And this year’s Oscar goes to… BigML machine learning
Last night’s Oscar winners were completely in line with the predictions announced on March 1 by BigML, the machine learning company where I am a strategic advisor. And when I say “in line” I really mean that: strike; 100% on the nail, in one category after another. The candidates for each and every one of the six major categories Big ML’s algorithm picked as favorites (best film, best director, best actor, best actress, best supporting actor and best secondary actress) all came in.
Magic? Luck? There was 28,125 combinations, which means a 0.00003556 probability of guessing them all by chance. Obviously, no magic, no luck, but machine learning: the result of working with the right algorithm. And such algorithm, once correctly fed with data and fine tuned, was able to predict the voting intention of the more than 7,000 members of the Academy.
Feed in the information: details about the film (duration, budget, genre, etc.), its IMDB ratings, and nominations in previous awards (Golden Globes, BAFTA, Screen Actors Guild and Critics Choice), applied to the same database from 2000 to 2017 used in last year’s predictions. In total, more than 100 pieces of data per movie. This year, the scores awarded by IMDB users were eliminated from the equation because they are complex to acquire and had little impact last year anyway. Finally, the results of the predictions were evaluated by applying them to nominees between 2013 and 2016, with very good results: the models were able to predict the winners of each category in four consecutive years with very few errors.
Seeing the result and knowing how it was obtained gave me a strange sense of predictability, something between “Yep, that makes sense” and “I told you so.” Actually, it’s just about applying the right methodology to the right data: last year, ensembles, this year, deepnets, or deep neural networks. A model for each category, which takes about half an hour to be trained by testing dozens of different networks and that ends up creating a high-performance classifier. Obviously, it’s possible for a winner to emerge in a particular category that defies the predictions, winning against all odds… but the more data you have and the better it is, the more likely your algorithms are to get it right.
That said, BigML was just carrying out an exercise, one that many other companies do: we all know that when IBM, Google or Carnegie Mellon create algorithms capable of winning chess matches against grand masters or beating brain boxes at Jeopardy, Go or poker, what they are really trying to do is showcase the possibilities of their technology. They may garner headlines in the process, but what their work is really about is helping decision-makers begin to understand the potential of machine learning.
I repeat, this year’s Oscar predictions are not magic: the key to successful machine learning is in defining the objective, in obtaining the right data, in transforming it and in carrying out the processes required to obtain a model and then evaluating it. Sadly, real life is rarely like this: businesses looking to use algorithms to gain a competitive edge will find that data is hard to find or will be in the wrong format or incomplete and may not be able to be imported into a database easily. On many occasions, the objective may not even properly be defined. As said, it’s not magic, it’s hard work. Someone has to define the task, carry it out and then use the right tools, i.e., that are able to do the job and then interpret the results. But when you have the right data and the right tools, the results make perfect sense and can illustrate the huge potential of machine learning.
And this year’s Oscar goes to… machine learning.
(En español, aquí)
