Event-based vision has emerged as a promising paradigm for energy-aware artificial intelligence (AI), offering sparse, low-latency visual signals that reduce redundant data processing and support sustainable edge computing. However, the asynchronous and noise-prone nature of event streams creates challenges for conventional deep learning models, which are often too computationally intensive for low-power embedded platforms. This work presents a compact and configurable event-driven autoencoder that efficiently compresses neuromorphic data while preserving essential spatiotemporal structure for downstream inference. The architecture integrates lightweight convolutional encoding with robust performance under adaptive event thresholding and a minimal classifier head, enabling substantial reductions in computational cost without degrading recognition fidelity. Extensive evaluations on the Smart Event Face Dataset (SEFD) and Event-Based Crossing Dataset (EBCD) show that the proposed framework achieves competitive or superior accuracy compared to YOLOv9 while requiring up to 35.6× fewer parameters. To assess real-world sustainability, the model is deployed on resource-constrained hardware: a Raspberry Pi 4B and a NVIDIA Jetson Nano. On NVIDIA Jetson Nano, it delivers real-time throughput of 44.8 FPS. On a Raspberry Pi 4B CPU, the 50% autoencoder classifier consumes 16.19 J for the evaluated inference workload, corresponding to approximately 726.3× lower energy consumption than YOLOv9 under the same evaluation protocol. These results demonstrate the potential of compact event-driven models to advance environmentally conscious, low-power AI systems for high-speed perception in autonomous, mobile, and embedded computing environments.
Figures & tables
Fig. 1: Comparison of average model accuracy, latency, and energy consumption on the SEFD [ 28 ] dataset, evaluated on a Raspberry Pi 4B CPU.
Fig. 2: (a) The proposed event autoencoder uses a conventional convolution layer followed by ReLU and Max pooling layer on the encoder side, while the decoder uses a similar convolution layer followed by ReLU but an up-sampling layer; and (b) the autoencoder-based classifier concatenates the encoder with two fully connected layers.
Fig. 3: Comprehensive edge-computing testbed integrating Raspberry Pi 4B and NVIDIA Jetson Nano boards with a Raspberry Pi camera and USB-based power analysis tools, enabling systematic evaluation of computational load, energy consumption, and real-time performance.
Model
Params
Th
Accuracy
Precision
Recall
F1-Score
FLOPs
YOLOv4 [ 6 ]
60.3M
4
83.81
94.00
89.00
91.00
59.56B
8
80.33
89.00
89.00
89.00
12
79.98
91.00
86.00
89.00
16
74.49
93.00
79.00
85.00
YOLOv7 [ 60 ]
25.2M
4
86.71
94.00
92.00
93.00
43.61B
8
82.95
92.00
90.00
91.00
TABLE I: Performance comparison on the SEFD [ 28 ] test dataset under temporal thresholds Th∈{4,8,12,16} . Our autoencoder classifier achieves competitive performance on frame-level data with orders of magnitude fewer parameters than heavyweight baselines.
Model
Parameters
Th
Accuracy
Precision
Recall
F1-Score
FLOPs
YOLOv4 [ 6 ]
60.3M
12
88.00
96.00
80.00
87.00
59.56B
16
87.50
94.00
81.00
87.00
20
89.00
94.00
84.00
87.00
YOLOv7 [ 60 ]
25.2M
12
77.99
97.00
78.00
87.00
43.61B
16
77.99
97.00
78.00
86.00
20
79.35
97.00
79.00
87.00
TABLE II: Performance comparison on the EBCD [ 44 ] test dataset
Fig. 4: (a) Training curve of the full-capacity event autoencoder, yielding an average loss close to fifteen hundredths. (b) Performance of the reduced 50%–filter configuration, which achieves a lower average training loss close to one-tenth, illustrating the model’s resilience under substantial architectural compression.
Fig. 5: The reconstruction outputs of the proposed models exhibit excellent accuracy, comparable to their model sizes.
Fig. 6: AUROC performance of the proposed autoencoder-based classifier across SEFD threshold settings. All configurations achieve AUC values above 90%, with peak performance at Th=8
Fig. 7: Runtime evaluation of the proposed architectures compared with contemporary work on Raspberry Pi 4B and NVIDIA Jetson Nano (4GB). The results show reduced execution time and higher FPS, with the autoencoder classifier achieving 22.1 × the FPS of YOLOv9. Although YuNet is fastest, the proposed 50% filter classifier is 6.3% more accurate.
