Artificial Intelligence: Contemporary deep-learning and its promises

Time and again we've heard about the breakthroughs in the field of AI. All of us have come across technical jargon such as 'neural networks' and 'machine learning', but the actual real world applications that we have right now are thanks to a sub-field of machine learning: deep learning.

These two terms may sound related and as a matter of fact, they are: since they are a part of AI, both of them involve sorting data based on a specific set of attributes that separates different unique types of data. Hence using pre-acquired data we can train such systems to, based on the approach taken, sort through a sea of information to make predictions and choices.

The most prominent difference(based on the approach taken) between the two is that while machine learning generates outcomes based on how a model (a system based on any number of algorithmic approaches for classifying inputs to make a decision or prediction) is designed with algorithms to sort through different sorts of predefined characterizing feature of a given input, deep learning is designed such that the algorithms can help them to train the software so that it can classify based on things not noticeable to the human eye.

So for the most part deep learning is self-guided with regards to data analysis once put into action.

Data is passed through layers of processes for the purpose of interpretation and it may be anything as far as the type is concerned; ML is limited in that sense. So, this fundamentally means that we do not require any explicit programming in deep learning.

Semantics aside, deep learning requires a mind-boggling amount of information. This implies that only tech giants such as the likes of Google and Amazon, who can generate lots of data sheerly based on their users, can afford to use these techniques. Also, the recently emergent startups or anyone interested can't use deep learning because of the storage and computational resources they're required.

The pace at which new and more precise models are approaching is alarming. Also as I discussed in a previous article, debugging in such scenarios is nearly impossible; meaning that it's effectively a 'black box'. One of the factors that deter newcomers' is that there's no concrete guideline to follow.

That said, the advantages by far outweigh the disadvantages. Practically, deep-learning can be designed so that it can be molded to accommodate any variety of problems rather smoothly. It has the highest precision out of all such techniques. Medical companies may apply big data to surpass the bounds of our current knowledge. Companies, such as the San Diego based "12 Sigma technologies" are already dabbling in deep learning for oncology. There's also a lot of potential for drug discovery by modeling and simulating molecular interactions.

And of course, we're all familiar with deep learning used in social media to predict friends, videos, and posts etc.

There is extensive use of deep-learning for sentiment analysis, using general keywords and opinion of the masses to make predictions about markets(mainly Forex and stock) and future events.

Future has great potential for the advancement of humanity through the likes of deep-learning and ML, and there's no shortage of frameworks for you start with. With increasing computational power and the rise of promising new types of architectures (quantum optical and cryo to name a few), It's merely a matter of time before deep-learning based approaches are deeply(pun intended) ingrained into our true and tested arsenal of technology.







Muhammad Ali Irfan Khan

This post has been seen 20021 times.

More from Aly Sher Khan

Travel Diaries: Gulmit

Nestled between a narrow valley, Gulmit is a small village, located roughly...
Read More