In recent years, the incorporation of early exit layers into deep neural networks has allowed inference to terminate earlier while maintaining accuracy. However, the passive decision-making involved in the these static exit placement creates a dilemma: fine-grained placement may cause high performance and energy overhead due to frequent exit layer execution, while coarse-grained placement may miss early exit opportunities. Moreover, common energy-saving techniques like adjusting processor configurations are not applicable once inference begins. To overcome these challenges and improve computation and energy efficiency, we propose LEAP, a software-hardware co-design approach. On the software side, LEAP proactively predicts exit points at runtime, reducing computation by enabling early exits without requiring every pre-placed exit layer to be executed. On the hardware side, LEAP adjusts processor settings—such as frequency and voltage—based on single or multiple predicted exits to optimize energy consumption while adhering to latency requirements. Extensive experimental results show that LEAP significantly improves efficiency. Compared to standard inference, LEAP reduces computation by up to 76.4% and saves up to 83.2% in energy. Compared to state-of-the-art early exit methods, LEAP achieves up to 27.9% less computation and 57.1% more energy savings, while maintaining similar accuracy and latency.

Single-cell RNA sequencing (scRNA-seq) provides high-throughput information about the genome-wide gene expression levels at the single-cell resolution, bringing a precise understanding on the transcriptome of individual cells. Unfortunately, the rapidly growing scRNA-seq data and the prevalence of dropout events pose substantial challenges for clustering and cell type annotation. Here, we propose a deep learning method, scHSC, that employs hard sample mining through contrastive learning for clustering scRNA-seq data. Focusing on hard samples, this approach simultaneously integrates gene expression and topological structure information between cells to improve clustering accuracy. By adjusting the weights of hard positive and hard negative samples during the iterative training process, scHSC employs an adaptive weighting strategy to integrate contrastive learning with a ZINB model for single-cell clustering tasks. Extensive experiments on 18 single-cell RNA-seq real datasets demonstrate that scHSC exhibits significant superiority in clustering performance compared to existing deep learning-based clustering methods. scHSC is implemented in Python based on the PyTorch framework. The source code and datasets are available via https://github.com/fangs25/scHSC.

Optical sensors, such as the Far Infrared (FIR) sensor, have demonstrated advantages over traditional imaging. For example, 3D reconstruction in the FIR field captures the heat distribution of a scene that is invisible to RGB, aiding various applications like gas leak detection. However, less texture information and challenges in acquiring FIR frames hinder the reconstruction process. Given that implicit neural representations (INRs) can integrate geometric information across different sensors, we propose Implicit Neural Fusion (INF) of RGB and FIR for 3D reconstruction of invisible scenes in the FIR field. Our method first obtains a neural density field of objects from RGB frames. Then, with the trained object density field, a separate neural density field of gases is optimized using limited view inputs of FIR frames. Our method not only demonstrates outstanding reconstruction quality in the FIR field through extensive experiments but also can isolate the geometric information of the invisible, offering a new dimension of scene understanding.

Surface waves investigation and monitoring (SWIM) can provide global wave spectra, but under small sea conditions, the presence of parasitic peaks at low wavenumbers, surfboard effects, and residual speckle noise lead to performance degradation of SWIM wave height spectrum products. To reduce the impacts of the above factors on the SWIM wave height spectrum, in this article, a convolution neural network (CNN) method based on BU-Net is proposed for calibrating SWIM omnidirectional wave height spectra with buoy measurements under sea states (wind wave mainly/swell mainly) and sea surface conditions (wind speed from 9 to 19 m/s, and significant wave height (SWH) from 0.8 to 3.4–4.2 m). The calibration results show that the impact of the above factors on the SWIM omnidirectional wave height spectrum can be corrected. The correlation coefficients between the corrected SWIM beams 6°, 8°, and 10° and the buoy mean omnidirectional wave height spectrum are all greater than 0.90, and the relative error of the peak wavenumber is within 10%. The relative error of the integrated energy is mostly less than 20%. In addition, the performance of spectral integration parameters (effective wave height$H_{s}$, and energy wave period$T_{m-10}$) of each spectral beam of SWIM has been verified using Meteo-France WAve Model (MFWAM) reanalysis data. The validation results show that RMSE of$H_{s}$and$T_{m-10}$, for the corrected SWIM beam 6° (8°, 10°) under wind wave sea conditions are 0.31 m (0.32, 0.25 m) and 0.50 s (0.51, 0.49 s), respectively; those for swell cases are 0.16 m (0.16, 0.13 m) and 0.87 s (0.77, 0.72 s), respectively.

