Jul 22, 2021 By Team YoungWonks *
What is Deep Learning? At a time when terms such as Artificial Intelligence (AI) and Machine Learning (ML) have become buzzwords in the tech world and the world at large, this is also a commonly asked question.
Let’s start by taking a quick look at what ML and AI mean. Artificial Intelligence, also known as machine intelligence, is intelligence displayed by machines, in comparison to the natural intelligence that is exhibited by humans and other animals. In the computing world, AI refers to the ability of a computer program or machine to perform computations, i.e. think, learn and make calculated decisions. Such a program or machine is fed large amounts of data which it then analyzes and processes and after this, it thinks logically to perform human actions.
ML, meanwhile, is essentially a branch of AI. It is a field that relies on the use of statistical methods so as to empower computer systems and grow their computing power to a point where they can progressively improve their performance on a specific task, without being explicitly programmed. It mainly involves the study and building of Machine Learning algorithms that can learn from and make data/ sample-driven predictions. Read our blog here if you wish to know more about ML and AI: https://www.youngwonks.com/blog/Machine-Learning-and-Artificial-Intelligence---An-Introduction-For-Absolute-Beginners.
This then brings us to our main question: What is deep learning?
What is Deep Learning?
Deep learning is a machine learning technique (and thus a subset of Machine Learning) teaching computers to do what humans can do naturally: learn by example. Also known as deep neural learning or deep neural network - because most deep learning methods use a neural network architecture - here a computer model or program learns to carry out classification tasks by scanning images, text or sound. It basically emulates the workings of the human brain in terms of processing information and coming up with patterns that can be used for decision making. Such deep learning models have shown their potential to deliver with an accuracy that humans - more often than not - cannot. A large set of labeled data and complex neural network architectures - made up of a number of layers - are used to train these models, after which the machine learning model is able to learn from unstructured or unlabeled data. Object detection, even facial recognition, speech recognition, and language translation are some of the tasks that can be carried out using deep learning.
The term ‘deep’ in deep learning alludes to the number of hidden layers in the neural network. Traditional neural networks are made up of 2-3 hidden layers, whereas deep networks can have as many as 150. And while deep learning can be unsupervised, semi-supervised and supervised, it is supervised learning that is the most common type; this is when the model scans labeled training data consisting of a set of training examples.
Deep Learning and Big Data: How Does Deep Learning Work?
As mentioned earlier, deep learning basically helps make sense of large sets of data. Such large, diverse sets of information that grow at ever-increasing rates are known as big data.
An application of data science, big data covers the volume of information, the velocity or speed at which it is created and collected, and the variety or scope of the data points. Big data is put together thanks to data mining, a process used by companies to turn raw data into useful information (with data scientists doing the statistical analysis to determine which ML approach should be used). Cloud computing is also used for deep learning models since it allows large datasets to be easily hosted on the cloud, and lets these models scale efficiently and at lower costs using GPU processing power.
Bear in mind that big data can be unstructured or structured. Structured data refers to information already managed in a defined structure (think databases and spreadsheets); it is usually numeric in nature. Unstructured data is unorganized information that does not fit a predetermined model or format. It could even be data collated from social media sources, which in turn help companies gather information on their consumer needs.
In case of deep learning, the data is usually unstructured. The data could be sourced from social media, internet search engines, e-commerce platforms, and other areas. But given that this data is unstructured, it would take a lot of time for humans to analyze it and cull out relevant information. This is where deep learning comes in, more and more companies are realizing this now and hence turning to AI systems for automated support.
Deep learning makes use of something called hierarchical neural networks to process / comprehend reams of data. Within these neural networks, neuron codes are linked together. And the hierarchical structure of deep learning enables it to adopt a nonlinear approach, in turn enabling it to scan data across a series of layers which further integrate with subsequent tiers of additional information.
So a deep learning model is actually trained to work with neural network architectures such that it can cull out features from the data in the network without any manual assistance.
Convolutional Neural Networks (CNN or ConvNet) is among the most popular types of deep neural networks today. A CNN entwines learned features with input data, and uses 2D convolutional layers, which in turn makes this network ideal for processing 2D data, such as images. In doing so, it basically helps do away with the need for manual feature extraction in tasks related to image classification. Interestingly, although the network has trained on a collection of images, it is not pre trained to look for certain features. Thanks to its training, it reads and detects different features of an image using several (from tens to hundreds) hidden layers. Each hidden layer adds to the complexity of the learned image features, making the result that much more accurate.
For instance, the first hidden layer learns how to detect edges, and the last one learns how to find more complex shapes or even the exact shape of the object we may be trying to detect. And it is this automated extraction of features through the different layers that makes deep learning models so accurate when it comes to computer vision tasks such as object classification.
Let us look at another example of deep learning. Suppose it is used to detect fraud (another common use). In such a context, the algorithm will be built into the computer model such that it is able to process all transactions taking place on the digital platform, see patterns in the data set, and point out any inconsistencies as per these patterns. As opposed to a traditional approach which would entail looking for the amount of the fraudulent transaction, the deep learning nonlinear technique will refer to not just the amount, but also factors such as the IP address, credit score, time, geographic location, type of retailer, sender details, and any other aspect that can hint at the fraudulent activity. So in the first layer of its artificial neural network, the amount sent will be analyzed. In the second layer, this information will be built upon and the IP address will also be included. In the third, the credit score will be added to the existing data, and so it will continue till a final decision is reached on the basis of a scan across all the levels of this neuron network and it is the final layer that sends a signal to an analyst who can then freeze the user’s account for further investigations.
Difference between Machine Learning and Deep Learning
Deep learning can be described as a specialized form of ML. To understand the differences between the two, let’s look at their role in the object identification in the images example. In the case of ML, relevant features are manually extracted from images and are used to build a model classifying the objects in the image. With deep learning, relevant features are automatically culled out from the images. Moreover, here we see end-to-end learning, in that the network is given raw data and a task to carry out (say, image recognition), and it learns how to do this on its own.
Another key difference between ML and deep learning is that deep learning algorithms scale with data. So deep learning networks typically become better even as the volume of data added goes up. Meanwhile, ML that can’t scale with huge amounts of data can be called shallow learning since these are ML methods whose performance plateaus at a certain level upon the addition of more data to the network.
It is also important to note that while deep learning and reinforcement learning are not mutually exclusive; they are not the same either. While both systems learn autonomously, deep learning is about learning from a training set and then applying that to a new data set, while reinforcement learning is an area of ML that deals with learning dynamically by modifying actions on the basis of continuous feedback so as to maximize a reward.
Applications of Deep Learning Today
Today, deep learning technology is used across industries for various tasks covering predictive analytics, computer vision, object recognition and Natural Language Processing (aka NLP, an AI area that aims to understand and illustrate the cognitive mechanisms contributing to understanding and generating human languages). Let us look at the key areas where it is used:
1. Automated Driving
Deep learning techniques are being used by automotive experts to automatically detect the presence of pedestrians and objects (think stop signs and traffic lights, among others). This is set to reduce the number of accidents for self-driving cars.
2. Aerospace and Defense
Here, deep learning systems are used to recognize objects from satellites that locate areas of interest, and help mark out safe or unsafe zones for troops.
3. Industrial Automation
Worker safety in industries, especially near heavy machinery, is being enhanced using deep learning. Deep learning helps automatically detect people or objects that come within an unsafe distance of machines.
4. Electronics and Virtual Assistants
Deep learning is also being used in automated hearing and speech translation (think NLP). So the smart home appliances (think Alexa on your Amazon Echo device or Siri on your iPhone) that respond to your voice requests, carry out the assigned tasks and come to know your preferences are all thriving thanks to deep learning applications. Typically, a recurrent neural network (RNN) is used here.
Chatbots are used to quickly address simple customer problems. Essentially an AI application that chats online via text or text-to-speech; a chatbot uses machine learning and deep learning algorithms to produce and deliver different types of automated reactions to user inputs. It is commonly used in customer interaction, for marketing on social network sites, and for instant messaging clients.
6. Medical Research and Healthcare
Cancer researchers are carrying out automatic detection of cancer cells using deep learning. A team from UCLA has actually made an advanced microscope that uses deep learning and photonic time stretch to find cancer cells more efficiently.
Deep learning is also being used for making predictions about the future in areas such as medicine, weather forecasting, biology, supply chain management and stock prices forecasting, etc.
*Contributors: Written by Vidya Prabhu; Lead image by: Abhishek Aggarwal