastive learning framework, which automatically learns to understand similar and dissimilar data pairs. However, they are constrained by the request for cumbersome sampling policies and prior knowledge of constructing pairs. Also, few works have focused on effectively modeling temporal-spectral correlations to improve the capacity of representations. In this article, we propose the cross reconstruction transformer (CRT) to solve the aforementioned issues. CRT achieves time series representation learning through a cross-domain dropping-reconstruction task. Specifically, we obtain the frequency domain of the time series via the fast Fourier transform (FFT) and randomly drop certain patches in both time and frequency domains. Dropping is employed to maximally preserve the global context while masking leads to the distribution shift. Then a Transformer architecture is utilized to adequately discover the cross-domain correlations between temporal and spectral information through reconstructing data in both domains, which is called Dropped Temporal-Spectral Modeling. To discriminate the representations in global latent space, we propose instance discrimination constraint (IDC) to reduce the mutual information between different time series samples and sharpen the decision boundaries. Additionally, a specified curriculum learning (CL) strategy is employed to improve the robustness during the pretraining phase, which progressively increases the dropping ratio in the training process. We conduct extensive experiments to evaluate the effectiveness of the proposed method on multiple real-world datasets. Results show that CRT consistently achieves the best performance over existing methods by 2%-9%. The code is publicly available at https://github.com/BobZwr/Cross-Reconstruction-Transformer.

Active inference and deep generative modeling for cognitive ultrasound.

Ultrasound has the unique potential to offer access to medical imaging to anyone, everywhere. Devices have become ultra-portable and cost-effective, akin to the stethoscope. Nevertheless, and despite many advances, ultrasound image quality and diagnostic efficacy are still highly operator- and patient-dependent. In difficult-to-image patients, image quality is often insufficient for reliable diagnosis. In this paper, we put forth the idea that ultrasound imaging systems can be recast as information-seeking agents that engage in reciprocal interactions with their anatomical environment. Such agents autonomously adapt their transmit-receive sequences to fully personalize imaging and actively maximize information gain in-situ. To that end, we will show that the sequence of pulse-echo experiments that an ultrasound system performs can be interpreted as a perception-action loop: the action is the data acquisition, probing tissue with acoustic waves and recording reflections at the detection array, and perception is the inference of the anatomical and or functional state, potentially including associated diagnostic quantities. We then equip systems with a mechanism to actively reduce uncertainty and maximize diagnostic value across a sequence of experiments, treating action and perception jointly using Bayesian inference given generative models of the environment and action-conditional pulse-echo observations. Since the representation capacity of the generative models dictates both the quality of inferred anatomical states and the effectiveness of inferred sequences of future imaging actions, we will be greatly leveraging the enormous advances in deep generative modelling (generative AI), that are currently disrupting many fields and society at large. Finally, we show some examples of cognitive, closed-loop, ultrasound systems that perform active beamsteering and adaptive scanline selection, based on deep generative models that track anatomical belief states.

Frequency-domain sound field from the perspective of band-limited functions.

A model that approximates the sound field well is useful in various fields, such as acoustic signal processing and numerical simulation. We have proposed an effective model in which the wideband instantaneous sound field is regarded as an element of a spherically band-limited function space, using the reproducing kernel of that space. In this paper, the frequency-domain sound field is regarded as an element of some band-limited function space, and a representation of the field as a linear combination of the reproducing kernel in that space is proposed. This model has the strongest representational capacity of all function systems when we know only the sound pressure information at arbitrary positions. The proposed model can be considered a generalization of the existing three-dimensional sound field model using the reproducing kernel of the solution space of the Helmholtz equation to the spatial dimension. One of the advantages of capturing the frequency-domain sound field in this way is the simplicity achieved for the estimation formula of the wavenumber spectrum. Two numerical simulations were conducted to validate the proposed methods.

Predicting future evapotranspiration based on remote sensing and deep learning

Study regionThe watersheds of the four flux sites in the United States were selected as the study areas for this research. Study focusThis study validates the efficiency of Convolutional Long Short-Term Memory Network (ConvLSTM) models for site-scale ETa prediction. We enhanced the ConvLSTM model by adding a Spatial Pyramid Pooling module (SPPM) and a Multi-head Self-Attention Module (MSA-Module), creating the Multi-head Self-Attention ConvLSTM (MSA-ConvLSTM) model, which we applied to predicting regional-scale actual evapotranspiration (ETa). This study aims to investigate whether the MSA-ConvLSTM model can enhance the accuracy of predicting regional-scale ETa, considering multiple feature variables. Furthermore, we evaluated different performance indicators, discussed possible reasons for errors in regional ETa prediction, and conducted sensitivity analysis of the model characteristics. New hydrological insights for the regionThe MSA-ConvLSTM model accurately predicts the future state of ETa. The average R2 was 0.81, which is 11.6 % and 5.5 % higher than those of the ConvLSTM and Self-Attention ConvLSTM (SA-ConvLSTM) models, respectively. The average RMSE is 11.94 mm/m, which is 21.5 % and 13.7 % lower than ConvLSTM and SA-ConvLSTM, respectively. The average MAE is 9.46 mm/m, which is 21.3 % and 13 % lower than ConvLSTM and SA-ConvLSTM, respectively. Incorporating of a multi-head self-attention module enhances the model’s capacity for comprehensive understanding of input data features. This improvement allows the model to better adapt to feature relationships at varying scales and angles, enhancing its representational capacity and enabling effective adaptation to complex environmental changes.

Ponte: Represent Totally Binary Neural Network Toward Efficiency

In the quest for computational efficiency, binary neural networks (BNNs) have emerged as a promising paradigm, offering significant reductions in memory footprint and computational latency. In traditional BNN implementation, the first and last layers are typically full-precision, which causes higher logic usage in field-programmable gate array (FPGA) implementation. To solve these issues, we introduce a novel approach named Ponte (Represent Totally Binary Neural Network Toward Efficiency) that extends the binarization process to the first and last layers of BNNs. We challenge the convention by proposing a fully binary layer replacement that mitigates the computational overhead without compromising accuracy. Our method leverages a unique encoding technique, Ponte::encoding, and a channel duplication strategy, Ponte::dispatch, and Ponte::sharing, to address the non-linearity and capacity constraints posed by binary layers. Surprisingly, all of them are back-propagation-supported, which allows our work to be implemented in the last layer through extensive experimentation on benchmark datasets, including CIFAR-10 and ImageNet. We demonstrate that Ponte not only preserves the integrity of input data but also enhances the representational capacity of BNNs. The proposed architecture achieves comparable, if not superior, performance metrics while significantly reducing the computational demands, thereby marking a step forward in the practical deployment of BNNs in resource-constrained environments.