arXiv subjects

Milin Zhang

Publications and source records attributed to Milin Zhang.

15 recordsLinked to original sources

arXiv2026

BoxMind: Closed-loop AI strategy optimization for elite boxing validated in the 2024 Olympics

Competitive sports require sophisticated tactical analysis, yet combat disciplines like boxing remain underdeveloped in AI-driven analytics due to the complexity of action dynamics and the lack of structured tactical representations. To address this, we present BoxMind, a closed-loop AI expert system validated in elite boxing competition. By defining atomic punch events with precise temporal boundaries and spatial and technical attributes, we parse match footage into 18 hierarchical technical-tactical indicators. We then propose a graph-based predictive model that fuses these explicit technical-tactical profiles with learnable, time-variant latent embeddings to capture the dynamics of boxer matchups. Modeling match outcome as a differentiable function of technical-tactical indicators, we turn winning probability gradients into executable tactical adjustments. Experiments show that the outcome prediction model achieves state-of-the-art performance, with 69.8% accuracy on BoxerGraph test set and 87.5% on Olympic matches. Using this predictive model as a foundation, the system generates strategic recommendations that demonstrate proficiency comparable to human experts. BoxMind is validated through a closed-loop deployment during the 2024 Paris Olympics, directly contributing to the Chinese National Team's historic achievement of three gold and two silver medals. BoxMind establishes a replicable paradigm for transforming unstructured video data into strategic intelligence, bridging the gap between computer vision and decision support in competitive sports. Code and data is available at https://github.com/gouba2333/BoxingWeb.

cs.AI↗

arXiv2024

Semantic Edge Computing and Semantic Communications in 6G Networks: A Unifying Survey and Research Challenges

Semantic Edge Computing (SEC) and Semantic Communications (SemComs) have been proposed as viable approaches to achieve real-time edge-enabled intelligence in sixth-generation (6G) wireless networks. On one hand, SemCom leverages the strength of Deep Neural Networks (DNNs) to encode and communicate the semantic information only, while making it robust to channel distortions by compensating for wireless effects. Ultimately, this leads to an improvement in the communication efficiency. On the other hand, SEC has leveraged distributed DNNs to divide the computation of a DNN across different devices based on their computational and networking constraints. Although significant progress has been made in both fields, the literature lacks a systematic view to connect both fields. In this work, we fulfill the current gap by unifying the SEC and SemCom fields. We summarize the research problems in these two fields and provide a comprehensive review of the state of the art with a focus on their technical strengths and challenges.

cs.LG↗

arXiv2024

Resilience and Security of Deep Neural Networks Against Intentional and Unintentional Perturbations: Survey and Research Challenges

In order to deploy deep neural networks (DNNs) in high-stakes scenarios, it is imperative that DNNs provide inference robust to external perturbations - both intentional and unintentional. Although the resilience of DNNs to intentional and unintentional perturbations has been widely investigated, a unified vision of these inherently intertwined problem domains is still missing. In this work, we fill this gap by providing a survey of the state of the art and highlighting the similarities of the proposed approaches.We also analyze the research challenges that need to be addressed to deploy resilient and secure DNNs. As there has not been any such survey connecting the resilience of DNNs to intentional and unintentional perturbations, we believe this work can help advance the frontier in both domains by enabling the exchange of ideas between the two communities.

cs.CR↗

arXiv2024

Resilience of Entropy Model in Distributed Neural Networks

Distributed deep neural networks (DNNs) have emerged as a key technique to reduce communication overhead without sacrificing performance in edge computing systems. Recently, entropy coding has been introduced to further reduce the communication overhead. The key idea is to train the distributed DNN jointly with an entropy model, which is used as side information during inference time to adaptively encode latent representations into bit streams with variable length. To the best of our knowledge, the resilience of entropy models is yet to be investigated. As such, in this paper we formulate and investigate the resilience of entropy models to intentional interference (e.g., adversarial attacks) and unintentional interference (e.g., weather changes and motion blur). Through an extensive experimental campaign with 3 different DNN architectures, 2 entropy models and 4 rate-distortion trade-off factors, we demonstrate that the entropy attacks can increase the communication overhead by up to 95%. By separating compression features in frequency and spatial domain, we propose a new defense mechanism that can reduce the transmission overhead of the attacked input by about 9% compared to unperturbed data, with only about 2% accuracy loss. Importantly, the proposed defense mechanism is a standalone approach which can be applied in conjunction with approaches such as adversarial training to further improve robustness. Code will be shared for reproducibility.

