This paper introduces a novel framework for designing efficient neural network architectures specifically tailored to tiny machine learning (TinyML) platforms. By leveraging large language models (LLMs) for neural architecture search (NAS), a vision transformer (ViT)-based knowledge distillation (KD) strategy, and an explainability module, the approach strikes an optimal balance between accuracy, computational efficiency, and memory usage. The LLM-guided search explores a hierarchical search space, refining candidate architectures through Pareto optimization based on accuracy, multiply-accumulate operations (MACs), and memory metrics. The best-performing architectures are further fine-tuned using logits-based KD with a pre-trained ViT-B/16 model, which enhances generalization without increasing model size. Evaluated on the CIFAR-100 dataset and deployed on an STM32H7 microcontroller (MCU), the three proposed models, LMaNet-Elite, LMaNet-Core, and QwNet-Core, achieve accuracy scores of 74.50%, 74.20% and 73.00%, respectively. All three models surpass current state-of-the-art (SOTA) models, such as MCUNet-in3/in4 (69.62% / 72.86%) and XiNet (72.27%), while maintaining a low computational cost of less than 100 million MACs and adhering to the stringent 320 KB static random-access memory (SRAM) constraint. These results demonstrate the efficiency and performance of the proposed framework for TinyML platforms, underscoring the potential of combining LLM-driven search, Pareto optimization, KD, and explainability to develop accurate, efficient, and interpretable models. This approach opens new possibilities in NAS, enabling the design of efficient architectures specifically suited for TinyML.
Over the past decade, recommender systems have experienced a surge in popularity. Despite notable progress, they grapple with challenging issues, such as high data dimensionality and sparseness. Representing users and items as low-dimensional embeddings learned via neural networks has become a leading solution. However, while recent studies show promising results, many approaches rely on complex architectures or require content data, which may not always be available. This paper presents Interact2Vec, a novel neural network-based model that simultaneously learns distributed embeddings for users and items while demanding only implicit feedback. The model employs state-of-the-art strategies that natural language processing models commonly use to optimize the training phase and enhance the final embeddings. Two types of experiments were conducted regarding the extrinsic and intrinsic quality of the model. In the former, we benchmarked the recommendations generated by Interact2Vec's embeddings in a top-$N$ ranking problem, comparing them with six other recommender algorithms. The model achieved the second or third-best results in 30\% of the datasets, being competitive with other recommenders, and has proven to be very efficient with an average training time reduction of 274\% compared to other embedding-based models. Later, we analyzed the intrinsic quality of the embeddings through similarity tables. Our findings suggest that Interact2Vec can achieve promising results, especially on the extrinsic task, and is an excellent embedding-generator model for scenarios of scarce computing resources, enabling the learning of item and user embeddings simultaneously and efficiently.
In neural network compression, most current methods reduce unnecessary parameters by measuring importance and redundancy. To augment already highly optimized existing solutions, we propose linearity-based compression as a novel way to reduce weights in a neural network. It is based on the intuition that with ReLU-like activation functions, neurons that are almost always activated behave linearly, allowing for merging of subsequent layers. We introduce the theory underlying this compression and evaluate our approach experimentally. Our novel method achieves a lossless compression down to 1/4 of the original model size in over the majority of tested models. Applying our method on already importance-based pruned models shows very little interference between different types of compression, demonstrating the option of successful combination of techniques. Overall, our work lays the foundation for a new type of compression method that enables smaller and ultimately more efficient neural network models.
Learning to control high-speed objects in the real world remains a challenging frontier in robotics. Table tennis serves as an ideal testbed for this problem, demanding both rapid interception of fast-moving balls and precise adjustment of their trajectories. This task presents two fundamental challenges: it requires a high-precision vision system capable of accurately predicting ball trajectories, and it necessitates intelligent strategic planning to ensure precise ball placement to target regions. The dynamic nature of table tennis, coupled with its real-time response requirements, makes it particularly well-suited for advancing robotic control capabilities in fast-paced, precision-critical domains. In this paper, we present SpikePingpong, a novel system that integrates spike-based vision with imitation learning for high-precision robotic table tennis. Our approach introduces two key attempts that directly address the aforementioned challenges: SONIC, a spike camera-based module that achieves millimeter-level precision in ball-racket contact prediction by compensating for real-world uncertainties such as air resistance and friction; and IMPACT, a strategic planning module that enables accurate ball placement to targeted table regions. The system harnesses a 20 kHz spike camera for high-temporal resolution ball tracking, combined with efficient neural network models for real-time trajectory correction and stroke planning. Experimental results demonstrate that SpikePingpong achieves a remarkable 91% success rate for 30 cm accuracy target area and 71% in the more challenging 20 cm accuracy task, surpassing previous state-of-the-art approaches by 38% and 37% respectively. These significant performance improvements enable the robust implementation of sophisticated tactical gameplay strategies, providing a new research perspective for robotic control in high-speed dynamic tasks.
