Who are you?I'm currently a CS PhD student at Stanford, advised by Prof. Fei-Fei Li. Our lab hosted the ImageNet challenge that ultimately fueled the deep learning epic as you see today.
Where have you been?I graduated summa cum laude with Bachelor in Computer Science degree from Columbia University. I was the Valedictorian of Columbia Engineering Class of 2016.
I consider myself exceptionally lucky and honored to have worked with many stellar institutions and brilliant people. Those experiences were so transformative that they reshaped my career and philosophy forever. My list of internships and collaborators includes but is not limited to:
- Google: chief scientist Fei-Fei Li, head of R&D Jia Li, and research scientist Mei Han. Researched on meta-learning techniques for neural architecture and hyperparameters. Jun. - Sept. 2017.
- Stanford NLP Group: Tim Shi and Prof. Percy Liang. Collaboration with Andrej Karpathy and John Schulman at OpenAI. Developed the concept and framework for World of Bits, an ambitious plan to set AI agents free on the internet. The white paper was accepted to ICML 2017. I also presented at the Open Source Software for Decision Making (OSS4DM) conference at Stanford. Sept. 2015 - Apr. 2017.
- OpenAI: summer internship on machine translation and then joined the World of Bits team. Fun fact: I talked to Elon Musk at a team meeting! Face-to-face! Jun. - Sept. 2016.
- MILA (Montreal Institute for Learning Algorithms): Prof. Yoshua Bengio and Prof. Aaron Courville. Research project on the Ladder Network semi-supervised learning model. Paper accepted to ICML 2016. Sept. 2015 - Mar. 2016.
- Baidu AI Lab: summer internship with research scientist Dario Amodei, AI Lab director Adam Coates, and Chief Scientist Andrew Ng. Developed and maintained DeepSpeech 2, a large-scale end-to-end RNN speech recognition system. Set a new record for character error rate in accented Chinese speech recognition. Paper accepted to ICML 2016. May - Sept. 2015.
- C++11 library design at Columbia: final project advised by Prof. Bjarne Stroustrup, inventor of the C++ programming language. Jan. - May 2015.
- Columbia NLP Group: Avner May, Prof. Michael Collins, and IBM researcher Brian Kingsbury. Worked on the Babel speech recognition initiative, which achieved state-of-the-art results on Bengali and Cantonese speech datasets. Paper accepted to ICASSP. May - Dec. 2014.
- Stanford Autonomous Driving project: Tao Wang and Prof. Andrew Ng. June - Dec. 2014
- Columbia Computer Vision Lab: Prof. Shree Nayar and Prof. Daniel Hsu. Employed computer vision techniques to analyze gravitational lens effect in astrophysics imaging. Jan. - June 2014.
- Columbia CRIS Lab: Prof. Venkat Venkatasubramanian. Developed an ontology-based knowledge engine for chemical engineering. Paper accepted to AIChe 2014. Sept. 2013 - May 2014.
Where are you heading?I'm broadly interested in any machine learning and AI topics. I focus more on deep learning, reinforcement learning, meta-learning, and computer vision, while I also have working experience in natural language processing, speech recognition, and some basic quantum computing.
What's your favorite quote?
A good traveler has no fixed plans, and is not intent on arriving.
— Lao Tzu.
— Lao Tzu.
There is an extremely fascinating algorithm in our brain, designed by Nature and implemented over billions of years by evolution. As AI researchers, our goal is to "export" that mysterious source code to the digital world. In chasing this grand vision, we become obsessed with benchmarking everything. We measure the progress towards artificial general intelligence (AGI) by a table of numbers, which we hope will increase monotonically with more compute power and clever tricks (formally known as "grad student descent"). We assume it will eventually lead us to the holy grail.
We hardly ever questioned the assumption. Yes, history does show that we can succeed to some degree by climbing the "benchmark gradient". I do not doubt the achievements we have made, but I am unconvinced that the gradient is sustainable in the long run. When we have one eye on the goal, we would have only one eye left on the path. I believe AI is not a journey upward, but a journey inward.
Perhaps we should cool down the craze, look more closely at other disciplines, inspect more deeply the machinery of our own intelligence, and search more creatively for paths that do not necessarily yield immediate gains in benchmark scores. We might discover a splendid forest, if we are willing to take a step back from the tree. When we do not aim at anywhere, we amplify our ability to reach everywhere.
I've always been an optimist on the prospect of AI. Many years from now, we will have solved AGI by a few strokes of inspiration. At that moment, I hope we can look back and smile at the good old times, "when we wandered, we were not lost. AGI was not that elusive after all."