Kai Zhen

Kai Zhen

About Me

I’m a Senior Applied Scientist at Amazon AGI. I work on ML/AI training and efficiency innovations. These days for LLMs, I train, ablate, and evaluate models on 100 - 10,000+ GPUs, blending speech and audio to the "new norm of interactive intelligence". I've served as the tech lead in neural efficiency, innovating and pushing the envelope of sub-8-bit (FP8, 5-bit, 4-bit, 1-bit) quantization-aware training/calibration and sparsification methods (2:4 sparsity, row-wise pruning) that are published and deployed in Echo devices worldwide: reducing model size and latency while improving accuracy for millions of users. I’ve published at ACL, EMNLP, Interspeech, ICASSP, IEEE SLT, SPL, T-ASLP, co-authored patents, and earned a Ph.D. in Computer Science & Cognitive Science from Indiana University supervised by Prof. Minje Kim, where I pioneered neural waveform coding (audio tokenization and decoding) inspired by human learning.

I sport indoor & outdoor; interact with nature: all as key for me, if not more, to approaching the meaning of life.

News

Sep 12, 2025: 3 papers accepted at EMNLP 2025 on Efficient and Robust LLM Pre-Training (with UCSB collaboration).
Aug 08, 2025: ACL’25 pruning paper featured on Amazon Science Blog: Prune Gently, Taste Often .
May 20, 2025: Intern project on LLM pruning accepted to ACL Findings.

Selected Publications

Neural Speech and Audio Codec:
[Psychoacoustics@SPL][NSC@TASLP][CMRL@Interspeech][CQ@ICASSP]
Deep Compression and EdgeAI:
[GQ@S8BQAT@SLT][CSP@ICASSP][S8BQAT@Interspeech]
LLM efficiency:
[Wanda++@ACL][Adazeta@EMNLP] [QuZO@EMNLP]

Something Fun

I like singing with or without audience (live band or shower taking).

From the Phantom of the Opera:

From "the Chinese Drama":

My vocal feature is well preserved in my neural audio codec :) More demos: neural-audio-coding.html