Replace a window with a digital display, put one or more cameras on the back of it, then apply transformations to the outdoor view: make a rainy day sunny, make a short day longer, or replace a brick wall with a stunning view. Difficulties: It seems like what you really need is a smart hologram?…

Follows along in the text as a group reads or performs a play.

Use a CycleGAN-like approach to transform scat-style singing into a musical instrument solo or into entire songs. Compare to TimbreTron

A system for novel-writing. Mostly just lets people set variables and intelligently refer to them within the text. This would require localization-like ability to adjust the case and number of the variables. Could provide a user interface that makes this easier for certain classes, e.g. Character, Location, Event, Relationship, etc. I’ve been using a combination…

Yakkity uses a non-ML algorithm which produces mediocre results. Improving the algorithm is the top priority. An easy win would be to train the weights by which it ranks candidates. This would require only labeling a few hundred candidates and performing least squares regression. A harder win would be to make a generative neural network…

Gather a bunch of SNES images. Downscale them to the NES resolution. Train a de-convolutional (transposed convolutional) network to translate the NES scale to the SNES scale. Then apply this network to NES games to get a SNES-style upsample. If necessary, embed this in a GAN until the upsampled NES games are indistinguishable from the…

Generate state-action-reward triples en masse and directly model the q-value function. Add Monte Carlo Tree Search. AlphaGo cost google around $35 million, it may be hard to reproduce that success. But, could I make it do something reasonable? https://github.com/joshhansen/Umpire

A vector space representing linguistic phonotactics across languages. Given a set of phonemes in a language’s phonemic inventory, but with N removed, predict which N phonemes are missing. Turn the weights used in the prediction into a vector. Then generate random vectors and convert them back into phonemic inventories somehow. See also Concept: Artificial language…

Flattening a vocabulary to a single number (word ID) such that the distance between ID’s best approximates the semantic distance between the words. Initialize: Refine:

In addition to the features outlined for Concept: Maximum Basic Goodness 2: