Abstract
In this chapter, we look at more advanced techniques typically used when developing CNNs. In particular, we will look at a very successful new convolutional network called the inception network that is based on the idea of several convolutional operations done in parallel instead of sequentially. We will then look at how to use the multiple cost function, in a similar fashion as what is done in multi-task learning. The next sections will show you how to use the pre-trained network that Keras makes available, and how to use transfer learning to tune those pre-trained networks for your specific problem. At the end of the chapter, we will look at a technique to implement transfer learning that is very efficient when dealing with big datasets.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
The original paper can be accessed on the arXiv archive at this link: http://toe.lt/4 .
- 2.
With computational budget we determine the time and hardware resources needed to perform a specific computation (for example, training a network).
- 3.
Remember in this case we have one weight and one bias.
- 4.
Remember in this case we have one weight and one bias.
- 5.
You can find all information on the dataset at https://www.cs.toronto.edu/~kriz/cifar.html .
- 6.
The code was inspired by http://toe.lt/7 .
- 7.
You can find more information at https://en.wikipedia.org/wiki/Multi-task_learning
- 8.
- 9.
You can check the official documentation for the function at http://toe.lt/5 .
- 10.
The term has been used by Yosinki in https://arxiv.org/abs/1411.1792 .
- 11.
You can find a very interesting paper on the subject by Yosinki et al. at https://arxiv.org/abs/1411.1792 .
Author information
Authors and Affiliations
Rights and permissions
Copyright information
© 2019 Umberto Michelucci
About this chapter
Cite this chapter
Michelucci, U. (2019). Advanced CNNs and Transfer Learning. In: Advanced Applied Deep Learning . Apress, Berkeley, CA. https://doi.org/10.1007/978-1-4842-4976-5_4
Download citation
DOI: https://doi.org/10.1007/978-1-4842-4976-5_4
Published:
Publisher Name: Apress, Berkeley, CA
Print ISBN: 978-1-4842-4975-8
Online ISBN: 978-1-4842-4976-5
eBook Packages: Professional and Applied ComputingApress Access BooksProfessional and Applied Computing (R0)