The rise of Machine Learning and Deep Learning

2024.06.10

Machine Learning

Abstract

Machine learning can be described as a part of AI which enables computers to learn new tasks and information automatically. This consists of developing models and algorithms that let computers draw conclusions or make forecasts from data patterns. Machine learning has gained popularity recently for its ability to handle datasets and make faster more accurate decisions compared to traditional methods. This paper discusses the comparison, between ML and DL, and the rise of ML in recent years and the increasing use of neural networks as well as the concepts of supervised and unsupervised learning.

Introduction

One aspect of ML is using algorithms to analyze data for making predictions/decisions. This process involves training a model with labeled datasets containing input data paired with output values. By identifying patterns and relationships within this data the model can then predict outcomes for data.

Machine learning applications can be seen in aspects like in speech recognition, image processing, natural language understanding and self driving vehicles. Additionally it plays a role in sectors, like healthcare, finance and marketing by analyzing data to support decision making processes.

Definition of Machine Learning
(Arthur, Samuel) described machine learning as the academic discipline that empowers computers to acquire knowledge and improve performance without the need for explicit programming. Tom Mitchell, after a period of time, provided a definition for machine learning. He stated that machine learning refers to a computer program's ability to acquire knowledge from experience E in relation to a specific task T, and evaluate its performance using a certain measure. If the performance of P on T, as quantified by p, enhances with experience E

Example:
classifying email as spam and not spam is a task T

Watching you label email as spam and not spam is experience E

The number of emails classified as spam and not spam that is a performance P
some basic examples of machines learning
Web search: an algorithm that rank the pages based on the input word. For example if you search the word “biology” on the web it will list results based on the word biology.
Most of us we use Gmail, so Gmail classifying email as spam and not spam is also supervised machine learning

Comparison between ML and DL
First both fall under the category of AI

An Example of ML vs DL

Picture a system that can identify basketballs in images, allowing us to look into the distinctions between machine learning and deep learning. For optimal performance, every system requires a well-designed algorithm for detection and a vast collection of images, including both basketball-containing and not basketball images, for thorough analysis.

Before image detection can occur in a machine learning system, a programmer must first define the characteristics or features of a basketball, such as its relative size and orange color. Once that's completed, the model has the capability to analyze the photos and provide images that include basketballs. With increased frequency of performing this task, the model is expected to improve. Just like a data scientist, a human can review the results and make adjustments to the processing algorithm in order to enhance accuracy.

Creating an Artificial Neural Network for the Deep Learning system requires the expertise of a skilled programmer, who meticulously designs multiple layers, each with its own unique purpose. Defining the characteristics of a basketball is unnecessary for a programmer. Just like a data scientist, the system is able to learn and determine the characteristics of a basketball when the images are inputted into the neural network layers. They utilize their expertise to analyze the images. Just like a data scientist, the Deep Learning system constantly evaluates the accuracy of its results and makes automatic updates to enhance its performance without any need for human intervention.

This example further illustrates the proper use of technology for a given task. Machine Learning is excellent for image detection, while Deep Learning may be too advanced (and challenging to set up and operate) for this particular application. Deep Learning is more suitable for tackling more intricate tasks. Building a Deep Learning system into an autonomous car's self-driving system can enhance its capabilities. It can be programmed to identify potential risks, such as balls bouncing into the road, and respond promptly to ensure safety.

Why ML is needed these days?

1. Growth of automation

2. Growth of the web

3. Big companies like silicon valley uses ML collect web data to mine users better and serve the users better preference

4. In medical areas, to read some diseases better

5. In biology: to read DNA sequences

6. Self customizing programs: watching YouTube, Amazon, Netflix, they give us recommendation based on our previous search as the computer learned what our preferences are from previous searches.

ML can be used in areas that we cannot program by our hand for example: if we want to a helicopter to fly by itself, this can be done by having a computer to learn by itself using algorithms

Self customizing programs: watching YouTube, Amazon Netflix, Taobao (Alibaba) they give you recommendation based on your previous search as the computer learned what your preferences are from previous searches

Deep learning

1. Introduction to deep learning

Deep learning is a fascinating field within AI that centers on developing algorithms that draw inspiration from the complex structure and functionality of the human brain.
Neural networks play a crucial role in this field, as they are intricate networks of interconnected nodes that mimic the information processing of neurons in the brain.