3D Gaussian Splatting (3DGS) has recently gained significant attention for high-quality and efficient view synthesis, making it widely adopted in fields such as AR/VR, robotics, and autonomous driving. Despite its impressive algorithmic performance, real-time rendering on resource-constrained devices remains a major challenge due to tight power and area budgets. This paper presents an architecture-algorithm co-design to address these inefficiencies. First, we reveal substantial redundancy caused by repeated computation of common terms/expressions during the conventional rasterization. To resolve this, we propose axis-oriented rasterization, which pre-computes and reuses shared terms along both the X and Y axes through a dedicated hardware design, effectively reducing multiply-and-add (MAC) operations by up to 63%. Second, by identifying the resource and performance inefficiency of the sorting process, we introduce a novel neural sorting approach that predicts order-independent blending weights using an efficient neural network, eliminating the need for costly hardware sorters. A dedicated training framework is also proposed to improve its algorithmic stability. Third, to uniformly support rasterization and neural network inference, we design an efficient reconfigurable processing array that maximizes hardware utilization and throughput. Furthermore, we introduce a $\pi$-trajectory tile schedule, inspired by Morton encoding and Hilbert curve, to optimize Gaussian reuse and reduce memory access overhead. Comprehensive experiments demonstrate that the proposed design preserves rendering quality while achieving a speedup of $23.4\sim27.8\times$ and energy savings of $28.8\sim51.4\times$ compared to edge GPUs for real-world scenes. We plan to open-source our design to foster further development in this field.
Dynamic Spectral Backpropagation (DSBP) enhances neural network training under resource constraints by projecting gradients onto principal eigenvectors, reducing complexity and promoting flat minima. Five extensions are proposed, dynamic spectral inference, spectral architecture optimization, spectral meta learning, spectral transfer regularization, and Lie algebra inspired dynamics, to address challenges in robustness, fewshot learning, and hardware efficiency. Supported by a third order stochastic differential equation (SDE) and a PAC Bayes limit, DSBP outperforms Sharpness Aware Minimization (SAM), Low Rank Adaptation (LoRA), and Model Agnostic Meta Learning (MAML) on CIFAR 10, Fashion MNIST, MedMNIST, and Tiny ImageNet, as demonstrated through extensive experiments and visualizations. Future work focuses on scalability, bias mitigation, and ethical considerations.
The continuous improvement in weather forecast skill over the past several decades is largely due to the increasing quantity of available satellite observations and their assimilation into operational forecast systems. Assimilating these observations requires observation operators in the form of radiative transfer models. Significant efforts have been dedicated to enhancing the computational efficiency of these models. Computational cost remains a bottleneck, and a large fraction of available data goes unused for assimilation. To address this, we used machine learning to build an efficient neural network based probabilistic emulator of the Community Radiative Transfer Model (CRTM), applied to the GOES Advanced Baseline Imager. The trained NN emulator predicts brightness temperatures output by CRTM and the corresponding error with respect to CRTM. RMSE of the predicted brightness temperature is 0.3 K averaged across all channels. For clear sky conditions, the RMSE is less than 0.1 K for 9 out of 10 infrared channels. The error predictions are generally reliable across a wide range of conditions. Explainable AI methods demonstrate that the trained emulator reproduces the relevant physics, increasing confidence that the model will perform well when presented with new data.