cs.LG↗

arXiv2024

Cepstral Analysis Based Artifact Detection, Recognition and Removal for Prefrontal EEG

This paper proposes to use cepstrum for artifact detection, recognition and removal in prefrontal EEG. This work focuses on the artifact caused by eye movement. A database containing artifact-free EEG and eye movement contaminated EEG from different subjects is established. A cepstral analysis-based feature extraction with support vector machine (SVM) based classifier is designed to identify the artifacts from the target EEG signals. The proposed method achieves an accuracy of 99.62% on the artifact detection task and a 82.79% accuracy on the 6-category eye movement classification task. A statistical value-based artifact removal method is proposed and evaluated on a public EEG database, where an accuracy improvement of 3.46% is obtained on the 3-category emotion classification task. In order to make a confident decision of each 5s EEG segment, the algorithm requires only 0.66M multiplication operations. Compared to the state-of-the-art approaches in artifact detection and removal, the proposed method features higher detection accuracy and lower computational cost, which makes it a more suitable solution to be integrated into a real-time and artifact robust Brain-Machine Interface (BMI).

eess.SP↗

arXiv2024

Stitching the Spectrum: Semantic Spectrum Segmentation with Wideband Signal Stitching

Spectrum has become an extremely scarce and congested resource. As a consequence, spectrum sensing enables the coexistence of different wireless technologies in shared spectrum bands. Most existing work requires spectrograms to classify signals. Ultimately, this implies that images need to be continuously created from I/Q samples, thus creating unacceptable latency for real-time operations. In addition, spectrogram-based approaches do not achieve sufficient granularity level as they are based on object detection performed on pixels and are based on rectangular bounding boxes. For this reason, we propose a completely novel approach based on semantic spectrum segmentation, where multiple signals are simultaneously classified and localized in both time and frequency at the I/Q level. Conversely from the state-of-the-art computer vision algorithm, we add non-local blocks to combine the spatial features of signals, and thus achieve better performance. In addition, we propose a novel data generation approach where a limited set of easy-to-collect real-world wireless signals are ``stitched together'' to generate large-scale, wideband, and diverse datasets. Experimental results obtained on multiple testbeds (including the Arena testbed) using multiple antennas, multiple sampling frequencies, and multiple radios over the course of 3 days show that our approach classifies and localizes signals with a mean intersection over union (IOU) of 96.70% across 5 wireless protocols while performing in real-time with a latency of 2.6 ms. Moreover, we demonstrate that our approach based on non-local blocks achieves 7% more accuracy when segmenting the most challenging signals with respect to the state-of-the-art U-Net algorithm. We will release our 17 GB dataset and code.

cs.NI↗

arXiv2024

LightSleepNet: Design of a Personalized Portable Sleep Staging System Based on Single-Channel EEG

This paper proposed LightSleepNet - a light-weight, 1-d Convolutional Neural Network (CNN) based personalized architecture for real-time sleep staging, which can be implemented on various mobile platforms with limited hardware resources. The proposed architecture only requires an input of 30s single-channel EEG signal for the classification. Two residual blocks consisting of group 1-d convolution are used instead of the traditional convolution layers to remove the redundancy in the CNN. Channel shuffles are inserted into each convolution layer to improve the accuracy. In order to avoid over-fitting to the training set, a Global Average Pooling (GAP) layer is used to replace the fully connected layer, which further reduces the total number of the model parameters significantly. A personalized algorithm combining Adaptive Batch Normalization (AdaBN) and gradient re-weighting is proposed for unsupervised domain adaptation. A higher priority is given to examples that are easy to transfer to the new subject, and the algorithm could be personalized for new subjects without re-training. Experimental results show a state-of-the-art overall accuracy of 83.8% with only 45.76 Million Floating-point Operations per Second (MFLOPs) computation and 43.08 K parameters.

