I am learning Convolutional Neural Networks and after studying basics such as their working, activation functions and other such details I tried to apply them on image category classification task.
I was unable to design an architecture which can perform reasonably good on my dataset and I got accuracies below 40% with upto 50 convolution, pooling, normalization and RLU layers in different configurations. When I tried GoogleNet and AlexNet, I got accuracies of 97.8% and 92.4% respectively. Can you guide me to some tutorial which can explain the considerations to design an effective neural network which can perform comparably.
Dataset: The dataset contains 10 classes and 50,000 images in total.
Dear Nisar Ahmed ,
This is an empirical and, hence, computationally expensive task. To handle this problem, recent studies are working on learning the network architecture, for example:
These previous questions/discussions can help you:
Read this book first
http://www.deeplearningitalia.com/wp-content/uploads/2017/12/Dropbox_Chollet.pdf
Hello,
See this link:
https://www.topbots.com/14-design-patterns-improve-convolutional-neural-network-cnn-architecture/
- Badges
- Science topic