Working Remotely as Machine Learning/Deep Learning Engineer
Hi, I want to share my experience while working remotely as a Machine Learning or Deep Learning Engineer in a company located in Jakarta Selatan, Indonesia. For finding the official site of my company that I worked for, you can click this link http://tigapilarmandiri.com to get more information.
I’ve been working on this company since Jan 2018 as Machine Learning/Deep Learning Engineer. Most of my work is in the Computer Vision field to solve a problem for detecting similarity or originality of physical vehicle identification number (VIN). In my country Indonesia, the process to check the VIN for machine number and body of vehicle number is still use manual check. To check the originality, we use pencil and do the drafting in a special paper that will stick in a machine or vehicle body and we begin to draft the area with a pencil. After that, the paper will be saved for later usage to compare with the new paper that uses the same process to take the VIN.
This paper then will be compared by the officer for the originality. Maybe you would ask a question? Why you do that to check the originality? Or does there were a problem with originality?
The answer is yes, we have a problem, and the problem is the thief who faked the VIN to get legality of their document to sell their theft car or motorbike. As a human, we can do error and we call it human-error, so to reduce the human error in the process of checking the originality, we want to have an assistant that is intelligent enough to help Police officer to identify fake or original motorbike or car by its physical check.
To solve this problem I use Deep Learning that has been proven to be working well in any application. At first, I used Keras to solve the problem but, I got only a little control to train my neural network, so I try to use tensorflow, but with tensorflow I got more problem. Tensorflow is not intuitive and the API is too low level, so I need a lot of time to develop and train my neural network from scratch if I use pure tensorflow. And then came the solution through pytorch as a deep learning framework, I very glad to have pytorch to solve my problem. It's really intuitive to develop and train my neural network with it.
There are three phases for developing originality of VIN, the first is getting the bounding box of a physical check and crop the image, the second is segmenting the image using UNet to extract the real text from pencil draft, the third is to find similarity of the old document with the new document using the siamese network.
We use 5000 training image that has been built by our data officer that was gathered from the Police Officer in Polda Metro Jaya Jakarta Indonesia. From the 5000 images that have been take and created, we filtered again that is clean enough to be training data to improve accuracy. After getting accuracy 99.6 percent with dice coefficients and binary cross entropy loss, I build the library that will be used later as web service.
The model than deployed in Google Cloud Compute Engine as web service with Django Web Service, then it uses as ML as Service for the front end users to check the originality of the document.
For me, it was a really exhaustive, fascinating, awesome, and stressful work. Why I said that first, because I work remotely and it means the amount of information is not good as it is when working on site, but thankfully I can adapt with it. Secondly, the challenge to solve is very hard, because we want very high accuracy to ensure the originality of the document is not easily cheated by a thief. And the third is, it's really awesome and fascinating because this field is my passion, I love to work with the neural network as my hobby is to explore mystery about the brain.
I really love about the brain and everything about it, because what we experiencing as life is inside the brain itself. The question about reality can be answered if we know more about the brain and how it really works, that why I love to dig deeper more into the neural network and the brain. For me learning about the brain and neural network is for quench the thirst of knowledge about brain inside of me.
