In this project, almost all of the projects involved some sort of deep learning. Deep learning has been pivotal in powering most of the advances in the last few years. However, there are obvious limitations to deep learning that we should understand before applying them to real-world situations. Here are some of them:
- Data-hungry: Usually, we don't have big datasets for every problem we want to solve using machine learning. On the contrary, deep learning algorithms only work when we have huge datasets for the problem.
- Compute intensive: Deep learning training usually requires GPU support and a huge amount of RAM. However, this makes it impossible to train deep neural networks on edge devices like mobiles and tablets.
- No prediction uncertainty: Deep learning algorithms are, by default, poor at representing uncertainty. Deep neural networks can confidently misclassify a cat image as that of a dog.
There is no notion of confidence intervals or uncertainty in predictions...