The Starting Number Nobody Talks About, and Why It Can Make or Break a Neural Network
中文摘要
即使架构、数据和学习率都完美，错误的初始化也可能导致神经网络无法学习，这是一个常被忽视的关键因素。
English Summary
Even with perfect architecture, data, and learning rates, improper initialization can cause neural networks to fail, making it a critical but overlooked factor.
Original Excerpt
A neural network can have the perfect architecture, the perfect data, and the perfect learning rate, and still refuse to learn for one… Continue reading on Medium »