Hey there!
So, since the last i posted anything on steemit, as i'm busy doing some research here in Taiwan..... But, the wait is over, the results of the AI Meets Beauty- Perfect Challenge results has been released publicly a few days ago!
If you've not heard about the competition, it's basically a challenge aimed to perform Object Recognition/Retrieval. Given a 'real-world' image, the task is to search through 500,000 images and find the EXACT match of the item.
This was a really difficult challenge, my teammates and i tried several different methods to see how we could synergize and make use of different features from the dataset to effectively label / classify objects into its specific domain.
On my end, i tried to use a Denoising Auto-Encoder network, with the aim of learning the latent features of each specific object, while hoping that the noise that i'm introducing would provide some kind of robustness towards the latent features that is learnt by the model.
P/s: sorry for the cheeky link above which will NOT work
Sorry for the low quality image as i was running this through teamviewer, hence the drop in quality...
So in the image above, basically, the picture on the right is the original object, while the picture of the left is the "reproduced" image after introducing some noise...
Since i had not much prior experience with machine learning programming, i was just scrambling some codes i found online, and trying my best to create a usable model.. The following images basically shows my network's configuration:
Oh yes, on hindsight, i'm aware my AE network is not deep enough, and probably did not contain enough latent features ... which meant i was encoding them into insufficient number of features.... but i was learning how to use the network... so i guess that's a lesson learnt?
At this point, you may be thinking, what am i talking about... what is an autoencoder, noise? what noise? i don't hear any music or loud noise around me...
Fear not!
This image would be able to explain what i'm talking about slightly better:
How about Auto encoder? what is that?
This picture would be able to roughly explain what it is:
Basically, the aim of an autoencoder is to take an input image, and output the exact/similar image.... while reducing the amount of data required to save the image... in other words, a dimension reduction method, so that only the "bare essential informations" are kept.
Anyways, back to the results....
So yea, the results were finally released.... and unfortunately, my teammates and i DID NOT manage to get the top 3 positions.... however, we managed to snag the 7th place.... so it's not too bad i guess?
Here's the people from Team ViPr : http://viprlab.github.io/people/
The variance between the winning team against the 2nd and 3rd place is HUGE .... thankfully, the 1st, 2nd and 3rd place winners would have to upload their codes to github (in order to win the prize) ... we'll be able to dissect their codes once it is made available....
Oh yea, if you would like to see my codes... i can probably upload them to github, just say the word, and i'll post it up :)
though, it may not be really useful as of now, some 
[p/s: i hope it's okay to tag utopian-io, haha... tell me if i should remove it :) since it was not really specified on their page, or at least, i didn't read it]
All in all... i'm thankful for the opportunity to pick up some machine learning programming skills which would prove to be useful in the coming years.
I guess that's all for now... i'll be sure to post up more things in the coming future, but for now.... back to reading research papers so that i can finally suggest some useful topic that my internship-supervisor would be interested in...
sigh
-- Signing off
BrightTux
P/s: if you're like me, only posting up things occasionally for now... do sign up for a smartsteem account using the banner below... and sell your votes, to gain some passive incomes ;)
winks.jpg
