Deep Learning – What I am excited about and what I am uncomfortable with?

It is 2022, the period of fourth industrial revolution – I don’t have to ask what I need, as soon as I enter a shop they know what I want to buy. I want to rent a house and my bot does the negotiation with bots of house owners and presents me with the best deal. As I enter my smart home, it creates an environment based on my mood. I switch on my TV and the TV knows exactly what I want to watch. My food is automatically prepared by smart oven based on the exact amount of nutrients I need. I don’t have to queue up in petrol stations, the driverless cars does itself. 

Back to 2018. Perhaps these are some of the reasons why everyone is so excited about what deep learning can do. We are living in the time where we’ve begun to see excitement and enthusiasm to go through the dirty csv files buried in the Unix servers or abandoned tables in some database for it can tell an untold story - story that can help organizations take data driven decisions, story that can act as competitive differentiatior.

Even for people who have been working in Analytics lifelong, #deeplearning brings lot of excitement – I attribute it mainly to the rise of open source, big data and cloud computing. Firstly, the pace at which open source evolves is incredible. The moment you build your models and take it ready for production, the next version releases. For eg. newer version for Apache Spark releases almost every quarter. Remember, “simple models with massive datasets are more powerful than complex models with less data”. The big data infrastructure, GPU’s based servers provide amazing capability in terms of processing massive datasets which was never possible earlier. Computation that took days would possible be done in minutes even for complex deep learning network using frameworks like Tensorflow and Keras. Dockerization allows data scientists to spin a Hadoop cluster in few minutes - bringing immense agility in model development. Serverless architecture is the next game changer – you only need to worry about the code not the server not the load balancer. AWS Lambda only charges 0.2$ per million request which makes server less architecture very lucrative. Moreover, every cloud provider now provides platform, frameworks and services for machine learning and deep learning. In short you've got computing power like never before at your disposal to implement something as complex as deep learning.

Fully connected deep learning network (Courtesy: sciencedirect.com)

What is most interesting about deep learning is rather than using one neuron to solve your problem you have multiple neurons across different layers within the network. It’s like rather than single intelligence you have many intelligence. It’s really fascinating to deep dive into every layer and observe how the input vector is getting transformed and all the more fascinating to imagine how frameworks like Tensorflow abstracts the complexity of forward and backward propagation. The deep learning network can be really complex and the one thing that I have observed about deep learning is no one really understands what actually happens at every layer. For image recognition, I’ve heard experts say at one layer the network learns about the edges in the images, next layer it learns about objects within the images and so on. If that be the case what happens when I use a 3-layer network vs 10-layer network for the same set of images? Perhaps one can justify the above hypothesis for 3-layer network but what about 10? How do we debug if we ever need to? What if out of every 1000 driverless car, two jumps out of the cliff. Can we arrive at a conclusion why did the two cars behave differently? This is something that makes me uncomfortable - the danger of not having complete control over what we are building.

Please feel free to post your comments, would like to know your views. More to come in 2018...

Disclaimer: Please note that all the views expressed in my post is solely my own, it neither reflects my current or previous employer's views or opinions.