Yuxiao Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Yuxiao Li has a total of 18 co-authors. Here are the top co-authors by number of shared publications:
- Yuan Shen
Tsinghua University
2 shared publications - Boyuan Dong
Department of Electronic Engineering, Tsinghua University, Beijing, China
1 shared publication - Can Wang
Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
1 shared publication - David D. Baek
MIT MIT MIT MIT
1 shared publication - Eric J. Michaud
Massachusetts Institute of Technology
1 shared publication - Hanfei Ma
Shanghai Academy of Spaceflight Technology
1 shared publication
Latest Publications
- Top Cited PapersOpen Access
The Geometry of Concepts: Sparse Autoencoder Feature Structure
Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: (1) The “atomic” small-scale structure contains “crystals” whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man:woman::king:queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently performed with linear discriminant analysis. (2) The “brain” intermediate-scale structure has significant spatial modularity; for example, math and code features form a “lobe” akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. (3) The “galaxy”-scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.
Top Cited PapersOpen AccessThe Geometry of Concepts: Sparse Autoencoder Feature Structure
Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: (1) The “atomic” small-scale structure contains “crystals” whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man:woman::king:queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently performed with linear discriminant analysis. (2) The “brain” intermediate-scale structure has significant spatial modularity; for example, math and code features form a “lobe” akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. (3) The “galaxy”-scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.
- 2 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Wireless Communications
- Vol. 23 (8) , 9113-9128
Multipath-Assisted Single-Anchor Localization via Deep Variational Learning
Location awareness plays an increasingly important role in wireless network applications. However, accurate localization in complex indoor environments remains challenging for existing radio frequency (RF)-based systems, among which the ultra-wide bandwidth (UWB) technology ranks to be the most promising one due to its capability in providing channel information with fine time resolution. In this paper, we propose a multipath-assisted single-anchor localization framework that can provide high-accuracy positional information in complex indoor environments. Specifically, a deep variational learning method is proposed to produce calibrated estimates of position-related parameters, including distance, time-difference-of-arrival and angle-of-arrival, which are then fed into a multipath-assisted single-anchor localization algorithm. The proposed method is implemented on self-built UWB transceivers and assessed with real-world data from an indoor measurement campaign. Extensive experimental results show that the proposed method outperforms conventional machine learning-based error mitigation approaches and can achieve 0.15m root mean square position error in non-line-of-sight scenarios.
- 2 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Wireless Communications
- Vol. 23 (8) , 9113-9128
Multipath-Assisted Single-Anchor Localization via Deep Variational Learning
Location awareness plays an increasingly important role in wireless network applications. However, accurate localization in complex indoor environments remains challenging for existing radio frequency (RF)-based systems, among which the ultra-wide bandwidth (UWB) technology ranks to be the most promising one due to its capability in providing channel information with fine time resolution. In this paper, we propose a multipath-assisted single-anchor localization framework that can provide high-accuracy positional information in complex indoor environments. Specifically, a deep variational learning method is proposed to produce calibrated estimates of position-related parameters, including distance, time-difference-of-arrival and angle-of-arrival, which are then fed into a multipath-assisted single-anchor localization algorithm. The proposed method is implemented on self-built UWB transceivers and assessed with real-world data from an indoor measurement campaign. Extensive experimental results show that the proposed method outperforms conventional machine learning-based error mitigation approaches and can achieve 0.15m root mean square position error in non-line-of-sight scenarios.
- 6 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Wireless Communications
- Vol. 22 (9) , 6252-6266
A Variational Learning Approach for Concurrent Distance Estimation and Environmental Identification
Wireless propagated signals encapsulate rich information for high-accuracy localization and environment sensing. However, the full exploitation of positional and environmental features as well as their correlation remains challenging in complex propagation environments. In this paper, we propose a methodology of variational inference over deep neural networks for concurrent distance estimation and environmental identification. The proposed approach, namely inter-instance variational auto-encoders (IIns-VAEs), conducts inference with latent variables that encapsulate information about both distance and environmental labels. A deep learning network with instance normalization is designed to approximate the inference concurrently via deep learning. We conduct extensive experiments on real-world datasets and the results show the superiority of the proposed IIns-VAE in both distance estimation and environmental identification compared to conventional approaches.
- 6 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Wireless Communications
- Vol. 22 (9) , 6252-6266
A Variational Learning Approach for Concurrent Distance Estimation and Environmental Identification
Wireless propagated signals encapsulate rich information for high-accuracy localization and environment sensing. However, the full exploitation of positional and environmental features as well as their correlation remains challenging in complex propagation environments. In this paper, we propose a methodology of variational inference over deep neural networks for concurrent distance estimation and environmental identification. The proposed approach, namely inter-instance variational auto-encoders (IIns-VAEs), conducts inference with latent variables that encapsulate information about both distance and environmental labels. A deep learning network with instance normalization is designed to approximate the inference concurrently via deep learning. We conduct extensive experiments on real-world datasets and the results show the superiority of the proposed IIns-VAE in both distance estimation and environmental identification compared to conventional approaches.
- 1 August 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 36, 414-419
A Trajectory Optimization Algorithm for Drone Target Localization and Tracking
Accurate positioning is the premise for the drone to successfully capture the target, this paper proposes a solution to the positioning failure caused by the relatively severe jitter of the drone and the target. In this paper, the Improved Kernel Correlation Filter (KCF) is used for target tracking, and the improved Kalman filter is used to smooth the input signal in real time, which reduces the continuous jitter of the drone. The advantage of this approach is that it will not cause misjudgment from the deviation where the measured value is too large. Further, the hysteresis is also significantly improved. The experimental works are carried out on a M100 UAV to verify the improved KCF and the improved Kalman algorithm, respectively. The results showed the effectiveness of the proposed approach.
- 1 August 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 36, 414-419
A Trajectory Optimization Algorithm for Drone Target Localization and Tracking
Accurate positioning is the premise for the drone to successfully capture the target, this paper proposes a solution to the positioning failure caused by the relatively severe jitter of the drone and the target. In this paper, the Improved Kernel Correlation Filter (KCF) is used for target tracking, and the improved Kalman filter is used to smooth the input signal in real time, which reduces the continuous jitter of the drone. The advantage of this approach is that it will not cause misjudgment from the deviation where the measured value is too large. Further, the hysteresis is also significantly improved. The experimental works are carried out on a M100 UAV to verify the improved KCF and the improved Kalman algorithm, respectively. The results showed the effectiveness of the proposed approach.
- 1 January 2012
- Published by Springer Nature
- p. 227-234
Research and Modeling on Reputation Information Broadcast on Social Media
- 1 January 2012
- Published by Springer Nature
- p. 227-234
Research and Modeling on Reputation Information Broadcast on Social Media