Disclaimer: The contents of this post, including the images, are from the lectures of “A First Course in Deep Learning” by One Fourth Labs. I am solely responsible for any mistakes that might have crept in while reproducing it.
As the name suggests, the purpose of this post is to burst the jargon bubble — “Deep Learning”, along with some of the related AI bubbles floating around in the industry as well as the academia. Probably when you heard about Deep Learning, you might have come across similar jargons like Data Science, AI, Machine Learning, Pattern Recognition etc.
Have you wondered: Are they all the same? Or related? Why are they known by different names? This post is for busting some of those bubbles and give a clearer picture so that you might be able to choose what is right for you.
Trending AI Articles:
1. Predicting buying behavior using Machine Learning
2. Understanding and building Generative Adversarial Networks(GANs)
3. Building a Django POST face-detection API using OpenCV and Haar Cascades
4. Learning from mistakes with Hindsight Experience Replay
Let us begin with the common confusion we all have.
Is AI the same as ML? How are they related to Data Science? Isn’t Pattern Recognition same as AI? What about PR in Computer Vision? Isn’t it AI? Where does Deep Learning fit in?
Can these definitions help us? Take a look!
Artificial intelligence (AI), sometimes called machine intelligence, is intelligence demonstrated by machines, in contrast to the natural intelligence displayed by humans and other animals.
Machine learning is a field of computer science that uses statistical techniques to give computer systems the ability to “learn” with data, and without being explicitly programmed.
Does AI not involve statistical techniques? Does AI not involve data? Is it AI only if it is explicitly programmed?
Data science is an interdisciplinary field that uses scientific methods, processes, algorithms and systems to extract knowledge and insights from data in various forms, both structured and unstructured, similar to data mining.
Is AI/ML not interdisciplinary? Does AI/ML not have scientific methods? Does AI/ML not deal with structured data? Is knowledge not important in AI/ML?
Pattern recognition is the automated recognition of patterns and regularities in data.
Doesn’t “recognition of patterns” imply intelligence? Doesn’t “recognition of patterns” allows you to extract insights from data? Doesn’t “recognition of patterns” implies learning from data?
Computer vision is an interdisciplinary field that deals with how computers can be made to gain a high-level understanding from digital images or videos and automate tasks that the human visual system can do.
Isn’t AI/DS not interdisciplinary? Of course, they are! Doesn’t high-level understanding show intelligence? Image and videos are data. So isn’t it data science? To automate tasks that human do implies AI, isn’t it?
Deep Learning is a new area of Machine Learning research, which has been introduced with the objective of moving Machine Learning closer to one of its original goals: Artificial Intelligence.
If these definitions still leave you confused, continue reading.
We know that humans are intelligent beings. So let’s try an alternative approach to resolve this confusion by thinking in terms of what humans can do, and can the machines do the same thing. We shall use this AI map as a lens to look through and differentiate these jargon bubbles.
The AI Map
The AI map is a framework that shows the abilities, tasks and methods that we come across in an AI system. To begin with, let us map all our AI jargons into one of these layers. Here is one shown for you.
Some of the abilities of human that make them intelligent are — the ability to see and comprehend, to speak and listen, to handle language and to plan/make a decision. When a machine gets these abilities we call it as Computer Vision, Speech, Natural Language Processing, Planning/Decision Making respectively (shown in the blue layer).
Having these abilities allows us to do certain tasks. For example, using our ability to see and comprehend, we can classify images, recognize digits, identify objects etc. So, to say that an AI system has computer vision, ideally, the system should have these capabilities like image classification, character recognition, object identification just to name a few. Similarly, speech ability allows machines to do speech synthesis, speech recognition, distinguish speaker etc. A system with NLP capability will be able to do document classification, spam detection, translation etc. Planning/Decision-Making ability helps in Autonomous driving, game playing, robotics etc. These are shown in the task (red) layer. Not all AI-driven systems have all these abilities or all the tasks listed within an ability. Such systems are known to have narrow AI, in contrast to the general intelligence a human-being can have. We are striving to build systems that can have general AI.
The methods/algorithms that the machine uses to achieve the above-mentioned tasks are shown in the green layer. One of the earliest method used by AI systems were expert systems, with every rule for decision making explicitly coded into the machine. Then came the machine learning techniques where the machines were trained to learn these rules instead of being explicitly coded (hence the term in its definition). The latest in this group is the Deep Learning models, which are complex models that can learn complex relation provided sufficient data and properly engineered.
So coming back to the jargon bubble, we are now able to fit a few of these into one of these layers.
The ones shown in green are the techniques used in AI, the red ones are the tasks that can be achieved using AI techniques and the blue ones are the abilities of an AI system. An AI system encompasses all these abilities, tasks and methods. So next time you see an AI jargon, try to fit it into this map and get it bursted.
Explicitly Programmed v/s Learning Based Systems
As pointed out earlier in the definition, we know Machine Learning systems are not explicitly programmed. Unlike expert systems where every possible rule is explicitly coded, Machine learning systems are provided with the input and output data, and a “suspected” function between them. The goal of ML algorithm is to learn the parameters of these function. That implies, the performance of our system depends on the training data as well as the function chosen. For example, a linear function might be a bad choice for data that aren’t linearly related. In that case, no amount of training or learning algorithm would help you in improving the “intelligence” of your system.
How is Pattern Recognition different?
Most AI systems require pattern recognition. For example, detecting an aeroplane in the image, identifying a spam email, recognizing the author of a document, all these are pattern recognition. To achieve this we might be using different methods (shown in the green layer). So in general, we can say, PR is an AI task that may span over different abilities and are solved by different methods.
Is Image Processing different from CV?
Yes, Image processing by a rule of thumb takes an image as input and produces an output image. This is not the case with computer vision where output would be an object category or marked portion of the image or maybe a gesture identified. Most CV system uses image processing for better results. Image processing uses the same or a subset of machine learning techniques that a CV uses.
What is Data Science?
Is it a subset of AI, a superset or neither?
When you are dealing with data and performing some systematic study through observation and experiment, then you can call your activity as Data Science. In that sense, many tasks in NLP, speech processing, Image processing etc could be termed as Data Science, because text, audio signals and images are also data, and you might be doing some scientific processing on it. But it is preferred to use a more specific term like NLP rather than Data Science or AI which may lead to more confusion. Data Science is commonly used when dealing with structured numeric data like database tables.
The DL landscape
We have seen that Deep learning is one of the methods used in the AI system. What makes it more attractive than the others is, they can be used for multiple abilities like vision, speech, language, decision/making etc. That means we could step closer to general intelligence, unlike the shallow AI. The downside of using DL is, it is highly data intensive. You need to have lots and lots of supervised data to get it right. That means you would need high storage and computational power (in terms of GPUs) to process these data. DL models are notoriously hard to interpret. I would compare it to the human instinct that tells you “this is the answer” but not being able to explain “why this is the answer.” DL systems are prone to adversarial attack. You could easily mislead your machine to classify a panda as a gibbon by adding a few random noises that do not cause any visible change in the original image. People are making progress in these matters though.
Why is this the right time to start learning “deep learning”? To quote Dr Mithesh Khapra’s words, “Deep Learning is one single technology that is the most prominent technology in machine learning right now, which is being used for a wide variety of task. Since you have a large amount of data coming in, and this being a technology that works well with data — this is the right time for Deep Learning”. I would like to add, this is one technology that can take you “far and deep” into the AI world.
To dig deeper, enrol to the “Deep Learning course” on One Fourth Labs. See you all there!