What are the limitations of deep learning?
There are a few on the top of my head:
1- Deep Learning usually requires more data to achieve higher accuracy than other ML models
2- It is very difficult to interpret the behavior of every layer once the network becomes a little deep. For example: in regression you can explain the effect of x1, x2 or any other x on y, but in deep learning we cannot explain why node 3 in layer 4 has a certain behavior because of all the inputs taken from layer 0 to layer 3
3- Implementing deep learning has more steps, even though there are good libraries like tensorflow and others, when you look at it from a company’s perspective usually python standard installation comes with sklearn which means you can start other models immediately while to use deep learning you need a little bit of time and some approvals to get the deep learning libraries in the system
4- Most of deep learning architectures are for supervised learning, this leaves a few important parts of machine learning like clustering
5- DL architectures solve particular problems really well, like image classification and predicting a sequence, they can even generate data that matches the pattern of another like GANs but they are not generalizable to every supervised learning problem. You might end up doing a ton of reading and experimentation before feeling confident about a proposed architecture for a specific problem
By: Abdallah Musmar