graph TD;
    DomainKnowledge[Domain Knowledge]-->MachineLearning[Machine Learning];
    MachineLearning[Machine Learning]-->StatisticalLearning[Statistical Learning];
    MachineLearning[Machine Learning]-->DeepLearning[Deep Learning];
    DomainKnowledge[Domain Knowledge]-->BackendDevelopment[Backend Development];
    DeepLearning[Deep Learning]-->ImageClassification[Image Classification];
    ImageClassification[Image Classification]-->LabelNoise[Label Noise];
    DeepLearning[Deep Learning]-->NaturalLanguageProcessing[Natural Language Processing];
    NaturalLanguageProcessing[Natural Language Processing]-->TopicModeling[Topic Modeling];
    NaturalLanguageProcessing[Natural Language Processing]-->LLMApplication[LLM Application];
    LLMApplication[LLM Application]-->RAG[Retrieval-Augmented Generation];
    LLMApplication[LLM Application]-->AIAgent[AI Agent];
    DeepLearning[Deep Learning]-->Multimodality[Multimodality];
    Multimodality[Multimodality]-->ImageTextClassification[Image-Text Classification];
    Multimodality[Multimodality]-->VisualQuestionAnswering[Visual Question Answering];

	Skills
Focus
OS
Languages
LLM Applications
IDEs
Tools For ML
Tools For DL
Web Deployment
Database
Version Control
DevOps
Self-developed Package
Statistic Tools
Visualization

English | 中文版

• Full-time
  LLM Algorithm Engineer @
  • TEX AI
  • Built & Optimized Search Agent
  • 2025.02 ~ Present
• Research
  • Interests: Agentic RL / LLM Post-training
  • Projects
    • Cross-modal medical image representation learning (USYD Engineering VRI Scholarship 2024 Winter)
    • Topic Modeling for the Evolution through Descriptions of Applications Longitudina (TopMEDAL)
• Internship
  • LLM Algorithm Intern @
    • AI Lab
    • 2024.08 ~ 2024.11
  • Vacation Research Intern @
    • Faculty of Engineering
    • 2024.06 ~ 2024.07
  • AIGC Algorithm Intern @
  • AI Tech
  • 2023.12 ~ 2024.02

Photo by Google DeepMind on Unsplash

My NumPy-based projects have been successfully integrated into my own open-source Python library, named MLForce. This library is also readily accessible on the PyPI Community.

A robust implementation of multilayer perceptrons, entirely built upon the powerful NumPy library.

Advantages of our implementation:

• As simple as Keras

  import numpy_keras as keras

• Built-in autograd engine

  import numpy_keras.autograd as keras

• Easy to construct

  layers = [
        Input(input_dim=2),
        Dense(units=4, activation='leaky_relu', init='kaiming_normal', init_params={'mode': 'out'}),
        Dense(units=3, activation='hardswish', init='xavier_normal'),
        Dense(units=2, activation='relu', init='kaiming_normal', init_params={'mode': 'in'}),
        Dense(units=1, activation='tanh', init='xavier_uniform')
    ]
  mlp = MultilayerPerceptron(layers)

• Stable to train

  mlp.compile(optimizer='Adam', metrics=['MeanSquareError'])
  mlp.fit(X, y, epochs=3, batch_size=8, use_progress_bar=True)

  

  Loss
• Great results

  

  Decision boundary
• Capability of dealing with complex datasets (10 classes, 128 features, 50,000 samples)

  

  Smooth optimization procedure in 600 epochs

This project implements nine different Non-negative Matrix Factorization (NMF) algorithms and compares the robustness of each algorithm to five various types of noise in real-world data applications.

• Well-reconstructed effects

  

  Image reconstruction
• Sufficient experiments

  We conduct a seires of experiments, thus when developing your own algorithms, these results could act as a baseline. The results of the experiments (2 datasets × 5 noise types × 2 noise levels × 5 random seeds implicitly) are displayed in the repository.

• Flexible development

  Our development framework empowers you to effortlessly create your own NMF algorithms with minimal Python scripting.

• Mature pipeline

  Our framework offers well-established pipelines, accommodating both standard and customized NMF tests.

  For personalized NMF models, the nmf parameter accepts a BasicNMF object. You can seamlessly insert your own NMF model into our pipeline to evaluate its performance.

• Multiprocessing experiments

  We've harnessed the power of multiprocessing for extensive experiments, significantly enhancing efficiency. This approach has halved the overall experiment duration, reducing it to 30% ~ 50% of the time it would take to run each experiment sequentially.

  For a comprehensive analysis of your algorithm, our platform enables conducting multiple experiments across various datasets:

  from algorithm.pipeline import Experiment
  exp = Experiment()
  exp.choose('L1NormRegularizedNMF')
  exp.execute()

• Interactive algorithm interface

  

  

  Demo

  Note that the initial parameter in these experiments can also be BasicNMF object, allowing the direct integration of your custom NMF model for thorough evaluation and testing.

This project aims to reproduce various convolutional neural networks and modify them to our specific requirements.

This project involves a multi-label multi-classification problem. We deployed four pre-trained image models and two pre-trained text models. To enhance performance, we developed 12 multi-modal models using self-attention and cross-attention mechanisms. The project poster showcases some valuable techniques and intriguing discoveries.

This project is an experimental repository focusing on dealing with datasets containing a high level of noisy labels (50% and above). This repository features experiments conducted on the FashionMNIST and CIFAR datasets using the ResNet34 as the baseline classifier.

The repository explores various training strategies (Trainer objects), including ForwardLossCorrection, CoTeaching, JoCoR, and O2UNet. Specifically, for datasets with unknown transition matrices, DualT is employed as the Transition Matrix Estimator.

• Meaningful Loss Trends

  

  Loss Trend 1

  

  Loss Trend 2
• Persuasive Results

  FashionMNIST0.5

  Actual Transition Matrix			Estimated Transition Matrix
  0.5	0.2	0.3	0.473	0.209	0.309
  0.3	0.5	0.2	0.306	0.485	0.232
  0.2	0.3	0.5	0.221	0.306	0.460

  FashionMNIST0.6

  Actual Transition Matrix			Estimated Transition Matrix
  0.4	0.3	0.3	0.407	0.295	0.298
  0.3	0.4	0.3	0.297	0.394	0.308
  0.3	0.3	0.4	0.301	0.310	0.388

A PyTorch-based implementation that leverages Transformer architectures to enhance the handling and design of tabular data.

A framework for multimodal-multilabel-multistage classification utilizing advanced pretrained models like CLIP and BLIP.

Diagrams of implementation:

Photo by Growtika on Unsplash

• Model weights: XavierSpycy/Meta-Llama-3-8B-Instruct-zh-10k
• Finetuning framework: LLaMA-Factory | PEFT | Unsloth
• Qauntization framework: llama.cpp | AutoAWQ | AutoGPTQ
• Deployment framework: llama.cpp | ollama | TensorRT-LLM & Triton | vLLM
• RAG framework: LangChain | LlamaIndex

If you have any questions or need further information, please don't hesitate to open an issue: Ask a Question.

jixu9182@uni.sydney.edu.au

Jiarui XU