What is Deep Learning?

Data science is revolutionizing many fields; from robotics to medicine, and everything in between. This revolution is partly due to advances in research, computing power, interests within the field, and the data science toolbox. Often, persons think of data science as extreme advances within artificial intelligence (AI); as in, eventually giving robots the ability to complete human-dominated tasks all on their own.

As much as this could be an aspect of data science, it is not all there is to data science. Rather, AI is part of the data science toolbox. Areas such as machine learning (ML) and AI have grown to become popular aspects of data science because they are incredibly powerful tools.

These tools are powerful because they learn and adapt to optimize the outcome of a situation which they are tasked with. This is important because although humans can learn and adapt to optimize outcomes, machines currently have the upper hand at completing this on a larger scale.

Many problems are quite complexed, and it would be unreasonable to ask a human to work through one such problem. Rather, humans should leverage their knowledge of a situation and combine this with both computing power and data to achieve substantial results. At the intersection of this is Deep Learning!

So, what exactly is deep learning? Broadly speaking, it is an application of AI. Since deep learning is a subset of AI, we must first understand AI and what it seeks to achieve. AI is any technique which enables a computer to mimic human behaviour. As the name suggests, it is a branch of computer science which emphasizes the development of intelligence within machines. In this case, intelligence can be considered to be the ability to process information which can be used to inform future decisions.

Therefore, the goal of AI is to develop efficient algorithms which can process information that can inform future decisions. ML is often used to achieve this goal. ML is a subset of AI and it provides systems with the ability to automatically learn and improve from experience without being explicitly programmed.

Now ‘learning’ has been mentioned a lot. However, machines are not reading books, conducting research nor asking questions to learn as humans do. Rather, ML algorithms use computational methods to understand information directly from data without relying on a predetermined equation as a model.

To do this, the algorithms are made to determine a pattern in data and develop a target function which best maps an input variable, x, to a target variable, y. Deep learning is a subset which takes this idea further. The goal of deep learning is to automatically extract the most useful pieces of information needed to inform future decisions.

The idea of taking ML a step further can seem a bit abstract and thus, blurring the difference between ML and deep learning a bit. The idea is that typical ML algorithms attempt to define a set of rules within data, and these rules are usually hand-engineered. Because of this, ML algorithms can also be less ideal than expected when placed in outside of a development environment.

Consider ‘ThisIsPizza’, which is my fictional app to detect whether the object in a picture is a pizza slice or not. Now, having an app and an algorithm which can accurately determine whether an object is pizza is important because one would not want to eat a triangular object which looks like pizza but is not pizza. Pizza is very complexed, and recall that machines are better at dealing with complexed situations than humans are. Therefore, the classification rule could then be if there is a triangular object with at least tomato sauce, cheese, pepperoni, and a crust at the base of the triangle, then it is a pizza. But then, the obvious question would be, how to determine whether something is tomato sauce, cheese, and pepperoni?

The idea of deep learning is that these features will be learnt just from raw data. There would be no need to define pepperoni as a red-ish circular image. And could you imagine defining cheese or tomato sauce? Rather, the deep learning model will develop a hierarchical representation of lines, curvature, and other features which can be used to distinguish cheese from tomato sauce, and then combine these features to then detect higher-level features such as pizza slices.

Although ThisIsPizza is being used as a toy example, the concepts explained are utilized in everyday applications to overcome multiple ‘real-world’ challenges. Understanding the challenges faces by ThisIsPizza, along with how deep learning can be used to overcome these challenges, is essential to understanding what exactly deep learning is.

To use a more concrete example, let us consider Apple’s Face ID. This is a facial recognition system that can be used to unlock Apple devices, as well as securely act as an authentication system when using services such as Apple Pay. Without going into too much detail, the Face ID system uses powerful cameras to detect and map a user’s face. But how does the software know that what it sees is actually a face? Said differently, it would be a major problem if a user accidentally scanned their leg and payment then goes through.

To determine whether something is a face, the algorithm might try to recognize a mouth, eyes, and nose. Once these are present, then the algorithm might classify the image as a face. But again, the question comes up about how to distinguish a mouth, eyes, and nose. So, to then distinguish these features further, we might say that a mouth is a pair of lines with a particular orientation and that these lines should not be situated above the nose. These rules can continuously become more complexed, and they would need to be created for every sub-item of interest.

Therefore, the key idea of deep learning is that these features need to be learned just from raw data. The algorithm would learn this by being fed thousands of images of faces, and it will then develop a hierarchical method of recognizing faces. First, it might try to detect low-level features like lines, edges, and corners. Next, it can use these to detect mid-level features such as mouths, eyes, and noses. Then, composing these together to detect high-level features such as facial hair or dimples.

So, what exactly is deep learning? It is a powerful process which enables a computer to mimic human behaviour by automatically extracting the most useful pieces of information needed to inform future decisions. Data science is revolutionizing many fields and this is partly due to the advances in incredible tools such as AI and computing power. Real-world problems are quite complexed, so let’s try to solve them using deep learning!

References:

machinelearningmastery.com/what-is-deep-learning/

neuralnetworksanddeeplearning.com/

mathworks.com/discovery/deep-learning.html

Other Useful Material:

towardsdatascience.com/what-is-deep-learning-and-how-does-it-work-2ce44bb692ac

www.youtube.com/watch?v=6M5VXKLf4D4

deeplearning.mit.edu/

https://www.inertia7.com/tristn