Fig. 8: Energy consumption and runtime during inference on a Raspberry Pi 4B. The proposed 50% Autoencoder Classifier and 50% Autoencoder require only 16.19 J and 30.48 J to process 100 images, achieving more than a 2× energy reduction over their full-size counterparts (36.37 J and 64.46 J). Their runtimes are similarly low at 4.8 s and 8.6 s. In contrast, lightweight baselines such as MobileNet-v1 and YOLOv7 consume 3234.95 J and 8440.31 J (476.2 s and 1250.0 s), while the most demanding model, YOLOv9, reaches 11,759.39 J and 2000.0 s—over 700× the energy usage of the 50% Autoencoder Classifier.
Model
Parameters
Th
Accuracy
Precision
Recall
F1-Score
FLOPs
Autoencoder Classifier 25% (Ours)
132K
4
88.79
86.48
82.97
84.69
429M
8
86.30
86.33
86.25
86.29
12
86.46
85.11
88.39
86.71
16
84.56
82.32
88.02
85.07
TABLE III: Ablation on the analysis of a further reduced autoencoder classifier, evaluated on the SEFD [ 28 ] dataset across temporal thresholds Th∈{4,8,12,16} .
Time-series anomaly detection is increasingly important in IoT systems, sensor networks, and edge monitoring applications, where models must operate under strict constraints on memory, latency, and power consumption. While recent deep-learning approaches have improved detection accuracy, many remain computationally expensive and often fail to capture subtle anomalies due to limited multi-scale sensitivity. Autoencoders are widely used for anomaly detection because they reconstruct normal patterns well, leading to elevated reconstruction errors for anomalous inputs. Their simplicity and efficiency also make them suitable lightweight backbones for handling multi-scale inputs. To address these challenges, we propose a Lightweight MultiScale AutoEncoder (LMSAE) network for univariate time-series anomaly detection, designed to be compact and computationally efficient. LMSAE leverages the Discrete Wavelet Transform (DWT) to extract multi-scale features and employs a multi-scale loss function to improve sensitivity to subtle or hidden anomalies. Experiments on benchmark datasets demonstrate competitive or superior detection performance despite using significantly fewer parameters and a model size of less than 500 KB. LMSAE also achieves low-latency, low-power inference on the NVIDIA Jetson Nano, with 9x reduction in inference latency and 2x reduction in power consumption, making it ideal for edge deployment.
Raheen Junaid Wani, Smruti R. Sarangi
School of Information Technology Indian Institute of Technology, Delhi New Delhi, India 110016 · Department of Computer Science Indian Institute of Technology, Delhi New Delhi, India 110016
Audio-visual event recognition (AVER) has achieved significant performance improvements through transformer-based multimodal architectures. However, the high computational complexity, large memory footprint, and inference cost of these models hinder their deployment on edge and resource-constrained devices. This paper presents an efficient compression framework for hybrid cross-attention-based audiovisual event recognition by combining architectural model compression, knowledge distillation, and dynamic INT8 quantization. A high-capacity teacher model integrates VideoMAE for visual representation learning, the Audio Spectrogram Transformer (AST) for audio feature extraction, and a hybrid cross-attention fusion network for multimodal feature integration. A lightweight student model is constructed by reducing the hidden feature dimension, the number of attention heads, and the feedforward network size while preserving the overall network architecture. The student model is trained using knowledge distillation to effectively transfer discriminative knowledge from the teacher. Finally, dynamic INT8 post-training quantization is applied to further reduce the model size for efficient deployment. Experimental results on the Audio-Visual Event (AVE) dataset show that the proposed framework reduces the number of trainable parameters in the multimodal fusion module by 59.06%, with only a 2.14% decrease in classification accuracy compared with the teacher model. Furthermore, dynamic INT8 quantization reduces the model size from 10.71 MB to 2.04 MB while maintaining competitive recognition performance. These results demonstrate that the proposed framework provides an effective trade-off between recognition accuracy and computational efficiency, making it a promising solution for deployment on resource-constrained edge AI platforms.
Parinaz Binandeh Dehaghani, Danilo Pena, A. Pedro Aguiar
University of Porto, Porto, Portugal · ResoSight, Montreal, Canada
Face recognition has become a cornerstone of modern AI applications, yet conventional approaches often rely on computationally intensive models deployed in cloud environments, leading to increased network traffic, high energy consumption, and a heavy carbon footprint. This work introduces a sustainable, edge-deployable face recognition framework based on Vector-Quantized Variational Autoencoders (VQ-VAE), which generates compact and semantically rich latent representations of facial images. By leveraging the compression capacity and reconstruction quality of VQ-VAE embeddings on the edge and combining them with the power of pre-trained face embeddings in a knowledge distillation setup, our system achieves comparable accuracy to state-of-the-art face embedding models while significantly reducing memory and computation requirements on the edge, making it suitable for low-power edge devices. The integration of VQ-VAE compression minimizes network overhead while keeping the matching accuracy high by retaining only the most informative facial features in the latent space. As a result, the reconstructed images preserve the key identity characteristics, improving the robustness and overall performance of the face embeddings.