Demystifying Artificial Intelligence, Machine Learning, Deep Learning, and Data Science (AI-ML-DL-DS)
We all have been hearing the buzzwords AI, Machine Learning, Deep Learning sometimes mixed with data science used interchangeably around us. We tag our posts sometimes without knowing the actual differences and similarities between them. I was also in a similar thought process a few weeks back when I started reading about these terminologies. Here is a summary that I have tried to put together using explanatory images for the ease of understanding and remembrance.
Below is a self-explanatory diagram that talks about the relation & difference among the terminologies and the timeline when they started becoming popular.
Understanding them with a common example (self-driving cars)
This term, AI was initially coined by John McCarthy in 1956 when he talked about machines that can do tasks that are meant for human intelligence. AI is broadly classified into two parts. AGI and ANI i.e. Artificial General Intelligence and Artificial Narrow Intelligence. Much of the advancement which has been done until now is mostly in the ANI space. Smart speakers, self-driving cars, AI in farming and factories, etc. AGI has not seen many concrete advancements and it is still an area of research where machines will be able to imitate human wisdom, reasoning, and intelligence. This might be decades away. For the sake of understandability, I will take a running example across all sections to help relate and differentiate the ideas.
Taking self-driving cars as a running example, AI (specifically ANI) deals with overall safe driving of the car thinking of all obstacles and making decisions on what to do when, while also considering the next steps that a nearby vehicle or a human in the surrounding environment may take and how to deal with that.
Coming to machine learning (ML), it is a set of algorithms that parse data, learn from that, and then apply what they’ve learned to make decisions. Arthur Samuel coined the phrase in 1959, defining it as, “the ability to learn without being explicitly programmed.” Machine learning is nothing but realizing AI without specifically the rules into the system. It is a way to teach the machines to learn and identify the rules for a particular task, similar to how humans do while feeding huge amounts of data to the software algorithm and creating and tweaking the algorithm such that it itself adjusts its parameters and improves.
In our running example of self-driving cars, machine learning deals with the computer vision part. It is fed with lots and lots of data (as we have sometimes seen self-driving cars roaming around and collecting data) with manual tagging and telling the machine that this is a car and this is a pedestrian. After some time, the machine knows which a car is and which a pedestrian is. One such technique (sliding window) for how it does so is also discussed in my previous writing here.
Deep learning is one of the streams of machine learning. Other approaches include decision tree learning, clustering, Bayesian networks, etc.
Deep learning was inspired by brain structure where multiple neurons connect with each other to learn something new and transport the message from one place to another while learning on the way. Neural Networks are algorithms that, so-called, mimic the human brain. However, it is not at all how our brain functions since we have not yet identified the functioning of a human brain! In an artificial neural network, the neurons are connected in layers and become deep as we increase the number of layers as shown in the figure above. Each layer tries to learn a specific or a set of features, such as curves/edges in an image. That is why the complex ones are also called a deep neural network. The depth is created by using multiple layers as opposed to a single layer.
The convolutional neural network is used in the above particular example. The circles represent the neurons in a neural network.
Sometimes, a pre-trained network can also be used to quickly re-train the network for a different scenario with just a small data-set. For example, a well-trained neural network to identify cars can be quickly trained to identify golf cars which are very different from a normal car.
In our self-driving cars example, one way to achieve computer vision is using deep neural networks. Multiple layers breakdown the features and characteristics of cars and then learn from it. One such visualization is show below here.
Data science is a wide-ranging field that encompasses the collection, analysis, arrangement, and interpretation of huge data with varied applications. It is a study of squeezing out business insights from a given data set. The data may need to be initially cleaned, labeled, organized, etc. before proceeding to the next steps. For all these steps, one can use ML/DL algorithms to find a pattern in data and classify huge data sets using unsupervised learning or basic deep learning techniques. And this is why it spans across all the segments as shown in the first figure of this post.
Another way to understand the difference is the basic steps involved. ML/DL pertains to collecting the data, training the network, and then deploying it for use to give input and get an output. However, data science pertains to collecting data, analyzing it, and then coming up and suggesting the hypothesis.
From our running example of self-driving cars. A data scientist, after analyzing the data from multiple cities, can find out which city will be best to train the cars initially depending upon the quality of data that a self-driving car may get. A data science person can also find out the answer to whether one should proceed ahead with a particular project/sub-project or not, depending upon the results of testing from different cities.
What they can and cannot do
Due to the recent emergence of high data processing capabilities, AI has created a hype that killer robots will take over the world anytime soon now. This is not going to happen at least for multiple decades or hundreds of years. Let’s try to understand what an AI can and cannot do as of now with some examples.
An algorithm can classify if the ‘customer care’ call is for complaints or returns departments but it can never empathize with the customer with the customer’s problem (except some pre-specified lines or sentences fed to it). A self-driving car can understand if the ‘thing’ in front of it is a car or not. It can obviously reduce the number of accidents due to human error to a large extent. However, as of now, it cannot understand the intention of hand-signaling pedestrian, whether a construction worker is asking to stop, a hitchhiker is asking for a lift or a bicyclist is giving a turn signal (see below). Since it is a safety-critical application, the car has to be very accurate and precise in terms of what it understands.
Moreover, AI needs hundreds and thousands of images of toys to understand that a newly given object is a toy. However, a human child can quickly identify even a new kind of toy that is presented to him/her the first time. Professor Andrew in one of his videos talks about coffee mugs. A human child need not be said a thousand times that ‘this is a coffee mug’ before recognizing a coffee mug. However, a machine needs them. These facts also signify that we have not yet understood how the human brain functions. To imitate it is a very distant future story (even if there is a minuscule chance).
I hope you would have got a basic idea of the difference between the famous buzzwords in this arena and the potential of these technologies as of now. The AGI (Artificial General Intelligence) is still an underdeveloped field and almost more than 98% of AI on which companies are creating products and benefitting economically is supervised machine learning and that too mostly using deep learning. A fair understanding of ML /DL should be good enough to estimate the possibilities of AI that currently exist.