Spiking Neural Networks (SNNs) are highly regarded for their energy efficiency, inherent activation sparsity, and suitability for real-time processing in edge devices. However, most current SNN methods adopt architectures resembling traditional artificial neural networks (ANNs), leading to suboptimal performance when applied to SNNs. While SNNs excel in energy efficiency, they have been associated with lower accuracy levels than traditional ANNs when utilizing conventional architectures. In response, in this work we present LightSNN, a rapid and efficient Neural Network Architecture Search (NAS) technique specifically tailored for SNNs that autonomously leverages the most suitable architecture, striking a good balance between accuracy and efficiency by enforcing sparsity. Based on the spiking NAS network (SNASNet) framework, a cell-based search space including backward connections is utilized to build our training-free pruning-based NAS mechanism. Our technique assesses diverse spike activation patterns across different data samples using a sparsity-aware Hamming distance fitness evaluation. Thorough experiments are conducted on both static (CIFAR10 and CIFAR100) and neuromorphic datasets (DVS128-Gesture). Our LightSNN model achieves state-of-the-art results on CIFAR10 and CIFAR100, improves performance on DVS128Gesture by 4.49\%, and significantly reduces search time most notably offering a $98\times$ speedup over SNASNet and running 30\% faster than the best existing method on DVS128Gesture. Code is available on Github at: https://github.com/YesmineAbdennadher/LightSNN.
Feedforward neural networks are widely used in autonomous systems, particularly for control and perception tasks within the system loop. However, their vulnerability to adversarial attacks necessitates formal verification before deployment in safety-critical applications. Existing set propagation-based reachability analysis methods for feedforward neural networks often struggle to achieve both scalability and accuracy. This work presents a novel set-based approach for computing the reachable sets of convolutional neural networks. The proposed method leverages a hybrid zonotope representation and an efficient neural network reduction technique, providing a flexible trade-off between computational complexity and approximation accuracy. Numerical examples are presented to demonstrate the effectiveness of the proposed approach.
In some game scenarios, due to the uncertainty of the number of enemy units and the priority of various attributes, the evaluation of the threat level of enemy units as well as the screening has been a challenging research topic, and the core difficulty lies in how to reasonably set the priority of different attributes in order to achieve quantitative evaluation of the threat. In this paper, we innovatively transform the problem of threat assessment into a reinforcement learning problem, and through systematic reinforcement learning training, we successfully construct an efficient neural network evaluator. The evaluator can not only comprehensively integrate the multidimensional attribute features of the enemy, but also effectively combine our state information, thus realizing a more accurate and scientific threat assessment.
Structured pruning and quantization are fundamental techniques used to reduce the size of deep neural networks (DNNs) and typically are applied independently. Applying these techniques jointly via co-optimization has the potential to produce smaller, high-quality models. However, existing joint schemes are not widely used because of (1) engineering difficulties (complicated multi-stage processes), (2) black-box optimization (extensive hyperparameter tuning to control the overall compression), and (3) insufficient architecture generalization. To address these limitations, we present the framework GETA, which automatically and efficiently performs joint structured pruning and quantization-aware training on any DNNs. GETA introduces three key innovations: (i) a quantization-aware dependency graph (QADG) that constructs a pruning search space for generic quantization-aware DNN, (ii) a partially projected stochastic gradient method that guarantees layerwise bit constraints are satisfied, and (iii) a new joint learning strategy that incorporates interpretable relationships between pruning and quantization. We present numerical experiments on both convolutional neural networks and transformer architectures that show that our approach achieves competitive (often superior) performance compared to existing joint pruning and quantization methods.
In this paper, we introduce **harmonic loss** as an alternative to the standard cross-entropy loss for training neural networks and large language models (LLMs). Harmonic loss enables improved interpretability and faster convergence, owing to its scale invariance and finite convergence point by design, which can be interpreted as a class center. We first validate the performance of harmonic models across algorithmic, vision, and language datasets. Through extensive experiments, we demonstrate that models trained with harmonic loss outperform standard models by: (a) enhancing interpretability, (b) requiring less data for generalization, and (c) reducing grokking. Moreover, we compare a GPT-2 model trained with harmonic loss to the standard GPT-2, illustrating that the harmonic model develops more interpretable representations. Looking forward, we believe harmonic loss has the potential to become a valuable tool in domains with limited data availability or in high-stakes applications where interpretability and reliability are paramount, paving the way for more robust and efficient neural network models.
