LAURENCE FANG

03 / AI ALGORITHM INTERNSHIP

Work where methods
meet the field.

燧人(珠海)医疗科技有限公司COMPUTER VISION · MULTIMODAL AI · EXPERIMENTAL ML
01

INTERNSHIP TRACK

Visual Classification Model Reproduction & Evaluation

EfficientNetV2 98.10% · StarNet 98.05%

WORK PERFORMED

  1. 01Dataset preparation, training-script adjustment and evaluation workflows.
  2. 02Pretrained weights, fine-tuning, learning-rate, optimizer, scheduler, AMP, EMA and checkpoint-resume exploration.
  3. 03Accuracy, F1, confusion matrices and inference / compute comparison.
The explored model routes are not presented as a claim that every model was fully trained under one identical experiment.
02

INTERNSHIP TRACK

Single-Video Multimodal Feature Extraction

TEXT 50 × 768 · AUDIO 925 × 25 · VISUAL 232 × 177

WORK PERFORMED

  1. 01Selected-modality extraction using Chinese BERT, OpenSMILE, TalkNet ASD, OpenFace and FFmpeg.
  2. 02Padding / truncation, valid-length and status recording, with partial failures kept explicit.
  3. 03Machine-readable NPZ and JSON outputs plus human-reviewable HTML reports.
Missing transcripts, faces or failed modalities are recorded as unavailable evidence; the pipeline does not fabricate them.
03

INTERNSHIP TRACK

INN / Feature Engineering Experiments

MNIST 90.58% · OOF +1.94 PP

WORK PERFORMED

  1. 01Python / C++ interfaces, training, prediction, model save/load and file-based workflows.
  2. 02Raw pixels, HOG + PCA, Gaussian convolution features, patches, One-vs-Rest and Bayesian ensembles.
  3. 035-fold out-of-fold evaluation, mutual-information weighting and difficult-class-pair analysis.
Leakage-like Heart Disease results caused by duplicate samples were excluded rather than presented as success.