emotionihilist

Netflix marketing team: So for the poster, we were thinking-

Last week I trained a neural network to generate new messages for candy hearts. They were, well, unique.

My training data was really small, though - I could only find about 360 existing messages total. So, when I decided to generate some more messages, I decided to add my favorite neural network-generated messages to the original dataset, increasing the total to almost 500. This is still really small for a neural network, but a noticeable upgrade.

The first thing I noticed is the definite upswing in the number of messages involving bears. In fact, I’m seeing worrying signs of a bear-based feedback loop that might lead to 100% bear content after a few more iterations of this.

BE MY BEAR LOVE BEAR STACK BEAR LOOK BEAR WINK BEAR HOME BEAR TWEET BEAR TAME BEAR DEARY BEAR BEAR BEAR TIME BEAR

In addition to the bears, the neural network still managed to produce some complimentary messages.

ME LOVE HAVE SWEET TANK OOO NICE CUTEY LIDS YOU RANK BEST MANE HEAT TEAM HOT GIVE

And some that are less easy to interpret.

BEE O O MAGE LOVE IN A FAN BOOK WEAR ME LET’S RIND CLOUD ME YAK O WAY FRARR ME YOU ARE BARE SWEET BOG BEELT POSWRORD U?GHCLENCY U

Use these with caution.

MEAT LOVE TOOL BIT DIRTY LOVER LOVE BAN SICK WINK OY LOVE BUM MY HAG YOU REAR FAN ME F LOVE LOVE BAG OR MY BUN COIN 2 LOVE SWEET MEAT

Did you miss the um, not-so-PG hearts from last time? Fill out the form here and I’ll send them to you.

Also, if you need more love help from the neural network, check out the pick up lines it wrote.

The question of the day: can a computer learn to sound metal?

Thanks to HellBlazer of http://www.metal-archives.com/, who provided a dataset of over 100,000 bands, subgenres, and countries of origin, I had an opportunity to find out.

I gave the dataset to an open-source neural network framework that I’ve previously trained to generate recipes, Pokemon, knock-knock jokes, pick up lines, and D&D spells. As usual the instructions were only to learn what the dataset is like and try to make more of the same. With over 100,000 entries to chew on, the neural network managed to produce results that were… well, surprisingly metal.

I give you: Band names, so far untaken, generated by neural network