Purpose: Surgical performance depends not only on surgeons' technical skills but also on team communication within and across the different professional groups present during the operation. Therefore, automatically identifying team communication in the OR is crucial for patient safety and advances in the development of computer-assisted surgical workflow analysis and intra-operative support systems. To take the first step, we propose a new task of detecting communication briefings involving all OR team members, i.e. the team Time-out and the StOP?-protocol, by localizing their start and end times in video recordings of surgical operations. Methods: We generate an OR dataset of real surgeries, called Team-OR, with more than one hundred hours of surgical videos captured by the multi-view camera system in the OR. The dataset contains temporal annotations of 33 Time-out and 22 StOP?-protocol activities in total. We then propose a novel group activity detection approach, where we encode both scene context and action features, and use an efficient neural network model to output the results. Results: The experimental results on the Team-OR dataset show that our approach outperforms existing state-of-the-art temporal action detection approaches. It also demonstrates the lack of research on group activities in the OR, proving the significance of our dataset. Conclusion: We investigate the Team Time-Out and the StOP?-protocol in the OR, by presenting the first OR dataset with temporal annotations of group activities protocols, and introducing a novel group activity detection approach that outperforms existing approaches. Code is available at https://github.com/CAMMA-public/Team-OR.
Convolutional Neural Networks (CNNs) are the backbone of many deep learning methods, but optimizing them remains computationally expensive. To address this, we explore multiscale training frameworks and mathematically identify key challenges, particularly when dealing with noisy inputs. Our analysis reveals that in the presence of noise, the gradient of standard CNNs in multiscale training may fail to converge as the mesh-size approaches to , undermining the optimization process. This insight drives the development of Mesh-Free Convolutions (MFCs), which are independent of input scale and avoid the pitfalls of traditional convolution kernels. We demonstrate that MFCs, with their robust gradient behavior, ensure convergence even with noisy inputs, enabling more efficient neural network optimization in multiscale settings. To validate the generality and effectiveness of our multiscale training approach, we show that (i) MFCs can theoretically deliver substantial computational speedups without sacrificing performance in practice, and (ii) standard convolutions benefit from our multiscale training framework in practice.
Backpropagation is the standard method for achieving state-of-the-art accuracy in neural network training, but it often imposes high memory costs and lacks biological plausibility. In this paper, we introduce the Mono-Forward algorithm, a purely local layerwise learning method inspired by Hinton's Forward-Forward framework. Unlike backpropagation, Mono-Forward optimizes each layer solely with locally available information, eliminating the reliance on global error signals. We evaluated Mono-Forward on multi-layer perceptrons and convolutional neural networks across multiple benchmarks, including MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100. The test results show that Mono-Forward consistently matches or surpasses the accuracy of backpropagation across all tasks, with significantly reduced and more even memory usage, better parallelizability, and a comparable convergence rate.
Quantized training of Large Language Models (LLMs) remains an open challenge, as maintaining accuracy while performing all matrix multiplications in low precision has proven difficult. This is particularly the case when fine-tuning pre-trained models, which can have large weight and activation outlier values that make lower-precision optimization difficult. To address this, we present HALO, a novel quantization-aware training approach for Transformers that enables accurate and efficient low-precision training by combining 1) strategic placement of Hadamard rotations in both forward and backward passes, which mitigate outliers, 2) high-performance kernel support, and 3) FSDP integration for low-precision communication. Our approach ensures that all large matrix multiplications during the forward and backward passes are executed in lower precision. Applied to LLAMA-family models, HALO achieves near-full-precision-equivalent results during fine-tuning on various tasks, while delivering up to 1.41x end-to-end speedup for full fine-tuning on RTX 4090 GPUs. HALO efficiently supports both standard and parameterefficient fine-tuning (PEFT). Our results demonstrate the first practical approach to fully quantized LLM fine-tuning that maintains accuracy in 8-bit precision, while delivering performance benefits. Code is available at \url{https://github.com/IST-DASLab/HALO}.
