“What is the unique insight?” This question kept coming up when introducing our startup idea. We have three, but to be fair each by themselves are not unique (ask any data scientist). The uniqueness comes from combining them.
1. Most machine learning problems don’t require complex models
There are a lot of buzz and excitement around state of the art machine learning models and methods (i.e. Deep Learning, neural networks, ensemble models etc. ) but the fact is most problems can be, and should be, solved with simple models. (think logistic regression, random forest, SVM).
Focusing on simple models has the additional benefits of the results being more understandable (‘interpretability’) and, requiring less data and resources to train, which happens to perfectly fit the needs of our target audiences, small to medium size companies.
2. Most problems don’t require 95-99% accuracy.
Sure, there are problems where low performance of a model is not an option (i.e. medical), but vast majority of problems, especially entry level business problems, will do with 80%, or sometimes 70% or even 60%. Often machine learning models are competing with random or “human intuition” and 60% is still 10% better than a coin toss and can create revenue (‘value’).
3. Machine Learning Model performance follows the (hyper) Pareto Principle(AKA, 80:20 principle)
Its really REALLY hard to hit 99%. (In fact, its so hard to reach high accuracy that when data scientists find that their model is above 95% they assume something is fundamentally wrong.) A kin to the speed of light, it is impossible to reach 100% accuracy in machine learning and you need to pour in exponentially ever more resources as you get closer to 100%. But hitting 80% is actually pretty easy. Most data scientist will tell you that hitting 80% with the first trained model isn’t uncommon. It takes 20% of the effort to reach 80% performance. No, in fact, commonly it actually takes far less than 20%.
Adding these 3 points together, we believe we have a winning thesis that allows us to build and provide machine learning models for use 10 times faster and 10 times cheaper then conventional methods, and, as a result, make the power of machine learning accessible to smaller companies that can’t afford to hire an army of data scientists.
Is our thesis correct? We believe it is and we hope to build a great company around it and grow in the process.
image from : Rawpixel.com