Over the last few years the number of startups pitching machine learning as part of their special sauce has increased dramatically. Just a few days ago Tomasz Tunguz asked the question Is Machine Learning Overhyped? and argued it wasn’t.
Machine learning is the subfield of computer science that gives computers the ability to learn without being explicitly programmed
So, if a computer is the figure out something without being told exactly what to do, there must be data with patterns that provides the basis for the learning. Here’s a Udacity video titled A Friendly Introduction to Machine Learning.
How does machine learning work?
Take a given set of data that’s known or correct (e.g. these are 10,000 pictures of cars) and have the software classify what it finds (e.g. size, shape, color, etc.) in this “seed” or “learning”, or “training” data (also called “supervised learning”). Now, give the software the overall data to analyze (e.g. 1 million pictures of anything) to find which ones have a high probability of being like what the classifier found (e.g. they have a car in the picture). Historically you’d have to “tell” the software what a car looks like in a photo. With machine learning, you give it examples to learn from for classification and then have it apply a probability of a match to each item in the overall data.
What’s a business example of machine learning in action?
Take SalesLoft and sales engagement software. In SalesLoft, sales reps are engaging with their leads via phone, email, social, etc. A common business use case is optimizing for when prospects are most likely to answer phone calls and then to optimize the engagement to coordinate the reps to make calls at those times. Using the SalesLoft data, machine learning would be trained with historical data of all phone calls that turned into conversations (as opposed to no answer, voicemail, etc.) with a variety of fields like time of day, day of week, day of month, day of quarter, timezone, job title, industry, etc. Next, the machine learning would trained on all the historical calls that weren’t answered using the same fields (the “error” data). Finally, the machine learning would process the “good results” and the “bad results” to find patterns and recommend the optimal times of day to make phone calls such that the quality and efficiency of time spent on the phone increased dramatically.
Why machine learning now and not 10 years ago?
Computing power has improved, data processing is better, proven open source machine learning libraries are available, and awareness has increased. Put another way, the challenge and complexity of doing machine learning has gone down 100x. With a lower barrier to entry and higher value, more startups are going to implement machine learning.
Machine learning is real and opens up a remarkable number of new opportunities to use computers to provide insights in a way that wasn’t possible before. Look for machine learning to grow in importance and the startup world to capitalize on it.
What else? What are some more thoughts on machine learning and the startup world?