苏剑林(Jianlin Su)

Bert4Keras. I made a cleaner way to use transformer AI models in Python.

ReRoPE. I built a technique that extends how much text AI models can read without retraining.

SimCSE Chinese. I tested a sentence-matching model on five Chinese datasets.

FSQ. I built a Keras implementation of Finite Scalar Quantization for compressing AI models.

CoSENT. I made a sentence comparison tool that outperforms industry standard approaches.

VAE. I built a simple image generation tool using variational autoencoders.

Word Discovery. I built a tool that finds new Chinese words in text faster than published research methods.

T5 in Keras. I simplified how to use T5 language models in Keras for real projects.

BERT Whitening. I made a tool that improves how AI models store text summaries.

RAdam. I wrote a smarter optimizer that trains AI models faster and more reliably.

Chinese Gen. I collected Chinese text generation models with setup guides.

Perturbed Masking. I built an unsupervised Chinese word segmentation and syntax analyzer using BERT.

LaBSE. I converted a multilingual AI model that compares sentences across 109 languages.

KG-2019. I built a model that extracts relationships from text automatically.

SPACES. I built a model that summarizes long legal documents automatically.

Attention. I wrote working code for the attention mechanism that powers AI models.

CLUE Benchmark. I built a testing suite that measures how well AI models understand Chinese.

Text Match. I built a model that matches short messages by what they mean.

CRF. I wrote a compact tool that labels sequences of text automatically.

Entity Linker. I made a model that identifies entities in text and links them to database records.

BERT in Keras. I built ready-to-run examples of BERT language model tasks in Keras.

TextScan. I built a tool that reads Chinese text directly from images.

Capsule. I built a pure Keras capsule network implementation with faster routing.

Infomax. I built a tool that finds the most useful patterns in data automatically.

Lookahead. I made a training technique that helps machine learning models learn better.

Tiger. I built an optimizer that trains AI models while using less computer memory.

NLP Zero. I made a text understanding toolkit based on a simple math principle.

Margin Softmax. I built an AI training method that improves how models recognize similarity.

P-tuning. I built a simple experiment applying P-tuning methods to Chinese language AI models.

R-Drop. I built a method to make AI language models learn better on Chinese text tasks.

Gradient Accumulator. I made a tool that lets you train AI models with bigger effective batches on smaller hardware.

Flow. I built Keras implementations of three flow-based image generation models.