eess.SP↗

arXiv2024

A Closed-loop Brain-Machine Interface SoC Featuring a 0.2$μ$J/class Multiplexer Based Neural Network

This work presents the first fabricated electrophysiology-optogenetic closed-loop bidirectional brain-machine interface (CL-BBMI) system-on-chip (SoC) with electrical neural signal recording, on-chip sleep staging and optogenetic stimulation. The first multiplexer with static assignment based table lookup solution (MUXnet) for multiplier-free NN processor was proposed. A state-of-the-art average accuracy of 82.4% was achieved with an energy consumption of only 0.2$μ$J/class in sleep staging task.

eess.SP↗

arXiv2024

Multi-Channel Multi-Domain based Knowledge Distillation Algorithm for Sleep Staging with Single-Channel EEG

This paper proposed a Multi-Channel Multi-Domain (MCMD) based knowledge distillation algorithm for sleep staging using single-channel EEG. Both knowledge from different domains and different channels are learnt in the proposed algorithm, simultaneously. A multi-channel pre-training and single-channel fine-tuning scheme is used in the proposed work. The knowledge from different channels in the source domain is transferred to the single-channel model in the target domain. A pre-trained teacher-student model scheme is used to distill knowledge from the multi-channel teacher model to the single-channel student model combining with output transfer and intermediate feature transfer in the target domain. The proposed algorithm achieves a state-of-the-art single-channel sleep staging accuracy of 86.5%, with only 0.6% deterioration from the state-of-the-art multi-channel model. There is an improvement of 2% compared to the baseline model. The experimental results show that knowledge from multiple domains (different datasets) and multiple channels (e.g. EMG, EOG) could be transferred to single-channel sleep staging.

eess.SP↗

arXiv2023

Adversarial Attacks to Latent Representations of Distributed Neural Networks in Split Computing

Distributed deep neural networks (DNNs) have been shown to reduce the computational burden of mobile devices and decrease the end-to-end inference latency in edge computing scenarios. While distributed DNNs have been studied, to the best of our knowledge, the resilience of distributed DNNs to adversarial action remains an open problem. In this paper, we fill the existing research gap by rigorously analyzing the robustness of distributed DNNs against adversarial action. We cast this problem in the context of information theory and rigorously proved that (i) the compressed latent dimension improves the robustness but also affect task-oriented performance; and (ii) the deeper splitting point enhances the robustness but also increases the computational burden. These two trade-offs provide a novel perspective to design robust distributed DNN. To test our theoretical findings, we perform extensive experimental analysis by considering 6 different DNN architectures, 6 different approaches for distributed DNN and 10 different adversarial attacks using the ImageNet-1K dataset.

cs.LG↗

arXiv2023

SiMWiSense: Simultaneous Multi-Subject Activity Classification Through Wi-Fi Signals

Recent advances in Wi-Fi sensing have ushered in a plethora of pervasive applications in home surveillance, remote healthcare, road safety, and home entertainment, among others. Most of the existing works are limited to the activity classification of a single human subject at a given time. Conversely, a more realistic scenario is to achieve simultaneous, multi-subject activity classification. The first key challenge in that context is that the number of classes grows exponentially with the number of subjects and activities. Moreover, it is known that Wi-Fi sensing systems struggle to adapt to new environments and subjects. To address both issues, we propose SiMWiSense, the first framework for simultaneous multi-subject activity classification based on Wi-Fi that generalizes to multiple environments and subjects. We address the scalability issue by using the Channel State Information (CSI) computed from the device positioned closest to the subject. We experimentally prove this intuition by confirming that the best accuracy is experienced when the CSI computed by the transceiver positioned closest to the subject is used for classification. To address the generalization issue, we develop a brand-new few-shot learning algorithm named Feature Reusable Embedding Learning (FREL). Through an extensive data collection campaign in 3 different environments and 3 subjects performing 20 different activities simultaneously, we demonstrate that SiMWiSense achieves classification accuracy of up to 97%, while FREL improves the accuracy by 85% in comparison to a traditional Convolutional Neural Network (CNN) and up to 20% when compared to the state-of-the-art few-shot embedding learning (FSEL), by using only 15 seconds of additional data for each class. For reproducibility purposes, we share our 1TB dataset and code repository.

