Hello, I am a master student in Shandong Province, China. My research direction is image segmentation. I try to apply deep learning to image segmentation, but I have read a lot of literature and never understand the principle of deep learning, However, it is not always the case. What is the specific work of deep learning and what has been learned. If deep learning is a black box, how can we optimize it? How can we optimize a black box model to improve its accuracy? This is the question I have been wondering. If you don't know how to work, how to optimize the structure? I have read a lot of literature, but I just simply introduced the structure and gave a structure diagram and experimental results diagram. I feel this is unscientific. Hope to get your reply
Have you looked into backpropagation yet? This is one of the commonly used optimization techniques when training a neural network. It will probably be a good starting point for gaining some intuition about how NNs get trained.
In principle, you could look at NNs as non-linear functions with a large amount of parameters rather than black boxes. The goal of training such networks is finding the optimal parameter values such that a given loss function is minimized. This, in turn, should translate to better performance on a given task.
I think you shouldn't get too fixated on the "black box" term. That's just what people say because there're many params involved in most of the big models.
But the basic building blocks are often simple enough to get. I'd suggest to start with building very simple CNNs and RNNs. While doing so, try to really understand what individual components do. I'd highly recommend "Grokking Deep Learning". This book will guide you through building many popular architectures from the ground up.
Also, there're projects that could help visualise features learned by a model. For example, https://github.com/slundberg/shap
But I'd suggest learning how to build simple networks from scratch first.
- Badges
- Science topic
- Similar topics
- Papers