The ease of making a compact tech

Cheers to the first publication of mine on medium!

We all are very influenced by Artificial Intelligence and a majority of us know Machine Learning. It seems we are running after it head over heels! But, how far we come to apply these heavy training and big size models in IoT based computational units? IoT has limited resources, and achieving higher accuracy with a smaller network is highly expensive. So, how close we could compact our tech with good accuracy ?

According to a research in IEEE 2019 titled: “ Compacting Deep Neural Networks for Light Weight IoT & SCADA Based Applications with Node Pruning” ,the study of ‘Pruning filters’ for not compromising with the accuracy of ML models to deploy in IoT based units.

Here I’ll proceed with a little more brief of this study (hoping you are familiar with the basic terminology of AI and ML)

What’s Pruning filter? It is a way to remove out the lesser useful filters from a well-trained model, like removing layers (nodes) of Kernels. This paper were mainly focused on the convolution layers in particular.

Now, we can understand the key idea of “Compressing neural networks and optimizing node pruning along with Supervisory Control and Data Acquisition(SCADA) which gives industrial process to control any sequences of mechanical process.”

SCADA is a collection of system software and hardware elements and also gathers real-time data. Nowadays, it depends on accurate data sensing with limited resources in real-time
The main idea of the neural network pruning is to halt the training based on accuracy metrics and later the filters were pruned based on the Average Percentage of Zeros(APoZ) ranking of each layer filters (it have its calculations).

So, to get more involved and to understand the complete method of these experiments, reading of this paper will complete worth it.