cs.NI↗

arXiv2023

BeamSense: Rethinking Wireless Sensing with MU-MIMO Wi-Fi Beamforming Feedback

In this paper, we propose BeamSense, a completely novel approach to implement standard-compliant Wi-Fi sensing applications. Wi-Fi sensing enables game-changing applications in remote healthcare, home entertainment, and home surveillance, among others. However, existing work leverages the manual extraction of channel state information (CSI) from Wi-Fi chips to classify activities, which is not supported by the Wi-Fi standard and hence requires the usage of specialized equipment. On the contrary, BeamSense leverages the standard-compliant beamforming feedback information (BFI) to characterize the propagation environment. Conversely from CSI, the BFI (i) can be easily recorded without any firmware modification, and (ii) captures the multiple channels between the access point and the stations, thus providing much better sensitivity. BeamSense includes a novel cross-domain few-shot learning (FSL) algorithm to handle unseen environments and subjects with few additional data points. We evaluate BeamSense through an extensive data collection campaign with three subjects performing twenty different activities in three different environments. We show that our BFI-based approach achieves about 10% more accuracy when compared to CSI-based prior work, while our FSL strategy improves accuracy by up to 30% and 80% when compared with state-of-the-art cross-domain algorithms.

cs.NI↗

arXiv2022

Single-channel EEG completion using Cascade Transformer

It is easy for the electroencephalogram (EEG) signal to be incomplete due to packet loss, electrode falling off, etc. This paper proposed a Cascade Transformer architecture and a loss weighting method for the single-channel EEG completion, which reduced the Normalized Root Mean Square Error (NRMSE) by 2.8% and 8.5%, respectively. With the percentage of the missing points ranging from 1% to 50%, the proposed method achieved a NRMSE from 0.026 to 0.063, which aligned with the state-of-the-art multi-channel completion solution. The proposed work shows it's feasible to perform the EEG completion with only single-channel EEG.

eess.SP↗

arXiv2022

SaleNet: A low-power end-to-end CNN accelerator for sustained attention level evaluation using EEG

This paper proposes SaleNet - an end-to-end convolutional neural network (CNN) for sustained attention level evaluation using prefrontal electroencephalogram (EEG). A bias-driven pruning method is proposed together with group convolution, global average pooling (GAP), near-zero pruning, weight clustering and quantization for the model compression, achieving a total compression ratio of 183.11x. The compressed SaleNet obtains a state-of-the-art subject-independent sustained attention level classification accuracy of 84.2% on the recorded 6-subject EEG database in this work. The SaleNet is implemented on a Artix-7 FPGA with a competitive power consumption of 0.11 W and an energy-efficiency of 8.19 GOps/W.

cs.AR↗

arXiv2020

A Fully Integrated Sensor-Brain-Machine Interface System for Restoring Somatosensation

Sensory feedback is critical to the performance of neural prostheses that restore movement control after neurological injury. Recent advances in direct neural control of paralyzed arms present new requirements for miniaturized, low-power sensor systems. To address this challenge, we developed a fully-integrated wireless sensor-brain-machine interface (SBMI) system for communicating key somatosensory signals, fingertip forces and limb joint angles, to the brain. The system consists of a tactile force sensor, an electrogoniometer, and a neural interface. The tactile force sensor features a novel optical waveguide on CMOS design for sensing. The electrogoniometer integrates an ultra low-power digital signal processor (DSP) for real-time joint angle measurement. The neural interface enables bidirectional neural stimulation and recording. Innovative designs of sensors and sensing interfaces, analog-to-digital converters (ADC) and ultra wide-band (UWB) wireless transceivers have been developed. The prototypes have been fabricated in 180nm standard CMOS technology and tested on the bench and in vivo. The developed system provides a novel solution for providing somatosensory feedback to next-generation neural prostheses.

q-bio.NC↗