Experiences

Reading and implement the SOTA algorithm from papers to make definition for the real requirements.

Plan and evaluate the accuracy and response time by comparison with other products, and iteration update.

On the other hand, I have landing the SOTA algorithm of Natural Language Processing to improve the user experience of blockchain ranking analysis, service robots and intelligent voice assistants.

• Machine Reading Comprehension: fine-tune with the new translation (from English) and DRCD datasets, the results are EM: 0.8289 and F1-Score: 0.8996.

• Text Correction: integrate with NER, Pinyin and ConvS2S and then improve the response time from 3 sec to 0.03 sec, the results are precision 0.9158 to 0.9415, F1 0.9266 to 0.9483, EM 0.6425 to 0.7624.

• Name entity recognition: integrated with Pinyin and huggingface. The results are F1-Score: 0.84.

• Text Similarity: replace TextCNN with Albert, and then improve performance.

• Text Classification: replace Tencent Embedding to BERT, BiLSTM and Attention.

With respect to speech separation, enhancement, speaker recognition and speech recognition, which improve the quality of voice, and integrate with name entity recognition, classification, similarity, correction and machine reading comprehension, and make the robots understand human meaning.

• Speaker recognition: integrate sincnet with PLDA of kaldi, re-design datasets and model quantization, including the results are hit rate from 0.88 to 0.94.

• Speech enhancement: replace the LSTM of DEMUCS@FB Separator with TCN, model quantization, the size from 73.4 MB to 9 MB and RTF from 0.2 to 0.08.

• Speech separation: re-design datasets, improve the performance with Self-Supervised, model quantization, the size from 11 MB to 3 MB, SISDR from 12.28 to 10.99 and rtf from 2.4 to 0.27

• Speech recognition: re-design datasets, improve the performance based on NVIDIA nemo.

Develop cyber threat analysis open source toolkits which are integrated with computer vision and based on vulnerability and malicious attack. Including malware analysis, blockchain analysis, pop-ups recommendations, phone scams, and deceptive advertising to improve the performance of some products (Clean Master and Security Master).

• R2-D2: convert classes.dex to RGB and store it as a color image with the fixed size.

• C-3PO: collaborative filtering based on hybrid user behavior analysis.

• TWMAN: open source malware analysis / digital forensics toolkit.

I have presented my research at Defcon, RuxCon, OWASP Appsec, and HITCON. At the same time, I have customs training and support experience with national security departments, governments, finance companies and universities.

Here are some related news reports with my develop experence.

• Defending Taiwan, Fighting the Pandemic with Technology (link)

• NVIDIA Developer: AI Improves the Frequency and Quality of Mobile App Notifications (link)

• Career: Machine Learning Expert (link)