 Hello everyone, I'm really excited to be here in this conference My name is Lam. I'm from Vietnam from Saigon So actually I'm father of Saigon M's and currently I'm working in satify.com at the senior not a scientist so Yeah, okay, so Sorry, I have to put the smart words here because I think you to see a vectorization a lot of deep learning Noronet work yesterday and also have a good introduction from Martin. So In daily basis my work is related mostly on Text understanding machine learning and It turns out that we can apply a lot of technique in in text Mining and NLP to into other space like music fashion as well So I just want to do a quick survey. How many of you already know work to make Dr. Wegg area to make all anything to make Just raise your hand Okay, just a few of you. Okay, and how many of you already use some kind of music streaming service? Pandora Spotify Spotify yeah, wow. Yes, you I like music already Okay, so in this talk, I will give you Yeah, a very quick introduction of our work to make and how we can apply it into the music recommendation as well Okay, in this talk The breath representation about the music here is not cover on the screen You see there the lyrics if the sound special brand the way for is a cover what I mentioned about music here it just The correction of songs the sequence of songs like the sequence of course you see in the the work to back So it's mean work one word is mean one song So we try to build a map her to Convert the song I mean the word here into the vector space and you can see that take example for for the song from Adele make you feel my love and you can Convert it in to some vector is 120 a thousand dimension But actually you can see the vector is not just a number. It can be the beautiful avatar like this matrix Okay, because I will have the demo and I demo and like Cosplay for you. So I just Do very quickly recap about work to banks and if you guys want to dive into more detail just I Give you some more reference Okay, we'll do that was invented by Mikolo in Google is three years ago at I think And you see that the philosophy is very simple If you are tell me your friends, I can tell who you are. It's the same for the linguistic domain. We have very You know traditional Distributional hypothesis is you shall know a word by the company it keeps So it's long time ago. You see from 1957 So take take an example a smart way of discovery and Recommendation is my title of my my talk title and you can see that we can do work to back to learn The similarity between the contact and the given are the given context and the worst for example the discovery and in this contact we see that this smart is the Recommendation and how how can we embed it into the same one space that we call the vector space And how can we get this vector? Okay, so given a bunch of Sentences of bunch of documents a bunch of paragraphs and we can get converted into the dance Vector so I don't want to talk more about the vector and everything everything related to No, no network now because I think it's yeah, it's enough So the thing here is given one course you can head get 100 or 1,000 number It's me the vector the list of number So if you want to try to play around just get to this link. This is very well Explanation from wrong thing In Michigan University and he had very good You say the package for work and very visual inspector and you can try it on the browser It's really cool So why work to back is really cool Okay, it's very simple math even for the kids. You see that Vietnam we have very Very fabulous very Largest cave, but it's not in China We call it sundown and you can see it on the good morning America as well We just do that very simple mathematics and the second thing is With the enough data even the training data you can have them You can keep this the latent the hidden structure of Of the words you try to convert into the vector space take example for Analogies in this case the cat dog is some kind of animals the V-body Beethoven is some kind of classical music. Okay, this is should be different type of category The next thing is about Writing about what you think you can do the translation from one space to another space and With this the compact representation like Doc is just some number trees number like point two two point three five point eight five five seven point five something like that and The most important thing is it's really simple simple to get a similarity between these vector just not product to get the the score for for forgetting the similarity between the vectors Okay, so finally We have worse this song Sentences playlist your playlist or album maybe your play logs as well so Nothing new I get a lot of playlists million of playlists over the South Plough from the Spotify something like that and we can convert on the song into the vector space and Finally, this is the fun and profit part Okay, discovery and recommendation you see basically it's just the similarity The system between the V vector your vector and your your song your vector and your artist your vector and your album very easily okay, and Luckily last week I heard that there is the the top from the I heard radio They have the loser already the same approach in Spotify. They did it already three years ago or two years ago so Map everything you have now the song is a vector the play is a vector the artist is a vector The style music the general music is vector everything in into one universal vector space and you can Just query searching recommendation everything you can do with that So in this case you can see that