Why do you need gradient descent in the first place?
Did you ever have this question in your mind when you were learning the basic concepts of machine learning? Gradient descent is one of the basic algorithms we will encounter in initial learning, and it has been proved to be very effective in ML. But once you start reading this article, you will have questions about what you have learned before.
At first, this question was straightforward, and later it became a big discussion among Reddit executives. I like reading these comments very much. I believe that if you are interested in this field and have a strict attitude towards mathematics, you will find it very useful.
Reverse engineering of large-scale neural network
What should you do when the developers of complex and huge neural networks suddenly disappear, leaving no documents needed to understand the network? This is not an unfounded fiction, but a common situation that actually happens, and the questioner himself has encountered it.
This situation often happens to developers, but when it comes to deep learning, it is completely new. This paper introduces several methods that scientists can use to restore the initial design idea of depth model. Some answers are practical and some are absurd, but each one adds a new perspective. Maybe one day when you are in this situation, they will help you.
Debate on TensorFlow2.0API
A lot of comments on this topic caught my attention (there were 1 10 comments when writing this article). What topic in the world can cause such a big controversy? However, when you read on, it is incredible that the debaters have completely different opinions. Apart from TensorFlow being ridiculed as "not a good framework", many people are praising PyTorch (if you have used PyTorch, you will understand that this is not surprising).
It all started when FrancoisChollet posted his ideas on GitHub, which ignited a fire in the machine learning community.
Reinforcement learning based on predictive reward
Another OpenAI project in this paper, Huo Ying Computer Training, is considered as another major breakthrough. The title doesn't look so special, but you have to understand what the OpenAI team did. As a Reddit user pointed out, OpenAI's work has brought us one step closer to a machine that imitates human behavior.
It took about one year's training time to pass Montezuma's revenge game beyond the human level, which left a very deep impression on people!