DL’s ability to learn and improve automatically from experience has made it a powerful tool for developing advanced technologies and applications.

Deep learning has an advantage, in identifying patterns and extracting insights from data sets that might be overwhelming for analysts. This technology plays a role in developing self driving cars offering recommendations on streaming services and enhancing medical diagnostic tools. The training of learning models involves techniques such as unsupervised or reinforcement learning based on the specific problem being addressed. Popular frameworks like TensorFlow, PyTorch and Keras provide the tools and libraries for building and training networks. Deep learning models are characterized by their interconnected layers of nodes that enable them to understand data representations. The success of learning can be attributed to advancements in data accessibility and cutting edge algorithms that have significantly improved network performance. Despite its accomplishments challenges such as the need for labeled data, risks of over fitting and the complexity of interpreting models continue to pose obstacles, in the field of learning. Recent studies in this field are concentrating on enhancing network capabilities such as improving generalization, to data sets and adapting to dynamic environments.

In the coming years progress, in learning is anticipated to prioritize enhancing effectiveness cutting down on expenses and creating models that are easier to understand and interpret. The ongoing development of learning is foreseen to influence various sectors such, as healthcare, finance and technology.

Overall, deep learning is an incredible, innovative technology that could change how we work, communicate and interact in the world

In the past time electricity once transformed the industry of transportation, healthcare, communications and more. AI is the new electricity which is going to bring a huge transformation
we have defined machine learning as a science of getting computers to learn without being explicitly programmed. This can be done using learning algorithms called neural networks

CNN: convolutional neural networks (applied to images)

RNN: recurrent neural networks (applied to audios)

LSTM: long short term memory models

What are neural networks?

Neural networks are a stack of neurons

ReLu Function: rectified linear unit

Neural networks help us to get an output (y) by from an input (x)

Supervised learning: the idea is we are going to teach the computer how to do something. In short, the machine is trained with labeled data. For example, if we already labeled as spam and non spam after that the computer can differentiate emails as spam and non spam
Unsupervised learning: we input a lot of data so that the algorithm can find a patter for us. The algorithm learns by itself without pre-labeled data, we are going to let it learn by itself. For example (clustering in Google news, there are many new stories everyday many broadcasting companies like CNN , BBC, guardians post news of the same topic with different headlines and different links, so what Google news does is cluster this results which are talking about the same topic

Application of Machine learning

The rise of neural networks enables computers to be better in interpreting unstructured data.

For structured Data, meaning as mentioned above like the house price number of bedrooms like most data are structured. Unstructured data: audios, images, texts which are more complex compared to the structured data.

• Real estate and online advertising use the standard neural networks

• Image application use CNN (convolutional neural networks))

• Sequence data like audio to text, language translations use RNN (recurrent neural networks) Autonomous Driving (custom hybrid neural networks)

The rise of neural networks in recent years, why?

In the past time, using the traditional learning algo the more data we give the performance graph is like a plateau but in recent years we have more data because of digitalization, so many human activities on digital devices, people spend more time and we collect more data in these days

There are several reasons for the rise of neural networks in recent years:

1. Technological advancements; Faster and more powerful computers have enabled the training and deployment of networks allowing researchers to explore larger networks and advanced algorithms that enhance performance and accuracy.

2. Data availability; the abundance of data, in today’s era offers researchers a wealth of information to train networks resulting in more precise predictions and valuable insights.

3. Enhanced algorithms; Researchers have introduced improved algorithms like learning and reinforcement learning for training neural networks boosting their capabilities.

4. Industry acceptance; Many companies across sectors, including healthcare, finance and technology have embraced networks for diverse applications. This increased adoption has spurred research and development efforts, in the field of networks. Success, in the field of research; Neural networks have proven to be highly effective in tackling challenges and attaining cutting edge outcomes across areas like image and speech recognition natural language processing among others. This accomplishment has sparked curiosity and exploration into networks resulting in their growing appeal. To enhance performance, in networks one must consider training networks or incorporating more data.

We have to train a large set of data to improve neural networks performance.

Abbreviations

AI- Artificial intelligence

ANN- Artificial neural network

CNN- Convolutional neural network

DL- Deep learning

DNN- Deep neural network

ML- Machine Learning

ReLU- rectified linear unit

RGB- Red green blue color model