Neural-Networks-in-R

NOTES:

1. There are 2 notebooks here. I created both these notebooks on Google Colab. Google had experimented with the use of IR Kernel for a while on Colab. I feel Colab is the one of the best platforms for Machine Learning Tasks.

2. RStudio has come up with "Keras" package for R. The interface of R's Keras is similar to Python's Keras. I have attempted to solve some basic deep learning tasks in R using Keras. To use Keras within R, Tensorflow(R version) has to be installed. And for this purpose, the python virutal environment has to be created witin R. I have mentioned all these steps in the Notebooks.

NOTEBOOKS:

~R_auto_encoder

In this notebook, I have tried to simulate a simple AutoEncoder to capture the patterns in MNIST dataset. The Dataset can be obtained from within Keras. Even Colab provides a miniature version of the datatset to work with. The autoencoder in my notebook works good.

~Convolutional_Network_in_R

In this notebook, I have worked with the CIFAR datset (basically the smaller version cifar10). Firstly, I have used a small CNN model. I have mentioned the architecture of the model. This model was about 67% accurate ( which is good for my 1st deep learning model in R) but unsatisfactory. So, I resorted to Transfer Learning. I used VGG16 Model with predefined weights(of imagenet). The Input Layer and the Output block had to be changed to make the model work with the cifar10 dataset. After a bit if fine=tuning the model, prediction accuracy was about 90% (a good improvement).

Observations

During the course of creating these notebooks, I found that using Keras with R is as swift ( and very similar to ) as that in Python. But for using Keras in R, some steps are required to be followed.

In case of Deep Learning, Python is a bit faster than R and more robust, but R is as efficient and reliable as Python with regard to Deep Learning Projects.

~Sumit Bhattacharya