I love Adele I get a lot of song from Adele I love me classical music. I have one vector for classical music if I have my daughter I also generate the young vector young music vector for my daughter I also generate a lot of vector for playlist and for my music style as well So this is some kind of universe universal vector space, okay, you can see that there's a lot of different Type of vector here and we have a as a song level General level album artist and user is me. Okay, so you see Long what is your music vector? Just see the beautiful avatar. Is it my music vector and you can see it's on the my Twitter avatar as well So the application Maybe the next time you get this don't need to have the cover art. You just to generate generate the new avatar for the song So you see that I already mentioned a lot of application playlist generation as well is easy just get the vector of the playlist you like and Searching around a lot of play it very similar in this vector space and You also have a lot of similar artist as well Radio station you can use some sequential model as well in this case. You may heard about INN LSTM This actually already used in Spotify recommendation system the cool thing about the application is you can even you to Expand the search query For example, you like this song. Maybe you you don't know another song But you like this song and you can expand the similar similar song for compare to this song Okay, time to demo I will use the Jupiter notebook had Mentioned by if you cannot leave guitar that Okay for preparation. I have to lot some I think I have about 300,000 songs database after loaded just the second Okay, so it's just 200 20 28,000. Okay. I Yeah, I Conveyed into the dictionary to very easily to get the song information the metadata Yes, I love the beetle and this is the data free as a bird. It's also my favorite song Okay, just check it is really good metadata okay The secret thing here is I already have the magic Transformer is take me just couple hours to train the model and to get a million of the playlist and put it into the The model I already prepared with some parameters configurable and Take it take a time Watch you just lot. Okay. Very fast Okay, Lam loves the beetle music. Okay. I know You see This is the the song name is freeze as the bird and the song vector is I Try to convert the song vector as the song into the vector space with 100 dimension for you to easily Imaging and you see that 100 number here it can convert into the 10 point 10 10 10 by 10 picture Right, because what the hell where I hate math? I don't see a lot. I don't need to see a lot of numbers Okay with month some beautiful Library like my body she born you're not already know that today we can Can convert the vector into the image? Okay, and you see this is the the song vector The free as a bird vector in 100 dimension Okay, and I try to another some Adele songs This is the make you feel my love Okay How about classical music Paganini? Okay, and you see It's also another avatar. We call music avatar Okay But the thing here is long have listened a lot of song Classical Adele the beetle. How can we convert long style music into the vector? It's really easy. Yes Several ways to do that. Okay first You just some on the collect on the song you love and the song maybe you don't really love that for example in this game my song is three song and there is some Gang bang from Madonna. I don't like it. So I put it in people songs. It's not my song So I convert this using the magic transformer. We call the part PV is mean paragraph vector and I can get my vector is mean my song avatar Okay, I can write for you Very simple, right just plus just not just do a addition and subtraction But there is a thing that there's the way we can do that even better to keep the order because you may be like the hello the hello from Adele Free as a bird from the middle But you want to keep the sequence the order of the playlist it means the first time you have to play the hello and the second song should be Free as a bird or the the universe from bureau So in this game, we need to keep the order and I have to convert the whole the the whole sequence into the vector with the The order so I convert this using that in first vector if you try to do this You can read the game team doctor back and we can convert the whole Sequence into one factor. It's a single vector To keep the order as well. So it is my recommend the recommendation for me Lam, okay, I'm like Ali freedom Paganini. Okay, the rock or something like Blackbird as well because I like bird Okay, so you see this. This is my music So the cool thing about the work to make a neural network can easily explain to the human Because the neural excited by some thing you like in this way. I try to bring out Okay, we given the my music vector is my music avatar. Can you show me which not on Excited and you see that this is not on 99 the red color here 99 Okay, so it's mean my music is defined by some neural very attractive excited about this This does the music. I'm listening to okay. There's some another no wrong something is 12 there's maybe here and also 33 is maybe here. This is my excited no wrong about my music style Okay, now moving to Recommendation can you give me? Yeah the match the magic transformer. Can you give me more similar song like Paganini? Okay. Okay. Given what song I can get Okay, nice 18 or maybe hundred song is a classical music. You see the same vectorization and For the personal line playlist. Okay. I just add my favorite song Okay, another song. I don't really like it and I just