What is the difference between deep learning and machine learning? Do you know?

If you often want to make yourself aware of the difference between machine learning and deep learning, read this article and I will introduce you to the differences in plain language.

Machine learning and deep learning are getting more and more popular. Suddenly, whether you know or don't understand, everyone is talking about machine learning and deep learning. Whether you have taken the initiative to focus on data science, you should have heard of these two terms.

To show how hot they are, I searched for these keywords on Google trend:

If you want to make yourself aware of the difference between machine learning and deep learning, read this article and I will introduce you to the differences in plain language. The terms in machine learning and deep learning are explained in detail below. And, I compared the difference between the two, not to mention their respective usage scenarios.

What is machine learning and deep learning?

Let's start with the basics: What is machine learning? And what is deep learning? If you already know about this, feel free to skip this section.

What is machine learning?

In a nutshell, the definition of the widely cited machine learning given by Tom Mitchell gives the best explanation. Here is the content:

“A computer program can learn from experience E given a certain class of task T and performance metric P, and if its performance in task T happens to be measured by P, it increases with experience E.”

Is it very readable? Let's break down this description with a simple example.

Example 1: Machine learning and estimating weight based on person's height

Suppose you want to create a system that estimates body weight based on a person's height (perhaps you are interested in this matter for some reason). Then you can use machine learning to find out any possible errors and errors in data capture. First you need to collect some data, let's see what your data looks like:

Each point in the graph corresponds to a single data, we can draw a simple diagonal line to predict the weight based on height

For example this slash:

Weight (in kg) = Height (in cm) - 100

...these slashes help us make predictions. Although these slashes perform very well, we need to understand how it behaves. We want to reduce the error between prediction and reality. This is also a measure of its performance. .

Far more far, we collect more data and the model will get better. We can also improve our model by adding more variables (such as gender) and adding different predictive slashes.

Example 2: Hurricane Prediction System

Let's find a more complicated example. Suppose you want to build a hurricane prediction system. Suppose you have all the data about the hurricane that happened before and the weather information for the first three months of the hurricane.

What should we do if we want to manually build a hurricane forecasting system?

First of all, our task is to clean all the data and find the pattern in the data to find the conditions that produce the hurricane.

We can input model condition data (such as temperature above 40 degrees, humidity 80-100, etc.) into our system to generate output; or let our system itself generate appropriate output through these condition data.

We can import all previous data into the system to predict if there will be a hurricane in the future. The performance of the system is evaluated based on the values ​​of our system conditions (the system correctly predicts the number of hurricanes). We can continue to iterate the above steps multiple times by using the system predictions as feedback.

Let us define our prediction system based on the previous explanation: our task is to determine the meteorological conditions that may produce a hurricane. Performance P is how many times the hurricane is correctly predicted under all given conditions of the system. Experience E is the number of iterations of our system.

What is deep learning?

The concept of deep learning is not new. It has been around for several years. But with all the existing hype, deep learning is getting more and more attention. As we did in machine learning, let's take a look at the official definition of deep learning and then explain it with an example.

“Deep learning is a special machine learning that uses the nested concept hierarchy to represent and implement great functionality and flexibility. Each concept is defined as associated with a simple concept, and more abstract. The representation is calculated in a less abstract way."

This is also a bit confusing. Let's break down this concept with a simple example.

Example 1: Shape Detection

Start with a simple example and explain what is happening at a conceptual level. Let's try to see how to identify squares from other shapes.

The first thing in our eyes is to check if there are four lines in the picture (simple concept). If we find such four lines, we further check that they are connected, closed, and perpendicular to each other, and that they are equal (nested conceptual hierarchy).

So, we completed a complex task (identifying a square) and doing it with a simple, less abstract task. Deep learning essentially performs similar logic on a large scale.

Example 2: Cat vs. Dog

Let's take an example of animal identification, where our system must identify whether the animal in a given image is a cat or a dog. Read this article to see how deep learning is one step ahead of machine learning in solving such problems.

Comparison of machine learning and deep learning

Now that you have an understanding of machine learning and deep learning, we will learn some of the key points and compare the two techniques.

Data dependency

The main difference between deep learning and traditional machine learning is that as data size increases, so does its performance. When the data is small, the performance of the deep learning algorithm is not good. This is because deep learning algorithms require a lot of data to understand it perfectly. On the other hand, in this case, the traditional machine learning algorithm uses the established rules and the performance will be better. The figure below summarizes this fact.

Hardware dependency

The deep learning algorithm requires a lot of matrix operations, and the GPU is mainly used to efficiently optimize the matrix operations, so the GPU is the necessary hardware for deep learning to work properly. Compared to traditional machine learning algorithms, deep learning relies more on high-end machines with GPUs.

Feature processing

Feature processing is the process of putting domain knowledge into the feature extractor to reduce the complexity of the data and generate a better model for the learning algorithm to work. The feature processing process is time consuming and requires expertise.

