Vu Hong Quan

Vu Hong Quan

Data scientist

[En] Facebook’s research on decoding speech from EEG

less than 1 minute read

Published:

Facebook research on EEG and speed
At the end of this August, Facebook/Meta’s AI team published a preprint of their research on brain electrical signals (electroencephalogram-EEG).

Speech is a complex activity of intelligence, so studies of how speech is processed in the brain often has to be worked with invasive EEG signals. Skull surgery is needed to place electrodes directly into the brain. This method is somewhat horrible and cannot be widely spread. However, non-invasive EEG signals, measured in the scalp, are often very noisy and weak, and there are differences between individuals, even between sessions of the same person.

Facebook’s research takes a significant step forward in decoding speech from non-invasive EEG signals. The network architecture uses wav2vec 2.0, contrastive training with the CLIP loss function, and notably a set of CNN layers superimposed on Subject layers. The subject layer is described as performing per-electrode parameterization on Fourier space. Hopefully the source code of this article will be published soon.

Link: https://lnkd.in/gBhVHBKj

[En] Can deep learning conquer brain signal processing?

4 minute read

Published:

Poster
Deep learning (DL), the science based on the idea of how the brain works. Since Hinton claimed that he understood how the brain works and invented deep learning, the related sciences have changed their approaches dramatically. In 2012, the AlexNet, developed by Hinton’s students, achieved SOTA results in image processing. Since then, traditional image feature extraction methods have gradually been forgotten. Nowadays, when someone talks about image processing, people immediately think of large-scale deep learning networks with stacked CNN layers.

[Vi] Deep learning và xử lý tín hiệu điện não

5 minute read

Published:

Poster
Deep learning (DL), ngành khoa học dựa trên ý tưởng cách hoạt động của bộ não. Kể từ khi Hinton tuyên bố mình hiểu cách não bộ hoạt động, cùng với sự ra đời của deep learning, các ngành khoa học liên quan đã thay đổi cách tiếp cận của mình rất đáng kể. Năm 2012, mạng AlexNet, phát triển bởi các học trò của Hinton, đạt kết quả SOTA trong xử lý ảnh. Kể từ đó các phương pháp trích xuất đặc trưng ảnh truyền thống dần chìm vào quên lãng. Ngày nay khi nói đến xử lý ảnh, người ta nghĩ ngay đến mạng deep learning kích thước lớn với các lớp CNN xếp chồng.