Off late there have been a bunch of startups popping up across India that claim to have built and deployed algorithms that learn, become smarter & self improve with time. These cutting edge technologies are called “Machine Learning Algorithms”. More often than not, those three words are used to lure investors into investing in a startup/product with its core algorithm barely proven. It is used as a buzzword in pitch decks almost as “Invest-baits” for investors. Unfortunately, most investors buy into that because they do not understand what it means and they’ve been told by western counterparts that it’s the “next big thing!”
To give you a couple of examples of applications of machine learning, they are used in decision making, spam detection, data farm management to manage power consumption, prediction of gene causing autism, credit card fraud and even Snapchat’s filters to put dog ears on your face. Here’s a list of ten ways machine learning algorithms are being developed and here’s another six.
So what really is machine learning and how does it work? This post is my attempt to break down Machine Learning with an analogy investors understand. Here goes:
Think of an algorithm as a startup’s financial model. The financial model is built using certain assumptions with a bunch of formulae linked with each other to give you profit and loss, balance sheet and cash flows. This system is called a “Model”. Now, investors appreciate that the projections/predictions made in the Financial model are only as good as the Model designed for that particular company in that particular industry. e.g. A financial model for a lending company will not have the same assumptions and formulae as that of a car company. Building a near accurate Model is crucial to predicting what returns you are expected to make in the future. However, if your basis of building the model is inaccurate i.e. you use the wrong assumptions and formulae, your whole model fall apart. You will never be able to predict your likely returns.
Now think of a Machine Learning Algorithm in the same way. If the assumptions and formulae used to build the Model of the algorithm are incorrect, the whole machine learning algorithm falls apart. No matter how accurate the data is that you feed the algorithm, the algorithm will fine tune its assumptions and formulae incorrectly. The results achieved in such a situation will be completely out of whack (just as most financial models one receives from most investors these days).
My suggestion to investors who evaluate startups/products with a machine learning component to it:
Hammer down the model of the algorithm. Hammer down the assumptions made while building the algorithm. Stress test the model of the algorithm. Most entrepreneurs claim that the algorithm is the secret sauce of the product and hence they cannot share the information (hogwash). That’s just a way for them to get away with showing/telling you what is actually going on with their algorithm. Ask them what’s the accuracy level achieved with their algorithm. Most accuracy levels won’t be more than 50% (which is as good as flipping a coin).
Also published on Medium.