Rohit Ayyagari is a software professional and an independent researcher with passion towards Machine Learning, Deep Learning and Artificial Intelligence.
His article, Dargan, S., Kumar, M., Ayyagari, M. R., & Kumar, G. (2019). A survey of deep learning and its applications: a new paradigm to machine learning. Archives of Computational Methods in Engineering, 1-22. provides insights into how Deep Learning can be used to save cost and time in several industries like Healthcare, Manufacturing, Big Data etc.
Machine Learning and Artificial Intelligence were at the forefront of technological breakthroughs throughout this past decade. The exponential development of computational speed these past few years allowed machine learning to be more widely accessible than ever before, allowing data science and by extension artificial intelligence to develop at an incredible speed.
Machine Learning’s More Computationally Expensive Cousin
One of the major advantages of this technological improvement is the increased adoption of Deep Learning technologies in industries and workplaces. Deep learning requires gigantic amounts of data and thus is orders of magnitude more computationally expensive than machine learning. The recent huge computational power increase allowed deep learning to be more widely implemented. To understand the meaning of deep learning, we first need to understand machine learning.
Machine learning describes the use of statistical models to predict an outcome or solve a problem based on the data inputted to the model. What does this mean? It means that machine learning models use statistics and mathematics to predict an event, either an outcome or a solution to a particular problem based on a set of variables that the model dictates.
To simplify this further, let us say we would like to predict the salary of a new employee, we have data that contains the salary, skills, years of experience, industry and other similar data that belongs to 5000 employees. The machine learning model would “train” on this data and figure out the patterns that leads to an employee having a particular salary, it would then construct an equation based on this data.
For example, this equation could be that for an employee, his or her monthly salary would be equal to 1000 dollars plus the years of experience multiplied by 600 dollars, this means that an employee with 10 years of experience would have a salary of 1000 + 10 x 600 = 7000 dollars. This process is of course a drastic simplification of the actual process, but it gives you an idea of how machine learning models work. A model would use given data (the data of previous employees) to predict a particular outcome (the salary of a new employee).
This process can be adapted using different statistical methods and equations to suit a particular need or purpose.
What about deep learning? And how is it different from machine learning?
“Deep learning is the most effective, supervised, time and cost efficient machine learning approach.”
According to A Survey of Deep Learning and Its Applications: A New Paradigm to Machine Learning, one of the more widely cited research papers in the topic of deep learning applications, found here.
Compared to machine learning, deep learning is a lot more flexible and can be applied in a lot more fields. It also had several technological breakthroughs in the last 6 years, recently making it the better option over machine learning in many fields. To quote the paper, “The widely used domains of deep learning are business, science and government which further includes adaptive testing, biological image classification, computer vision, cancer detection, natural language processing, object detection, face recognition, handwriting recognition, speech recognition, stock market analysis, smart city and many more.”
We will go into some of its more widely used applications later but for now, how does deep learning differ in the way it works compared to machine learning?
How Does Deep Learning Work?
Deep learning uses artificial neural networks, which are modeled after the neurons in the human brain, this makes neural networks able to “learn” more like a human would learn. These neural networks are the major components of deep learning models, which is why deep learning models are often referred to as neural networks. The “deep” in deep learning refers to the fact that these neural networks are stacked in multiple layers, like neurons in a brain are stacked, which makes these deep learning models “deep”.
According to the paper, deep learning models function in two phases, the training phase in which each “layer” of the model takes data as input, figures out the patterns in the data, and gives out an output, which the subsequent layer takes as input, figures out more patterns on a deeper level, and gives out an output, and so on, until the data goes through the whole network. The second phase is the inferring phase in which the model makes conclusions based on the data it saw and assigns a label, or result, to the problem it is trying to solve.
Deep learning is more suited to complex problems than machine learning is due to its nature of being able to actually “learn” from the data. This nature also makes it easier to understand patterns on a deeper and more abstract way.
What kinds of complex problems can deep learning solve?
Deep Learning Applications
One of the more crucial uses of deep learning is in healthcare, in which deep learning models are used to identify and diagnose various kinds of tumors, cancers, and diseases. These models are sometimes as accurate as experienced doctors while being faster and, in some cases, more accurate as well. A model can be as fast as several hundred identifications a second, which is extremely fast. It also allowed the development of telemedicine, which can provide rural and less developed areas with medicinal assistance without having a doctor go there physically.
Deep learning models can be used to make cars autonomous, with data from multiple cameras that capture the surrounding environments and then input it into a deep learning model with various distinct models like object detection and recognition to Identify pedestrians and other cars, and so on. This combination of models could allow a car to be completely autonomous. We can expect to witness this technology in the upcoming years.
Natural Language Processing (NLP)
The understanding of natural human language is among the most complex tasks a machine can get used to. Context, semantics, tone, expressions, sarcasm, double meanings, and other vague parts of language can make it extremely hard for a model to predict how the conversation flows and what a person could mean in a sentence. Deep learning made unforeseen strides in NLP due to the huge increase in computational power as well as more advanced techniques. This allowed various fields to thrive, like text and document summarization, question answering, text classification, sentiment analysis, speech recognition, word spotting, and writer identification.
Deep learning is expected to continue to achieve greater and greater strides due to the exponential improvement in computing power, which will make numerous fields develop and in turn become faster and more efficient. It is exciting to see how these emerging fields will impact human society and consequently, the whole world.