15: Autoencoders
We start with a dive into convolutional autoencoders and explore the concept of convolutions. Convolutions help neural networks understand the structure of a problem, making it easier to solve. We learn how to apply a convolution to an image using a kernel and discuss techniques like im2col, padding, and stride. We also create a CNN from scratch using a sequential model and train it on the GPU.
We then attempt to build an autoencoder, but face issues with speed and accuracy. To address these issues, we introduce the concept of a Learner, which allows for faster experimentation and better understanding of the model’s performance. We create a simple Learner and demonstrate its use with a multi-layer perceptron (MLP) model.
Finally, we discuss the importance of understanding Python concepts such as try-except blocks, decorators, getattr, and debugging to reduce cognitive load while learning the framework being built.
Concepts discussed
- Convolutional autoencoders
- Convolutions and kernels
- Im2col technique
- Padding and stride in CNNs
- Receptive field
- Building a CNN from scratch
- Creating a Learner for faster experimentation
- Python concepts: try-except blocks, decorators, getattr, and debugging
- Cognitive load theory in learning