Day 65 of 100 Days of AI
I completed reading and working through the StatQuest book I bought on Day 2 of this challenge. “Completed” is a stretch word in this context, though! I could barely follow the final chapter on neural networks.
Still, I did get an appreciation for how much of an art form building neural networks can be. I also now understand that despite the massive benefits they bring, neural networks also have critical weaknesses.
For one, the massive neural nets used in state-of-the-art products today are almost like black boxes. Understanding what is happening in a neural net with billions of parameters is tough, if not impossible.
Also, neural networks might be good at fitting training data but in areas where there has been no training data at all, it’s hard to tell what the neural net will do and whether it will perform well. (The StatQuest book author Josh Starmer highlights this as a concern in the case of self-driving cars.)
Neural networks are still incredibly valuable. We just have to use them with an understanding of what they can and can’t do.