Build an AI Reader - Machine Learning for Hackers #7

[@nickriker8196]

Hands down best machine learning channel on KZbin.

[@SirajRaval]

thanks so much, i'm flattered!

[@SlumdogWunderKid]

Came here from Facebook. Keep making these videos! I haven't had time to go over Google's new NLP tool. and I am willing to bet all this AI open-sourcing is only going to increase in frequency.

[@SirajRaval]

Thanks PJ, you make a good point.

[@abdousayed3485]

Thank Siraj. I am a mechanical engineer who left his field to AI. I am learning a lot from you THUMBS UP

[@TamilLatest]

Great video! Had to put it at half speed around 3:10 so I could understand what was going on. You sound like a laid back surfer at half speed lol Tensorflow FTW!

[@SirajRaval]

interesting, thanks for letting me know i will slow down future videos a bit

[@chemdah]

Really impressed with your knowledge of NLP. How do you learn all this stuff so quickly? You must be a genius!

[@FG-Supercharged]

FYI... The Github repository location for Parsey has changed to: github.com/tensorflow/models/tree/master/research/syntaxnet

[@arunk123456789]

Awesome Video! Helps a lot! I am working on a sentiment analysis project. Make more of these!

[@SirajRaval]

thanks so much! will do

[@simonstrandgaard5503]

Awesome. Easy to understand. Perfect pacing.

[@SirajRaval]

Thanks so much :)

[@niko2756]

Is anyone else having trouble getting this code to work properly?

[@JaredScheib]

Awesome! Thanks for this - been wanting an intro to PMcP. But what the new facial hair as determined by an AI!?! ;)

[@SirajRaval]

Ha!! Thanks, the next one 'AI Barber'

[@JaredScheib]

:)

[@badashphilosophy9533]

Who knew there was so much that goes into understanding words. Dependancies, something i hadnt considered. I had this cool idea for optimizing language, it just takes applying the one we have to a heirarchical taxonomy of which already has a cipher or a syntax designated to each node. Y u ask? Because if our language was plotted to the tree based on meaning then all words in all languages with the same meaning no matter how theyre spelled, how thry sound or the grammar of the language all go in the same place on the tree, so if i know that 2a7b is 'encode', a german guy who knows that codar is 2a7b then we both know what we mean when we say 2a7b without learning eachothers lanfuage. It essentiallu becomes a universal language translator. But its also an intuitive tree heirarchy ordered by troponyms for verbs and hypernyms of nouns so its way more than that, instead of creating complex maps and metadata for each words to explain relationships we simply have clusters in clusters in clusters that all nestle by meaning, semantics and common use ie a more generalised form of dependancies where verbs for example are commonly used alongside certain kinds of nouns helping classify them, i mean u wouldnt decode a horse, u decode information, u interpret information to, hence decode and interpret go under communication but decodes broader so interpretation is a lower teir, u can navigate from the grandfather node to through the tree intuitively by opening the broader terms that house the mores specific terms. In other words the relational information is built into the map itself, the words and the cipher are all the same thing and thus the information about the relationships is self evident. The potential implications for computational efficiency for both brains and computers is theoretically insane in my non expert opinion, it also potentially bridges the gap between humans and machine, coz technically u could also call it a language even though its really a framework for language. Anyway i have 0 technical skill and it would take me years to get this off the ground due to my low knowledge and lack of resources i dont have a pc or anything but if it interests u and u can see how it might revolutionise humanity, ai and basically all our systems then u can use it, u seem liek a clever dude maybe it wouldnt be so hard for u to do and maybe u could find some beneficial uses for it.

[@DanielDugovic]

Thanks for doing these videos and for publishing the code. I think you overstated your claims: how using these tools would you extract intent (or perform sentiment analysis)? Parsey constructs a sentence diagram with POS tagging, but the rest is nontrivial!

[@SirajRaval]

thanks Daniel! we can extract sentiment trivially using a human labeled sentiment lexicon see Textblob python module or harder but funner way (deep learning)

[@DiogenesActual]

How do you know the API functions wihtin Syntaxnet? Anyone have documentation on the API? I've got it up and running but I'm trying to write my own pythons scripts that can use the parser.

[@rohitsaxena22]

Nice Vlog! Try to increase the length of the video and provide some prerequistes. For sentiment anlaysis, u recommend using the POS tags generated by SyntaxNet with SentiWordNet or using the intent output?