In machine learning, the characteristics of most applications require expert identification and then coding as a data type.

Features can make pixel values, shapes, textures, positions, and orientations. The performance of most machine learning algorithms depends on the accuracy of the extracted features.

Deep learning attempts to obtain high-level features directly from the data, which is the main difference between deep learning and traditional machine learning algorithms. Based on this, deep learning cuts the work of designing the feature extractor for each problem. For example, convolutional neural networks attempt to learn low-level features (boundaries, lines) at the front layer, then learn part of the face, and then describe the advanced face. For more information, read the interesting applications of neural network machines in deep learning.

Problem solving

When applying traditional machine learning algorithms to solve problems, traditional machine learning usually breaks down the problem into multiple sub-problems and solves the sub-problems one by one. Finally, the results of all sub-problems are combined to obtain the final result. Instead, deep learning advocates direct end-to-end problem solving.

for example:

Suppose there is a multi-object detection task that requires the type of object in the image and the position of each object in the image.

Traditional machines learn to break down problems into two steps: object detection and object recognition. First, use a bounding box detection algorithm to scan the entire image to find the area of ​​the object; then use the object recognition algorithm (such as SVM combined with HOG) to identify the object detected in the previous step.

Instead, deep learning directly computes the input data to produce an output. For example, you can pass the picture directly to the YOLO network (a deep learning algorithm), and the YOLO network will give the object and name in the picture.

execution time

In general, it takes a long time to train a deep learning algorithm. This is because there are many parameters in the deep learning algorithm, so the training algorithm takes longer. The most advanced deep learning algorithm, ResNet, takes two weeks to complete training, while machine learning training takes relatively little time and takes only a few seconds to a few hours.

But the time of the two tests is exactly the opposite. The deep learning algorithm requires very little time to run during testing. If compared to k-nearest neighbors (a machine learning algorithm), the test time increases as the amount of data increases. However, this does not apply to all machine learning algorithms because some machine learning algorithms have a short test time.

Interpretable

Crucially, we use interpretability as a factor in comparing machine learning and deep learning.

Let's look at an example. Suppose we apply deep learning to automatically rate articles. Deep learning can reach people's standards, which is quite amazing performance. But this still has a problem. The deep learning algorithm won't tell you why it gives this score. Of course, from a mathematical point of view, you can find out which deep neural network node is activated. But we don't know what models neurons should be, and we don't know what these neural unit layers should do together. So I can't explain how the results are produced.

On the other hand, in order to explain why the algorithm is so chosen, machine learning algorithms like decision trees give clear rules, so it is easy to explain the reasoning behind the decision. Therefore, algorithms such as decision trees and linear/logical regression are mainly used for industrial interpretability.

What areas is machine learning and deep learning used for?

Wikipedia's article on machine learning outlines all areas of machine learning. These include:

● Computer vision is used for license plate recognition and facial recognition applications.

● Information retrieval for applications such as search engines - including text search and image search.

● Marketing for applications such as automated email marketing and target group identification.

● Medical diagnosis such as cancer identification and abnormality detection.

● Natural language processing Applications such as emotional analysis and photo tagging.

The above diagram succinctly summarizes the application areas of machine learning, covering the broader topic of machine intelligence.

A prime example of a company that uses machine learning/deep learning is Google.

In the image above, you can see that Google is applying machine learning to its various products. The application of machine learning/deep learning is endless - you just need to find the right time!

Assault test

In order to assess whether you really understand the difference, we will conduct a test. You can post your answer in this post.

Be sure to include the following steps to ensure that you are fully answering each scenario.

● How will you use machine learning to solve the following problems?

● How do you use deep learning to solve the following problems?

● Conclusion: Which method is better?

scene 1

You must build a software component for autonomous vehicles. The system you build should take raw pixel data from the camera and predict what angle you should guide the wheel.

Scene 2

Given a person's credit and background information, your system should assess whether the person should be eligible for a loan.

Scene 3

You must create a system that translates Russian messages into Hindi messages so that Russian representatives can communicate with local people.

Find the discussion notes and opinions of various data scientists on the above issues.

Future trends

This article outlines machine learning and deep learning and their differences. In this section, I will share my views on the future of machine learning and deep learning.

• First, with the growing trend in the industry for data science and machine learning, it is important to value machine learning for every company that wants to survive. Apple is using machine learning in iPhone X, which marks the direction of this technology.

● In-depth study makes us surprised every day and will continue to do so in the near future. This is because deep learning is one of the best performing techniques proven to be the most advanced.

● Research on machine learning and deep learning will be ongoing. But unlike the previous years of research limited to academia, research in machine learning and deep learning will have explosive developments in both industry and academia. And with more funding than ever before, it is more likely to become the main theme of human development.