Fund Name Generator

Less Random More random
Short explanation:
  • Start with a long list of names as a corpus to train on and pattern after.
  • From each name generate many truncated string fragments of e.g. 1-50 characters starting from the first character. (i.e. from 'Federated' generate 'f', 'fe', 'fed', etc.)
  • Train a recurrent neural network to predict the last character in a chunk based on all the previous characters (about 72% accuracy).
  • To generate new fund names, start with some characters (or empty string).
  • Use the RNN to estimate the probability distribution of the next character based on the string so far.
  • Sample a character from the generated distribution. (Use the temperature to 'unflatten' the distribution. Temperature=1.0 uses the computed distribution, lower temperatures bias the distribution in favor of less random, more frequent outcomes.)
  • Add the sampled character to the end of the string.
  • Predict the next character using the updated string; update the string; repeat until an 'end of string' character is sampled.
Further reading: