Back to Home
The idea of Artificial Intelligence (AI) definitely came from some Lazy People. Kidding! But we can not deny that AI is making our lives easier. If we look around, we will see many intelligent applications to automate routine labor, understand speech or images, and make diagnoses.
At this time, you may hear some terms Machine Learning (ML), Deep Learning (DL), etc with AI. Even many people interchangeably use these words. But the thing is they are not the same. They all have their own fields with specific boundaries. AI is the container of ML, DL, and other similar terms. But one thing is for sure that they all are necessary for building up intelligent machines, and robots. Anyway, here, our focus will be only on the Deep Learning part.
Through this article, we will try to find the answer to the question “Do we actually need Deep Learning?”
☑️ The Plot of AI
☑️ Why do we need Deep Learning?
☑️ How Deep Learning Works in General
Note: “Machine”, “Computer” words are used interchangeably through the post
AI has been a hot topic of research since the 1950s. But during the early days of AI, it’s not that common. You know that the beginning of anything is always difficult. AI started its journey when people wanted to let machines think. You can find the history of AI in detail here.
Many of the early successes of AI took place in relatively formal environments. Computers or machines were not required to have much knowledge about the world. Generally, the machine was provided with some set of instructions or logical inference rules to perform any intellectual task. A machine can reason statements or classify anything using logical inference rules. This is actually known as Knowledge-Based or Rule-Based AI.
In Rule-Based or Knowledge-Based AI, you can classify records by using a collection of “if…then…” rules.
(Condition) → Y
Here,
➵ Condition is the conjunction of if-else rules
➵ Y is the class label
For example,
(Has Wings = Yes && Lay Eggs = Yes) → Birds
(Has Wings = Yes && Lay Eggs = Yes) this part is the condition and “Birds” is the class label. So if a thing has wings and can lay eggs, then it means the thing is a bird. So simple, right?
Still, this kind of Rule-Based AI research on populating rules is popular. If you look at this research paper which was published in 2019, you will see that at the end we extract similar rules (given below in Fig. 2) by a tool called Ant-Miner.
[ CKD is the short form of Chronic Kidney Disease. ]
CKD and NOTCKD are two classes which means the patient might have CKD disease and the patient might not have CKD disease respectively.
On the left side of the rule, there are some conditions checking for medical-related factors likehemoglobin, appetite, etc. And on the right side, we construct the class.
We can feed these rules to a machine or computer by some AI or ML algorithms, and train the machine with those rules. Later, for a new set of data, the machine can find out whether any patient might have been diagnosed with CKD or not by reasoning the rules it trained on earlier.
So in the end, if the trained machine is given that a patient has hemo = 5.6, then the machine figures out that the patient might have been diagnosed with CKD. Because the machine already knows from the training data (Rule 1) that if hemo<=12.9, it means CKD.
By the above discussion, we try to cover how a traditional AI or ML works. In short, we can say, many artificial intelligence tasks can be solved by designing the right set of features to extract for that task, then providing these features to simply a machine learning algorithm.
People always want machines to behave exactly how a human behaves. That’s not possible by integrating some rules into the machine. Because a person’s everyday life requires a huge amount of knowledge about the world. Computers need to capture this same knowledge in order to behave intelligently. However, for many tasks, it is difficult to know what features should be extracted.
What if we feed an elephant’s raw image to a machine?
Here rules or factors or features like color, the number of legs are not given, it’s just a raw image. The machine must figure out the factors and classify the image by itself. Can the machine classify the image?
We as a human know that an elephant has 4 legs, we might like to use the presence of legs as a feature or factor. But it’s hard to describe exactly what a leg looks like in terms of pixel values. Generally, the leg of the elephant is tall, fat. Also in an image, there may be shadows or mud on the leg, or the full leg might not be visible.
Also, it can be very difficult to extract such high-level, abstract features from raw images or data. For that reason, we basically need some magic box named Deep Learning. It allows the computer to build complex concepts out of simpler concepts. It extracts high-level features like edges, corners, etc from an image by itself (so which means we don’t have to extract the features) and then classifies images.
An example of the Deep Learning model is the feedforward deep network or multilayer perceptron (MLP). MLP is just a set of mathematical functions mapping some set of input values into output.
Each layer can be thought of as the state of the computer’s memory after executing another set of instructions in parallel. Network with more depth will have more sequential executions in parallel.
For an image, the function mapping from pixels to the classification of an object is very complex. Deep learning resolves this complexity by breaking the desired complicated mapping into a series of nested simple mappings.
If you look at Fig. 7, you will see the input layer is called the Visible Layer. And then series of hidden layers. These are called hidden because the data in those layers are not present in the raw image.
Now, you know how Deep Learning works with images in general.
So at this point of time, anyone must agree that we certainly need Deep Learning at least for extracting high-level details from data.