By adding exiting layers to the deep learning networks, early exit can terminate the inference earlier with accurate results. However, the passive decision-making of whether to exit or continue the next layer has to go through every pre-placed exiting layer until it exits. In addition, it is hard to adjust the configurations of the computing platforms alongside the inference proceeds. By incorporating a low-cost prediction engine, we propose a Predictive Exit framework for computation- and energy-efficient deep learning applications. Predictive Exit can forecast where the network will exit (i.e., establish the number of remaining layers to finish the inference), which effectively reduces the network computation cost by exiting on time without running every pre-placed exiting layer. Moreover, according to the number of remaining layers, proper computing configurations (i.e., frequency and voltage) are selected to execute the network to further save energy. Extensive experimental results demonstrate that Predictive Exit achieves up to 96.2% computation reduction and 72.9% energy-saving compared with classic deep learning networks; and 12.8% computation reduction and 37.6% energy-saving compared with the early exit under state-of-the-art exiting strategies, given the same inference accuracy and latency.

Deep learning approaches, such as convolution neural networks (CNNs), have achieved tremendous success in versatile applications. However, one of the challenges to deploy the deep learning models on resource-constrained systems is its huge energy cost. As a dynamic inference approach, early exit adds exiting layers to the networks, which can terminate the inference earlier with accurate results to save energy. The current passive decision-making for energy regulation of early exit cannot adapt to ongoing inference status, varying inference workloads, and timing constraints, let alone guide the reasonable configuration of the computing platforms alongside the inference proceeds for potential energy saving. In this paper, we propose an Energy Efficient Neural Networks (EENet), which introduces a plug-in module to the state-of-the-art networks by incorporating run-time power management. Within each inference, we establish prediction of where the network will exit and adjust computing configurations (i.e., frequency and voltage) accordingly over a small timescale. Considering multiple inferences over a large timescale, we provide frequency and voltage calibration advice, given inference workloads and timing constraints. Finally, the dynamic voltage and frequency scaling (DVFS) governor configures voltage and frequency to execute the network according to the prediction and calibration. Extensive experimental results demonstrate that EENet achieves up to 63.8% energy-saving compared with classic deep learning networks and 21.5% energy-saving compared with the early exit under state-of-the-art exiting strategies, together with improved timing performance.

Dimension reduction (DR) plays an important role in single-cell RNA sequencing (scRNA-seq), such as data interpretation, visualization and other downstream analysis. A desired DR method should be applicable to various application scenarios, including identifying cell types, preserving the inherent structure of data and handling with batch effects. However, most of the existing DR methods fail to accommodate these requirements simultaneously, especially removing batch effects. In this paper, we develop a novel structure-preserved dimension reduction (SPDR) method using intra- and inter-batch triplets sampling. The constructed triplets jointly consider each anchor's mutual nearest neighbors from inter-batch, k-nearest neighbors from intra-batch and randomly selected cells from the whole data, which capture higher order structure information and meanwhile account for batch information of the data. Then we minimize a robust loss function for the chosen triplets to obtain a structure-preserved and batch-corrected low-dimensional representation. Comprehensive evaluations show that SPDR outperforms other competing DR methods, such as INSCT, IVIS, Trimap, Scanorama, scVI and UMAP, in removing batch effects, preserving biological variation, facilitating visualization and improving clustering accuracy. Besides, the two-dimensional (2D) embedding of SPDR presents a clear and authentic expression pattern, and can guide researchers to determine how many cell types should be identified. Furthermore, SPDR is robust to complex data characteristics (such as down-sampling, duplicates and outliers) and varying hyperparameter settings. We believe that SPDR will be a valuable tool for characterizing complex cellular heterogeneity.

