Limitations of Deep Learning Models
Rather than spitting out everything Google threw at me, if I searched for limitations of Deep Learning, I want to share the difficulties that I faced in my small but happening professional life as Machine Learning Engineer.
I was developing a time series forecasting model for one client and when presenting them with the model, he asked me to explain the model. For this particular set of problem, I used Convolution-1D network and I failed miserably in explaining him the functioning of the Network. The ignominy of not being able to explain the work you do, day in and day out, caused me certain degree of existential crisis. My advise is, you need to be ready with some real world analogies, if you ever found yourself in such obnoxious situation.
Amount of Data
If you have handful of data, you should never take Deep learning algorithms into considerations. Though, neural network are hyped as something that imitates the functioning of brain, they aren’t as effective as their organic counterpart (not yet, in future maybe). A young human mind is able to differentiate b/w dogs and cats with very few samples, but neural network requires a good amount of dataset to be able to do so.
This criticism is not on deep learning alone but on machine learning as a whole. The relationship between the variables is only down to correlation or covariance only, and the inferences drawn about causality of the events is down to developer only. And human reasoning can be flawed. This could hurt model in the long run.
These are the only three I could come up with and because there aren’t any more. The every other limitation you will come across is down to the inability of the person who is working with neural networks, not with networks itself. I think Nate Silver sums it up perfectly, in his book “The Signal and the Noise”,
Data-driven predictions can succeed—and they can fail. It is when we deny our role in the process that the odds of failure rise. Before we demand more of our data, we need to demand more of ourselves.
By: Arjun Sehajpal