NFL draft picks, as described by a neural network
2 min read

NFL draft picks, as described by a neural network

NFL draft picks, as described by a neural network

Every year the NFL recruits new football players in a process known as the NFL Draft. Scouts watch the prospective players, and the NFL posts their lists of each player’s strengths and weaknesses.

These lists include all sorts of information, including the player’s agility, sportsmanship, size, grades, and even their body type in rather precise detail (”Robust build with burly booty”, said one scout. “Late wiggle is average”, says another.)

Could we one day send drone scouts instead? Autonomous flying vehicles that could analyze a player’s every micromovement and calculate their odds of success? Okay, that’s super hard - just finding a football field, pointing a camera, and tracking a single player is challenge enough. But as a first step, let’s see if an algorithm could at least sound like a scout.

I trained a machine learning algorithm on 3045 example attributes from the 2018 NFL draft picks. This kind of algorithm learns by example - I’ve used it to imitate Halloween costumes, metal bands, sports team names, and even guinea pig names.

At first, the algorithm struggled to get the grammar right. The draft picks are written in a particular sort of sentence structure called “understood subject”, which produces unintentionally hilarious results when an algorithm tries to imitate it. For example:

Too much his ability to find the ball
Has to be improve the frame
Trouble teams when he get into pocket
Has a blocks and allows a much benefitt
Completive and he very the seasons
Extremely can be a rusher
Has adequate at times
Plays with a throws and but is breaks
Sticky and he take on his hips
Can impressive tackle on the boundy for the gaps
Has backside and play the pocket and contact
Able to sink and powerful, but stride
Gives his size, in the pocket
Physical compressive against competition to take game
Needs to a running downhill
Extremely legs and edge down the field

After more training, the algorithm did a better job with the grammar, but still didn’t make all that sense. I’m not sure if these are strengths or weaknesses.

Slaps to improve athletic ability
Plays with a thwing and strong optimal power
Has desired -0 percent of his strength
Plays with backside to the ball
Has fluid in the run blocker
Has velocication
Doesn’t get in space
Bobwing plays on the pocket
Has ability to shed
Plays with take away
Doesn’t have the ball
Loves base
Has bowy long demeanor
Scrambly and herping body control as a pass protection
Has a percent of his backside
Will need to play with the ball
Expected to stay on touchdowns as meme inside coverage and lateral change of screams

Your NFL draft picks, folks!

Bonus experiment: I’m still having fun with textgenrnn, which lets me train first on one dataset, and then on another. I used it to mix tomatoes and metal bands; for this week, I wanted to find a dataset whose grammar was similar to the NFL draft attributes. I found one: Pokemon abilities. The resulting mashup:

Fire-type moves and arm moves and lateral moves.
Can maul strength 50.
Protection to strike downfield.
Prevents the Pokemon’s plays.
Powerful attacks against the ball
The Pokemon is protect.

Become an AI Weirdness supporter to get more of these as bonus content. Or become a free subscriber to get AI Weirdness posts in your inbox!

Subscribe now