This article is part 3 of 3 in the series MachineLearning

What is Deep Learning?

Data that has multiple levels of abstraction can be represented with the help of computational modeling, having multiple processing layers, and computational power, to achieve problem-solving objectives. This process is termed as Deep Learning, and it is redefining the insight generation out of abstract data e.g. speech, image, etc.

Advances in Computational Power

The first and foremost thing to understand is how the computational power in getting a boost every passing moment. From the era of mainframes to the incoming era of Quantum Computing, understanding of complex data is becoming a somewhat easy task (primarily due to the decrease in processing time and testing of the multitude of models). To test the ideas and to solve some very difficult problems, we definitely need an immense amount of computational power and resources. The first takeaway is that some of the Deep Learning methods require a huge amount of Computational Power. How huge is the question we need to ask?

My Teraflops per second are not going to flop me

As said earlier that Computing processing capabilities are now getting mature in terms of trillions of instruction processing per second, we see that if they are used in the right way, they will never disappoint us and our goals. We all have our goals in our life sciences projects, engineering projects, finance-related projects, management projects, bioinformatics, chemistry project etc. The technologies in computing world are converging in such a way that all of these fields can make use of it. And the method to do it Deep Learning. And Deep Learning has no limits. It’s high time to learn computer programming as well.

Computing Continuum

Smartphone, embedded systems, mainframes, cloud computing, quantum computing, servers, data centers, High-Performance Computing (HPC), and the like, all are being used to address the wide variety of problems in multiple fields and with diverse business models.

Hardware and network capabilities have evolved tremendously but this cannot be said about software development. Some highly lucrative mathematically algorithms have been around for quite some time but it was really hard to make use of them other than the theoretical sense. Due to deep learning, now we have a great coupling. This should increase not only the understanding of the world but also the velocity in the development of software itself.

Scientific Inquiry

Scientific Inquiry as seen from humans perspective is the method of understanding ideas. Through scientific inquiry, students develop the understanding of the complex ideas that are presented to them. From the machine’s perspective, things get really interesting. The computer has to do the same scientific inquiry, but the beauty of it is that computer has all the time in the world to do this job, and it is getting powerful and fast day by day. It seems like we are at the stage now where we can make use of it to understand at least two important facets of life i.e. image and speech. We won’t limit us to these two, so we will come to this topic again.

Computational Modeling & Simulation

Computational Modeling & Simulation plays an important role in doing scientific inquiry. This is another aid to us humans. Modeling & Simulation has been around for ages but now it is possible to make use of it at scale. Many unsolvable research problems which are really hard to solve in traditional theoretical and experimental approaches can be solved in the realm of computer-based modeling and simulation. The computer is now able to model many processes e.g. in basic energy sciences, chemistry, biology, nuclear physics, environmental research, fusion energy sciences etc.

We all have been mesmerized by the implemented petascale computing, but within next few years, it would be possible to do Terascale computing (or perhaps if you are reading this after 5 years then do let us know in the comments that you have passed the Terascale era as well). The continuous evolution of computational software, the mathematical ingenuity of new algorithms, and the computing capabilities are the precursors of the deep learning.

Machine Learning & the delicate Link to Deep Learning

Machine Learning is a precursor to Deep Learning. We can say that Deep Learning is one of the applications of Machine Learning, which can be used to solve bigger problems than what machine learning was first envisioned do (at-least in implementation).  Deep Learning is further being used as a feedback tool to further mature the Machine Learning itself. It’s the “smarter factor”, which is now in the spotlight.

Machine learning is being used in web search technology, recommendation engines, content filtration, pattern synthesis, text transcription, speech transcription, matching problems, object identification etc. All these applications make use of the technique called Deep Learning. So it is a feedback factor as well.

So if you want to do Machine Learning the right way, you need to know Deep Learning. There are a lot of online resources which can help you kick start in this area. Modern Computing itself is getting diverse that developers have to catch up or evolve with new capabilities.

Supervised Learning

Imagine you want to make your machine tell you about the object which is in front of its camera or screen view. One way to do it is to make your machine learn about the shapes and forms of the similar objects. This will become the training data for the machine. Once machine sees the patterns, it will then start to develop the notion of that thing or object. So if you present the same object with a different look, which the machine never saw in the past, it would be able to tell you that this object is the instance of one of the objects that it has seen earlier. This is called supervised learning.

To do this the machine can use the wide variety of algorithms which are specifically made to do the supervised learning. We will come to those algorithms in a technical way in one of the parts in this series.

Unsupervised Learning

I think you somewhat now know what it means (we as humans are much smarter than machines perhaps till now). The machine, when they are doing supervised learning, can come up with other layers of hidden knowledge which can become a feedback to learn more about dissimilar objects. And it is this notion which leads to the unsupervised learning. So machines get power to do feature detection of objects which are different than the ones for which it was getting trained. This becomes an important part of the cognitive abilities which can be imbued in the machines. Hence it gives a pathway to creating a “mind” in machines.

Conclusion

These concepts, techniques, and methodologies present highly attractive ways to solve many critical problems. Since the advent of computing, it is now much cheaper to solve more complex problems with these. Machine intelligence will get more mature as the time passes, but the real catalyst to shorten each time window the enhancements in the power of computing.

 

 

 

 

 

 

Series Navigation<< Part 2 – Blazingly Hot Applications of Machine Learning