print out. Okay. You see that. This is my recommendation recommendation for my personal playlist And also we can convert the playlist into the vector like coffee playlist. It's somebody I already created the The the coffee play it and I try to convert this playlist into the vector Okay, and I see how what is a similar song for this playlist and how does this compare to my music test? I Can load the escalator to load the functions Coside similarity and we can print out the similarity of my music style my music vector and the coffee playlist given by Some one on the universe. Okay, and it's quite not really matched with my music style So you imagine you can have a lot of application. Just do a lot at for your imagination So now for the live demo I can give you a web application demo as well Let's take a while Maybe I forgot the the port number and I check that Any question? Just let the feel free to ask me. Yes Okay, let's take a while have to wait a little bit Wait, maybe don't worry. Sometimes the demo is yeah, you see the demo. Okay, so I I try to hide something but it's not give me a chance to I stop it. I love it from my court Okay Okay, you see actually the console console is also the best way Very clear. Okay. Let's not lot model. Let's master that I want to take this chance to Introduce a little bit about I can do it Yes, just Okay, you see the burger is also bad the best option Okay They're not reachable. Sorry guys. Okay. I have to restart this one Actually, I also use the LDA a little bit in this model because I try to Fick figure out how many gender how many topic a team of the playlist they the the human already try to to tag No Thank God. I'm sorry It's used to be fine before the demo. Okay Okay, maybe I turn off the Yes, I tried but this is already local Okay, right on okay One more time. Do you guys know how to route this thing work locally just have me oh Why it's already used let me check. Yes, maybe a lot of Pollution there. Mm-hmm. I check first. Okay. Yes. I'm kind somebody idea. You see Okay, I hope this yeah, this time should be fine Okay, anyway, yeah, actually we are hiring in 75 and if you want to Talk about more about the 75 you can come and talk with my colleague David from 75. Oh, you see Okay, I hope this is time to be great. Who can help me? Yes, try it. I have to turn it Okay, get the same. Thank you. Hope it work. Thank you. This is really a silent demo I ever try Okay, hello everyone. Thank you so much This is the soul. Okay. There is two are two little demos here I try to visualize on the Disney. It's quite different compared to the PC. I mentioned a mean town Instead of keep the variants From the VC for the whole structure is to keep the local structure only it means to keep your neighborhood Okay, let's see is how this work Okay, and you can see that From Adele is the Lord of Adele and I tried to You know a little bit. Yes, this is the hello song and I project on the Disney and I see some more a Michael Burbley. No Justin Burbley by but no it's maybe people try to mix a lot of Adele and and other artists as well Okay, you can try it still now moving to some another song like Tracy Chapman, baby, can I hold you? Oh, and you see this is Mary Coray. Okay, Tracy Chapman here and Not this is beautiful. Hmm and how about the Classical music I hope to see some Vivaldi Beethoven. Okay, fine. That is it That means that the data we have is quite good In in terms of classical music you see a lot of Beethoven Instrumentals, yeah, I can you you mean the demand for you. Yeah, I some Vietnamese already Paganini Paganini Okay Did it for the mapping? So for projection into 2D, but okay, you can we can do the song similarly Simality reading search by using just a list by it's very easy for you guys to get the song name and the food information Okay, you see the hero and the story. No Justin Bieber is okay the boys like you okay this guy and Let's see some Yeah, you see the alligator from Paganini is a lot of Paganini violin there. Yes And maybe you can try your laptop as well if you put my Okay, I don't know. This is the local one. I turn off the Wi-Fi Okay, so that's it for my presentation. Thank you for your patience your Help as well, and I hope next time I will have more interesting demo for you Thank you. Any question, please So Okay, yes, thank you for the question really to really good questions. So I answer the second question first Okay for the subtraction actually yes here. Yes different ways as I mentioned The first is the addition you just put on the song It's one song is one unique idea and the one was in in our vocabulary already So you put on the song if it contributed the positive meaning and then do at the least of the negative song and Then all the the vector is plus and minus plus minus at the same vector already one vector Maybe 100 dimension add on I don't really mention. Yes, and that yes, do you right? It's a tickle sign similarity just an angular distance. Yes And the for the first one is what I do here is I collect a million. I think over a million's our playlist and I can have to convert it into the unique song ID You see maybe hello out early, but hello for ideally Should be covered by someone somebody else is it should should be another song as well So I have a sequence of song millions yes, and then I Try to put it on to the work to make and doctor back as well into LDA and get 100 dimension vector for each song Each playlist as well. So and the next thing you can do the addition and you try to direct the playlist yourself as well Yeah, it is answer the question Yeah, thank you Have a