Data Science, Machine Learning and Deep Learning – The buzz words of this decade are often misinterpreted and misunderstood and the hype around them does not help the cause.
While there is certainly quite a bit of overlap among the three, like all 3 cannot function without data, there are some differences worth noting.
Lets take cricket for example.
Data Science will have access to all sorts of cricketing data by collecting, storing, processing, visualizing and drawing inferences. Say for a match, it will tell us whether the fast bowlers or the spinners were more successful, whether the openers did well under certain conditions like dew or under lights, whether the fielding position change helped take a wicket or not and so on. In other words, we can visualize the data and draw inferences on what went well for a team and what did not based on this particular match and probably also compare it with historical data and draw further conclusions and thereby create new or edit current strategies for the upcoming matches.
Machine Learning is data hungry because the objective is different. Our objective here is to simulate on what will happen in future matches given the vast amount of historical data. We are not bothered about the underlying relation between the variables. All we need to know from our ML models is, for certain conditions will it work in our favor or not and thereby select the appropriate team. So for ML to work, we would need tons of cricket match data and after going through the initial Data Science steps of collecting, storing, processing and visualizing we are now ready to model the same. But that is not going to be easy as we need to try out multiple models, tweak hyperparameters and if necessary, go back to the previous steps of collecting, storing, processing and visualizing through the iterative process. Once we find a good model by training and testing it, we try to make as accurate as possible and finally deploy the same.
Deep Learning like Machine Learning is also data hungry but goes a step further. Imagine having a robotic bowler who is bowling on a real match and learning with every delivery on the fly just like a human bowler would on what went wrong or well and therefore based on the surrounding situation, what should the bowler bowl next? Bouncer, yorker, good length and so on. So for Deep Learning model to flourish, obviously we would need tons of historical data to train the neural networks.