The technical explanation about this app is the most interesting thing about it:
So the title should be pretty self descriptive for learning about the songs. We have a data set of playlists scraped from Spotify, with songs that appear above a certain threshold, and we train our network to learn embeddings just like word2vec, using playlists as sentences.To generate the playlist, we find points along the (euclidean) line between the normalized vectors of the two songs the user selects. We then find a small set of the closest (cosine distance) songs to each point on this line, and from that set identify the song with the minimal squared cosine distance from the two original songs.This yields...interesting results. Some playlists work out well. Some may not appear to make sense, but when listening to it the transitions work better than you'd expect. Some are just...strange - especially with rarer songs/genres (which makes sense).Our training data consists of 9k playlists of 30k songs, and we embed in a 200 dimension vector space. We have to keep a slightly higher learning rate than I usually see recommended (0.2 with an exponential decay). So sparsity of data is certainly an issue contributing to strange results.I have a suspicion that using a euclidean line through the vector space to draw the "path" of the playlist compared to the cosine distance for proximity comparisons may not be the best way to do this, but I can't really think of another way. My brain barely works in 3 dimensions, so 200 is a bit much.We would love to hear your thoughts! This is just an experiment - it originally started out with a focus on music recommendation, but has morphed into a playlist generator.
@chrismessina layman’s terms? Maybe say: We use deep learn8ing which is a subset of machine learning to train our network (our digital equivalent of how a bio brain learns but faster and moldable) to learn and automatically create these . The vectors mentioned are the inputs (like a bio brain’s ‘experience’) to the network and the other fancy words are the ways to variate how it may learn.
“Interesting results” mean the output just happens to do ‘x,y,&z’. We still don’t know exactly why. If we do, we should have written a peer reviewed paper on our ‘flavor’ of deep learning. 😄.
Just like Mozart made his music accessible for the Everyman and why he was so popular, and therefore legendary, it’s probably useful to speak to the Everyperson. Such as this: “We made a digital brain. Probably because we like music and everyone’s doing that now with Udemy’s Deep Learning course and Rival’s YouTube channel to reference. We applied it to mixes. Seems cool, so here’s some words to look up.” As an aside I like it!
Momixa is telling you their create mix out of two songs, so you think they will blend the songs together. Instead, it create playlists based on two picked songs.
I can see that - we will try to make it a little more clear that it is for playlist generation instead of making a DJ-type mix.
Thanks for taking the time to try it out!
Momixa is telling you their create mix out of two songs, so you think they will blend the songs together. Instead, it create playlists based on two picked songs.
Raycast
Wingman (YC S19)
Momixa is telling you their create mix out of two songs, so you think they will blend the songs together. Instead, it create playlists based on two picked songs.
Pros:nothing special
Cons:Deceptive baseline
Momixa