Machine Learning in general and Deep Learning, in particular, have helped us achieve things in recent days that were previously thought to be difficult or impossible to achieve.

At my workplace, we were investigating how we could make use of A.I in education and improve the lives of students. So, I along with my team started working on this idea to build a small PoC. This is when I stumbled across deep learning and its innate power. Here are my major experiences that I would like to outline.

(Be warned: If you are accustomed to reading high tech stuff, then this article isn’t for you. This is more diluted tech version of my experience.)

1. Data Augmentation

As they say ‘Data is Wealth’, it’s true to the core. But as you might think, the real world is not ideal, in the sense it does not behave as we expect it to do so. Now coming to our case, ‘machine learning needs lots of data to produce a very good model’, this is what I have been hearing since the days I came across the field of machine learning. What if I don’t have lot’s of data, what should I do? Should I abandon the project completely? No way!

That’s where this technique of data augmentation comes into play. Data augmentation is a simple technique or rather I would say ‘hack’ that lets you enhance the amount of existing data. So how does one do that? Let’s consider a simple case of image data-set. The way we enhance the amount of existing data is by applying some transforms on the images in the form of rotation, shifting, shearing, zooming, flipping etc to get new set of images. A typical example is shown in the image below where an image of squirrel (on left) is applied various random transformations to produce 4 images of the squirrel (on right).

Illustration of Data Augmentation Technique (Original photo by Caleb Martin on Unsplash)

After learning this technique, I was totally convinced with Francois Chollet’s (the primary author of Keras) words that –

Being able to make the most out of very little data is a key skill of a competent data scientist.

Just what a simple yet effective technique is data augmentation!

2. Power 👊 of DL in finding Patterns

One of the reasons, I think deep learning has become more popular is because it is very good at finding patterns. Its deep architecture makes it possible to identify good features and patterns that we can make use of. In my case, I was able to get very good results with a minimal amount of data, that too, without any data augmentation. Simple DL networks were able to pick up minute nuances between the various classes very well. This is almost far far better when compared to traditional machine learning algorithms.

3. Selfies Suck 😠

On a lighter note, yes taking selfies suck. For my project, the data-set wasn’t readily available in the market. I had to curate my own data-set. This involved taking selfies to build an image data-set. Taking selfies was almost a nightmare for me. I still wonder why I chose to collect the data-set hard way; instead, I could have recorded a video and extracted frames from it. Since I got irritated by taking selfies, I could only collect a small amount of data. With this small amount of data, I was able to experiment and realize that amazing results can be achieved with very minimal amount of data. I really wonder how those selfie-freaks pose for the camera with the varying styles of smile 😂.

Conclusion

Finally to conclude, the power of deep learning is just simply amazing. Whether helping in diagnosing cancer or building a self-driving car, the applications of DL are just humongous. I have just soaked my feet in the ocean, a long way to dive in!