Zeqi Li Email and Phone Number

Computer vision deep learning research in model compression and generative models

• Researched the field of meta learning by reading a series of insightful papers• Implemented 2 state-of-the-art meta-learning models, SNAIL and RL2, by directly following the corresponding papers in TensorFlow• Conducted various experiments including multi-armed bandit, tabular MDP and visual navigation on VizDoom to gain more insights and understanding on both models• Documented the findings and analysis in NeurIPS format to better convey the work and the contribution to the field

• Designed a program to classify different types of ferrosilicon product and record in & out transaction of product storage room, deployed to every production workshop in the company.• Automated the process of assigning each product to its appointed storage area set by the admin of each storage room• Realized dynamic adjustable classification criteria including element content and prioritization of element via external configuration file to satisfy the fast-changing need of the… Show more • Designed a program to classify different types of ferrosilicon product and record in & out transaction of product storage room, deployed to every production workshop in the company.• Automated the process of assigning each product to its appointed storage area set by the admin of each storage room• Realized dynamic adjustable classification criteria including element content and prioritization of element via external configuration file to satisfy the fast-changing need of the company• Developed memorization mechanism to restore program to the state before last closed and duplication avoiding mechanism to guarantee integrity of the recording Show less

• Designed a full image detection pipeline for cell junction detection used in automating biopsies• Implemented a complete framework on top of TensorFlow and TFlearn for preprocessing input images, configuring neural networks and tuning hyperparameters• Trained a 5-layer convolutional neural network on an oversampled and augmented dataset (positive-labeled images: 624, augmented positive-labeled images: 3120 and negative-labeled images: 3744) to learn the features of the cell… Show more • Designed a full image detection pipeline for cell junction detection used in automating biopsies• Implemented a complete framework on top of TensorFlow and TFlearn for preprocessing input images, configuring neural networks and tuning hyperparameters• Trained a 5-layer convolutional neural network on an oversampled and augmented dataset (positive-labeled images: 624, augmented positive-labeled images: 3120 and negative-labeled images: 3744) to learn the features of the cell junction• Fine-tuned the initial model to fit with the original dataset (around 60,000 images) to address data imbalance issues (positive-labeled images : negative-labeled images = 1 : 100)• Evaluated an optimized model based on metrics such as recall, precision and F1-score to comprehensively understand its performance. Final model reached precision: 98%, recall 97% and F1-score: 97%, which is beyond expectation by a considerable margin. Show less

• Designed a program to classify different types of ferrosilicon product and record in & out transaction of product storage room, deployed to every production workshop in the company.• Automated the process of assigning each product to its appointed storage area set by the admin of each storage room• Realized dynamic adjustable classification criteria including element content and prioritization of element via external configuration file to satisfy the fast-changing need of the… Show more • Designed a program to classify different types of ferrosilicon product and record in & out transaction of product storage room, deployed to every production workshop in the company.• Automated the process of assigning each product to its appointed storage area set by the admin of each storage room• Realized dynamic adjustable classification criteria including element content and prioritization of element via external configuration file to satisfy the fast-changing need of the company• Developed memorization mechanism to restore program to the state before last closed and duplication avoiding mechanism to guarantee integrity of the recording Show less