Accelerating model convergence in resource-constrained environments is essential for fast and efficient neural network training. This work presents learn2mix, a new training strategy that adaptively adjusts class proportions within batches, focusing on classes with higher error rates. Unlike classical training methods that use static class proportions, learn2mix continually adapts class proportions during training, leading to faster convergence. Empirical evaluations on benchmark datasets show that neural networks trained with learn2mix converge faster than those trained with classical approaches, achieving improved results for classification, regression, and reconstruction tasks under limited training resources and with imbalanced classes. Our empirical findings are supported by theoretical analysis.
3D color lookup tables (LUTs) enable precise color manipulation by mapping input RGB values to specific output RGB values. 3D LUTs are instrumental in various applications, including video editing, in-camera processing, photographic filters, computer graphics, and color processing for displays. While an individual LUT does not incur a high memory overhead, software and devices may need to store dozens to hundreds of LUTs that can take over 100 MB. This work aims to develop a neural network architecture that can encode hundreds of LUTs in a single compact representation. To this end, we propose a model with a memory footprint of less than 0.25 MB that can reconstruct 512 LUTs with only minor color distortion ($\bar{\Delta}E_M$ $\leq$ 2.0) over the entire color gamut. We also show that our network can weight colors to provide further quality gains on natural image colors ($\bar{\Delta}{E}_M$ $\leq$ 1.0). Finally, we show that minor modifications to the network architecture enable a bijective encoding that produces LUTs that are invertible, allowing for reverse color processing. Our code is available at https://github.com/vahidzee/ennelut.
For many complex simulation tasks spanning areas such as healthcare, engineering, and finance, Monte Carlo (MC) methods are invaluable due to their unbiased estimates and precise error quantification. Nevertheless, Monte Carlo simulations often become computationally prohibitive, especially for nested, multi-level, or path-dependent evaluations lacking effective variance reduction techniques. While machine learning (ML) surrogates appear as natural alternatives, naive replacements typically introduce unquantifiable biases. We address this challenge by introducing Prediction-Enhanced Monte Carlo (PEMC), a framework that leverages modern ML models as learned predictors, using cheap and parallelizable simulation as features, to output unbiased evaluation with reduced variance and runtime. PEMC can also be viewed as a "modernized" view of control variates, where we consider the overall computation-cost-aware variance reduction instead of per-replication reduction, while bypassing the closed-form mean function requirement and maintaining the advantageous unbiasedness and uncertainty quantifiability of Monte Carlo. We illustrate PEMC's broader efficacy and versatility through three examples: first, equity derivatives such as variance swaps under stochastic local volatility models; second, interest rate derivatives such as swaption pricing under the Heath-Jarrow-Morton (HJM) interest-rate model. Finally, we showcase PEMC in a socially significant context - ambulance dispatch and hospital load balancing - where accurate mortality rate estimates are key for ethically sensitive decision-making. Across these diverse scenarios, PEMC consistently reduces variance while preserving unbiasedness, highlighting its potential as a powerful enhancement to standard Monte Carlo baselines.
Improving diesel engine efficiency and emission reduction have been critical research topics. Recent government regulations have shifted this focus to another important area related to engine health and performance monitoring. Although the advancements in the use of deep learning methods for system monitoring have shown promising results in this direction, designing efficient methods suitable for field systems remains an open research challenge. The objective of this study is to develop a computationally efficient neural network-based approach for identifying unknown parameters of a mean value diesel engine model to facilitate physics-based health monitoring and maintenance forecasting. We propose a hybrid method combining physics informed neural networks, PINNs, and a deep neural operator, DeepONet to predict unknown parameters and gas flow dynamics in a diesel engine. The operator network predicts independent actuator dynamics learnt through offline training, thereby reducing the PINNs online computational cost. To address PINNs need for retraining with changing input scenarios, we propose two transfer learning (TL) strategies. The first strategy involves multi-stage transfer learning for parameter identification. While this method is computationally efficient as compared to online PINN training, improvements are required to meet field requirements. The second TL strategy focuses solely on training the output weights and biases of a subset of multi-head networks pretrained on a larger dataset, substantially reducing computation time during online prediction. We also evaluate our model for epistemic and aleatoric uncertainty by incorporating dropout in pretrained networks and Gaussian noise in the training dataset. This strategy offers a tailored, computationally inexpensive, and physics-based approach for parameter identification in diesel engine sub systems.