A Look Back at Deep Learning in 2021 (Application)
January 03, 2022 | 4 min read | 195 views
This post is a continuation of the previous one. In this post, we’ll look back at the updates in deep learning in 2021, with the focus on application projects. For research papers, please read the previous post.
Here are the four deep learning applications I selected:
- GitHub Copilot
- (Background Effect of) Google Meet
The function of a protein is dependent on its structure, so the protein structure is the one of keys to understanding what is going on inside an organism. However, it takes huge time and cost to analyze the protein structure through experiments, and it has been impossible to accurately predict from the protein’s amino acid sequence (protein folding problem)—until the advent of AlphaFold(2) .
I’m not going to discuss the details of AlphaFold2 here because I’m not really familiar with structural biology, but the following figure from DeepMind’s official blog ”AlphaFold: a solution to a 50-year-old grand challenge in biology” convinces us that this is an actual breakthrough.
There is already a database of “predicted” protein structure. I can’t wait to see how AlphaFold will accelerate life sciences and drug discovery.
GitHub Copilot is a super-smart code autocompleter. Based on the context including comments, GitHub Copilot suggests the best code snippets to complete your current code. The core machine learning model of Copilot is OpenAI Codex, a large language model trained on publicly available source code from GitHub .
As shown in the video below, if you want to write a function, all you have to do is write an appropriate comment and press the tab key. Copilot will complete the function for you.
Since Copilot is not yet a perfect programmer, you always have to write a test code to verify that the autocompleted code is correct (Copilot can help it, too!). Either way, it generally reduces your total coding time especially if you are not familiar with the language/framework.
It should be noted that Copilot is also provoking controversy. As the paper ”Extracting Training Data from Large Language Models” (introduced in the previous post) points out, Codex is highly likely to memorize its training data. Therefore, it is possible to extract personal information (e.g., API keys uploaded by mistake) or to abuse codes from a repository that is protected by its license. Also, it might help spread wrong, malicious, or inefficient code that exists somewhere in the public repositories.
Airfriend is a chatbot that you can nurture to be what you like. You can finetune the chatbot with your ideal responses in an online fashion, and you can also share your chatbot with your (physical) friends.
Currently, Airfriend only supports Japanese, but it has already gained 650 thousand users due to its natural conversation and customizability.
Another great thing about Airfriend is that it is developed by an individual (@_Ryobot). The chatbot model is a Transformer with one billion parameters, so I imagine it is very hard to keep the service running for thousands of users. Other languages will be supported “in the near future”.
The background effect of Google Meet is perhaps the most familiar example of machine learning on web browsers (WebML). As you may know, Google Meet can blur or replace the background. Other meeting apps such as Zoom and Microsoft Teams also support this feature, but one thing is different. Google Meet works on the browser.
For the segmentation model, Google Meet adopts MobileNetV3-small  exported to TFLite and accelerated by XNNPACK. The MobileNet outputs a low-resolution segmentation mask, which is to be refined in the subsequent module.
Thank you for reading! I’m happy if you find this annual post useful. I hope 2022 will be a great year!
 John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W. Senior, Koray Kavukcuoglu, Pushmeet Kohli, Demis Hassabis. ”Highly accurate protein structure prediction with AlphaFold“. Nature. 2021.
 Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, Wojciech Zaremba. ”Evaluating Large Language Models Trained on Code”. 2021.
 Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, Hartwig Adam. ”Searching for MobileNetV3“. ICCV. 2019.
Written by Shion Honda. If you like this, please share!