Startups are hot. Coming up with a good name for a startup can be difficult, however. After half an hour of internet strolling, I had built myself a decent dataset with around 1,300 names for startups. I then wrote a quick Python script to delete all the double entries:
The resulting dataset consisted out of precisely 1,250 startups. I’ve uploaded the dataset here in case you want to try the following RNN experiment for yourself. I realise the dataset is quite small, but we’ll give it a try anyway and experiment with different sequence lengths and batch sizes of 10-40 and 10-20 respectively. Now that we have some data to work with, we can focus on feeding the data into our recurrent neural network. Let’s do some training using Torch! Here’s an in progress sneak peek:
The model has been trained, so we can start sampling some new startup names:
Okay, they’re certainly not all gems, but I definitely do like some of them! You can find the final sample here. These are my favourites: Aust, Rale, Cingo, Hidesy, Insidabase, Zomiva, Frant, Last, Rindeo, Oride, Iversive, Omtile, Trenibo, Sole, Synact, Fomox, Pellod, Reale, Block, DeepApit, Tacka, Pack, Zovel, Hust, Greco, Nodius, Metio, Boostal, Blixn, Alse, RaxDack, Tides, Wayto and Mitar.
And there we go, we’ve just created a bit of (potentially valuable) data using an RNN. Feel free to actually use these names, by the way. Who wouldn’t want to have a startup that was named by an AI? Also, credits to Andrej Karpathy for his excellent post on RNN’s.