[@SirajRaval]

SyntaxNet with SentiWordNet. Thanks!

[@piyushsinha1362]

Nice one..I like the idea of building a political debate fact checker.I 'll definitely work on it. Thanks.. Keep posting more videos.

[@SirajRaval]

Thanks so much! Glad you liked the idea, that'd be so incredible send me a link if you get around to building it :)

[@konkoism]

Hey Siraj, I'm trying to understand this in the context of the previous episode, building a chatbot. There you mentioned that one could teach a program to respond to a human within a closed domain via a generative model. But you didn't parse the incoming text like you did here, did you? So where did the machine learn to understand the user's intent in order to - for example - return the correct IT support answer? Oh and thank you for the great video content :-)

[@SirajRaval]

you're right it didn't. see this explanation explosion.ai/blog/syntaxnet-in-context

[@cult-ur-fit]

Hey there I checked out the repo and I had a quick question: What is the tensorflow transformer library doing for us?

[@SirajRaval]

its transforming our inputs into a format that is more suitable for training in our neural network. we can tune those transformation parameters if we wanted to as well

[@GeorgeArellano]

Good Shit Siraj. I guess I have to check out Parsey now.

[@SirajRaval]

Thanks Jorge!

[@anhvth7342]

amazing, how long have you been with deep learning ?

[@SirajRaval]

since the robotics lab as an undergrad so about 4 years now on and off.

[@qwerty12341244]

great video! do you know of a way to "play" with Parsey without knowing much programming?

[@Unitazy]

it's a library for Tensor Flow so you need to have some basic Python skills. Worth learning tho!

[@SirajRaval]

what Vlad said!

[@ぶらえんぴん]

It is almost perfect if you can adjust your video onscreen examples "I book a ticket to google" to match the example you were talking about: "the man feed her cat food"

[@VictorGallagherCarvings]

Even more interesting would be not only a fact checker but a logic checker.

[@SirajRaval]

word

[@CasteloOfficial]

It's what I have searching for! Amazing! 😍❤️👍

[@reijin90]

Well done; the demo output couldve been clearer (readability and time on screen).

[@SirajRaval]

Thanks! Great feedback, i'll slow it down next time

[@stupify6751]

Sir , how can I build my own api.ai using Syntaxnet? where should I start?I am getting POS tagging but what next?

[@SirajRaval]

wrap syntaxnet in a django app. build API calls for functionality that you define. upload it to heroku.

[@stupify6751]

Sir, thank you very much for your response . if u have seen the video of api.ai demo, there they have shown a hotel booking scenario, the system is able to understand what the user is asking with very little training. The syntaxnet is providing the pos tagging,how do I get to know the meaning of the sentence?is it just simple comparison. sorry if this sounds dumb, I am a noobie to machine learning.

[@knighthawk095]

Can we use Parsey McParseface in Android ? If yes then how ? And did you solve your buffer overflow problem?

[@srikanta80]

this is awesome bro.... thanks for providing us with an insightful video.

[@SirajRaval]

Thanks for watching!

[@MegaSexygayguy]

I would subscribe twice if I could. Great stuff man

[@SirajRaval]

woot! thx

[@niko2756]

will i be able to get this to run in python 3? Regardless I really like this video I'm going to start working on this right now.

[@SirajRaval]

Yup!

[@Throwingness]

When I type... >>> import demo_lib into terminal, I get... Traceback (most recent call last): File "", line 1, in ImportError: No module named 'demo_lib' Tensorflow is installed and tested

[@Throwingness]

I'm learning now. 'Demo_lib' is a module? Where can I download this module to put into the Python folder, (I know where the python folder in my machines is). I don't see this module anywhere in the Tensorflow Github page.

[@SirajRaval]

it's a file included in the repository in the description

[@sourabreddy2k15]

you are really killing it bro :) , keep up the awesome work. Still can't understand why you left twilio for this thing , now that's some true passion . But now as you have left the job do some justice , bombard ML learners like me with awesome supply of tutorials . Also you could start a kickstarter for this dude , really good .

[@SirajRaval]

Thanks a lot, yeah i'm figuring out a business model. It'll happen. Kickstarter is a good potential idea, I'm sure something will come up. I'll keep pushing content for you :)

[@hexkcd7163]

Hey, how about Patreon?

