About Me

I am a recent graduate of Southeast University in China, having earned my bachelor's degree of Artificial Intelligence.

My research interest lies in machine learning modeling and natural language processing (NLP) with a focus on efficiency, safety, and explainability. More specifically, I am focusing on addressing fact-conflicting hallucinations in text generation to explore the faithfulness of explanations and cre- ating more efficient and accessible NLP.

Research Interest

  • design icon

    Enhancing Text Generation: Balancing Fluency and Fidelity

  • Web development icon

    Ameliorating Computational Efficiency and Safety through Interpretable Approaches

  • mobile app icon

    Addressing fact-conflicting hallucinations in LLMs

  • camera icon

    Trustworthy and Responsible ML & NLP Applications

Latest Publications

Daniel lewis
quote icon

Daniel lewis

Richard was hired to create a corporate identity. We were very pleased with the work done. She has a lot of experience and is very concerned about the needs of client. Lorem ipsum dolor sit amet, ullamcous cididt consectetur adipiscing elit, seds do et eiusmod tempor incididunt ut laborels dolore magnarels alia.

Resume

Education

  1. University of California, Los Angeles

    Mar.2022 - May.2022

    Exchange, Data Science, 97/100

  2. Southeast University

    Aug.2019 - Jun.2023

    B.S., Artificial Intelligence, 88.33/100

Research Experience

  1. Federated robust gradient difference compression

    Nov.2022 - Jun.2023

    Proposed FedGraD, which leverages gradient difference compression and combines robust aggregation rules in federated learning settings, avoids information loss, mitigates the communication bottleneck, and at the same time tolerates popular Byzantine attacks.
    Designed the pipeline of FedGraD, and analyzed four different attack setups on three datasets.

  2. Robust Spiking Neural Networks framework under noise and adversarial attacks

    April.2022 - Feb.2023

    Designed an inherently robust model with channel-wise activation rcalibration to overcome the performance flaws of Spiking Neural Networks by establishing inter-neuron connections and refining the diverse activation without additional training. Provide insights into the interpretability of attacks.
    Achieved SOTA performance on MNIST with 99.52% accuracy and successfully defeats 6 SOTA attacks on standard benchmarks, ranging from a single attack to multiple attacks.
    Explored the comparison of DNNs and SNNs, built the framework of CARE, and plotted the experimental results.

  3. Mutually interactive Emotion–Cause Pair Extractor via soft sharing

    May.2022 - Sep.2022

    Proposed an end-to-end soft-shared Emotion–Cause Pair Extractor framework and extends the framework to be compatible with language models. The performance is improved by 8.64% with essentially the same number of parameters; a level close to that of the SOTA method is achieved with 11.5% of the number of trained parameters.
    Organized the derivation of formulas, wrote the entire paper, revised it and added experiments after submission.
    Extended the framework from Bi-LSTM to BERT with experimental validation.

  4. Brain research based on deep learning interpretability

    Oct.2021 - May.2022

    Simulated the intra-brain features of monkeys during playing Pac-Man by deep learning model, and predicted monkeys’ decisions using ConvRNN. Predictive accuracy of monkey decision-making under three Pac-Man behaviors improved by 4% compared to previous methods combined.
    Gave an interpretable analysis of deep neural networks by generating the monkey’s attention through Grad-CAM.

  5. Time series data operation and maintenance analysis

    Jun.2021 - Oct.2021

    Designed the anomaly detection algorithm: using Fast Fourier Transform to compute spectral residuals against the anomalous data, and constructing a discriminator for spectral residual thresholds end-to-end for anomaly detection.
    Compared with the baseline method, it improves 36.1% and 68.8%on KPI and Yahoo datasets, respectively.
    Won Grand Prize of the 17th “Challenge Cup” National Extracurricular Academic and Technological Works Competition for College Students (Top 5 in China).

  6. Federated joint estimator of multi-sUGMs

    Dec.2020 - Jul.2021

    Trained the sparse Undirected Graphical Model methods over a massive network, and predicts several informative groups of connectomes based on the real-world dataset. While maintaining data security, the computation cost is reduced by 15x, 54x respectively compared to the Baselines.
    Analyzed and processed simulated and real data of neural connections in the brain using Nilearn.
    Plotted figures of experimental predicted results, and drew the methodological framework and schematic diagram of experimental operations. Analyzed and processed simulated and real data of neural connections in the brain using Nilearn.

My skills

  • Programming
    80%
  • Web Development
    70%
  • Communication
    90%
  • Collaboration
    80%

Contact

Contact Form