Clustering and cell type classification are a vital step of analyzing scRNA-seq data to reveal the complexity of the tissue (e.g. the number of cell types and the transcription characteristics of the respective cell type). Recently, deep learning-based single-cell clustering algorithms become popular since they integrate the dimensionality reduction with clustering. But these methods still have unstable clustering effects for the scRNA-seq datasets with high dropouts or noise. In this study, a novel single-cell RNA-seq deep embedding clustering via convolutional autoencoder embedding and soft K-means (scCAEs) is proposed by simultaneously learning the feature representation and clustering. It integrates the deep learning with convolutional autoencoder to characterize scRNA-seq data and proposes a regularized soft K-means algorithm to cluster cell populations in a learned latent space. Next, a novel constraint is introduced to the clustering objective function to iteratively optimize the clustering results, and more importantly, it is theoretically proved that this objective function optimization ensures the convergence. Moreover, it adds the reconstruction loss to the objective function combining the dimensionality reduction with clustering to find a more suitable embedding space for clustering. The proposed method is validated on a variety of datasets, in which the number of clusters in the mentioned datasets ranges from 4 to 46, and the number of cells ranges from 90 to 30 302. The experimental results show that scCAEs is superior to other state-of-the-art methods on the mentioned datasets, and it also keeps the satisfying compatibility and robustness. In addition, for single-cell datasets with the batch effects, scCAEs can ensure the cell separation while removing batch effects.

Developments of single-cell RNA sequencing (scRNA-seq) technologies have enabled biological discoveries at the single-cell resolution with high throughput. However, large scRNA-seq datasets always suffer from massive technical noises, including batch effects and dropouts, and the dropout is often shown to be batch-dependent. Most existing methods only address one of the problems, and we show that the popularly used methods failed in trading off batch effect correction and dropout imputation. Here, inspired by the idea of causal inference, we propose a novel propensity score matching method for scRNA-seq data (scPSM) by borrowing information and taking the weighted average from similar cells in the deep sequenced batch, which simultaneously removes the batch effect, imputes dropout and denoises data in the entire gene expression space. The proposed method is testified on two simulation datasets and a variety of real scRNA-seq datasets, and the results show that scPSM is superior to other state-of-the-art methods. First, scPSM improves clustering accuracy and mixes cells of the same type, suggesting its ability to keep cell type separation while correcting for batch. Besides, using the scPSM-integrated data as input yields results free of batch effects or dropouts in the differential expression analysis. Moreover, scPSM not only achieves ideal denoising but also preserves real biological structure for downstream gene-based analyses. Furthermore, scPSM is robust to hyperparameters and small datasets with a few cells but enormous genes. Comprehensive evaluations demonstrate that scPSM jointly provides desirable batch effect correction, imputation and denoising for recovering the biologically meaningful expression in scRNA-seq data.

Three-phase isolated matrix rectifier (IMR) is a kind of single-stage converters without dc-bus. IMR has the advantages of high transmission efficiency and flexible adjusta-bility of output voltage, which is very suitable for electrical vehicle (EV) chargers. But IMR generates harmonic pollution to the power grid because of its nonlinearity property, which in turn brings disturbances to the output side of IMR. This paper proposes an improved plug-in repetitive controller for the harmonic compensation to three-phase IMR. The proposed controller, which consists of a repetitive controller with a parallel forward path and a conventional PI controller can compensate the harmonic in power grid and attenuate the harmonic currents. For IMR operating at the power frequency, a fractional delay(FD) filter is utilized to approximate the characteristic of the delay and improve the steady-state tracking performance of the proposed controller. Simulation results verify that the proposed control scheme can compensate the harmonics of IMR effectively.

As the compensation and extension of static wireless sensor nodes, wearable helmets can build regional early warning network of personnel security. In this paper, a wearable helmet is presented towards early warning of leaking toxic gas in large-scale petrochemical plants for protecting the lives and safety of workers better.