[@SirajRaval]

Hex i think that's what i'm going to do. i'm working on one right now, will have a link to it in my next video :)

[@8Trails50]

tbh, i only watch cause of the memes

[@SirajRaval]

Haha! Thanks for the honesty

[@mandeepvratesh4702]

yeah... :p

[@moshfiq]

That is one important reason for me too. Lets eat grandpa. Lets eat, grandpa. :D :D

[@hiteshvaidya3331]

I want to build a algorithm to code convertor with this. How can I do it?

[@SirajRaval]

wow, that is quite an ambitious idea i like it. it sounds similar to language translation. here's what i would do for a prototype; Apply OCR via OpenCV to an image of the algorithm. That way you get a math equation from it. It would be extracted with computer vision on the image. This problem has been solved. Then its a bit easier; converting a string equation to code. I think sympy will help you do that nbviewer.jupyter.org/github/jrjohansson/scientific-python-lectures/blob/master/Lecture-5-Sympy.ipynb

[@maniacalcactus4705]

Dat deep learning meme though

[@SirajRaval]

ONE MORE TIME :)

[@vijaykoravi7583]

hey siraj can you tell us about replika.

[@luisrevilla9173]

Amazing videos. Please keep going.

[@SirajRaval]

Thanks Luis! Will do

[@RaisesCom2]

Best channel for ML. Get a better mic tho

[@SirajRaval]

Thanks!! Good call.

[@SunilLangtad]

where is that demo_lib package?

[@machapuis]

your code works :) thank you!

[@SirajRaval]

anytime!

[@tomyoung1394]

This is very helpful, thank you

[@SirajRaval]

thanks for watching!

[@hussain5755]

you are the best siraj

[@woodstoney]

You're not cracked after all...you're just...um...hyper...no...Enthused!! That's it! Cool videos you make...Keep going! ;)

[@SirajRaval]

Indeed, it is genuine enthusiasm. Thank you!

[@dominiquerouillard316]

I changed the speed to 0.5. Solved the problem.

[@SirajRaval]

lol im getting slower now

[@imaginist0806]

Nice! you provide really interesting information about AI... Try talking a bit slower, so that your information is conveyed even better. good luck!!

[@SirajRaval]

will do thx

[@griesrt]

This was filmed at LinkedIn

[@kinestheticlearning]

When you say "from left to right" and wave your hands in that direction starting from 3:00 it would be really nice if you could first 'flip' the part of video depicting you *before* you apply chromakey (and add other elements in video). Because that way you would be seen waving hands in the left (or right) direction, as it is from viewer's perspective; I'm just saying it because I fall into a category of people (that includes 19.5% college professors and 26.2% college students, according to this faculty.washington.edu/chudler/java/hands1.html ) who have trouble telling right from left; and adding perspective problem, it makes it even worse ;-)

[@SirajRaval]

this is great feedback, thanks i will do this next time

[@vivasjimmy]

nice thanks Sira

[@SirajRaval]

Anytime

[@adriannajera8013]

could it be used for spanish language?

[@SirajRaval]

yes

[@adamshamsudeen6808]

Is there any python version of this code?

[@SirajRaval]

yes github.com/livingbio/syntaxnet_wrapper

[@carbonchar]

Great Job (Y)

[@SirajRaval]

thanks

[@lkd982]

You are so damn smart, it ant natural. Are you just showing off? Btw, the production for these videos, including the memes and skits: did you do all of that, or do you get help? Again - it's not natural! Thanks for these, they are absolutely brilliant, it's a wonder you haven't been snaffled by the Illuminati! Q~~~~

[@SirajRaval]

thx its all me

[@lkd982]

Staggering indeed. But could you answer this? Why learn ML when Musk et al are warning about AI supremacy? And (as a less sinister prelude) even before that, won't there soon be ML models capable of doing their own domain specific ML? what is your understanding?

[@akpronet]

While i commend your efforts and I'm subbed. I can't help but seem to think you gloss over some important details. Parsey McParseface is for parts of speech understanding. That does not equate to identifying intent of a sentence, thats a much harder problem and while your demo shows a very simple example, real world scenarios will fail at recognizing the correct root word and in somecases there are multiple root words. I suppose anyone who has actually used pmcp would know this, but still its kinda misinforming people.

[@SirajRaval]

right, pcmp helps intent recognition tasks by providing POS tagging. perhaps i needed to make that clearer. thanks for the feedback and for subbing! getting better over time

[@christianhuerta4103]

How do you build a political fact checker?

[@Earzone63]

Google search has just been updated to do political fact checking now

[@gliasmarinho]

wtf is that