Are you planning to jump on the shiny Deep Learning bandwagon? Have you taken a sip of the tasty Deep Learning Kool-Aid?
There have been some incredible discoveries in the Data Science world over the past 70 years or so:
1950’s: The birth of Artificial Neural Networks, based on the mathematical model of a neuron ie the perceptron, by Frank Rosenblatt.
1980’s: The creation of the Backpropagation algorithm by David E. Rumelhart et al., and Paul Werbos.
1990’s: The introduction of Support Vector Machines by Bernhard Boser, Corinna Cortest and Vladimir Vapnik.
Early 2000’s: The development of Random Forests by Leo Brieman.
These days, we have Deep Learning, an exciting new prospect, allowing us to build Artificial Neural Networks of potentially millions of layers. Given vast amounts of data, and cheap computational power, we can now build models that require millions or more parameters to be estimated.
Given how powerful they are, why not use them to solve all problems?
Herein lies the main point behind this article: There is no ‘one method to rule them all’.
Before solving any problem, we must first deeply understand it, rather than trying a bunch of methods, or whatever is the latest fad, and hope it works assuming it will be the best solution. Once we truly understand the problem (and its constraints, limitations, nuances), can we only then begin to find the most suitable solution, starting with the easiest, and gradually increasing complexity until we attain convergence to our desired result, be it technical or business.
The important point is that some algorithms perform best in certain domains, others require vast amounts of data and computational power, while some even offer greater insight to how they work but can come at the cost of lower accuracy. Further, it’s always best to begin with the simplest techniques, ones that are easiest to explain, understand, validate and work with. For instance, Deep Learning tends to be computationally expensive and requires large volumes of data, hence won’t always be the best solution, or even viable.
In summary, we need to keep in mind the “Science” in Data Science. Remember, it’s our job to understand the problem, think about the solution and its application, in the context of the business it exists within, and then find the most suitable solution.
Deep Learning is incredibly powerful, with many amazing applications, but it isn’t a Machine Learning panacea…