Machine learning can be considered the most widely used branch of artificial intelligence that has spread to all levels of daily life. 

Understanding machine learning can help you better understand the world to come.

Artificial intelligence (AI) is everywhere. You may be using it right now and not know it yourself. One of the most popular applications of AI in custom software development is machine learning (ML).

Computers, software, and machine learning devices function similarly to the human brain and perform tasks through cognition.

Machine learning allows computers to perform human tasks. Machine learning, by performing tasks such as driving a car or translating speech, revolutionizes artificial intelligence and helps software understand a chaotic and unpredictable real world; But what exactly is machine learning and what are its uses?

In this article, we will give a comprehensive answer to this question.

What is machine learning?

By definition, machine learning is a branch of artificial intelligence (AI) focused on building applications that learn from data and increase their accuracy over time without the need for a programmer. In data science, an algorithm consists of a sequence of statistical processing steps.

In machine learning, algorithms are taught to find patterns and specifications in large amounts of data so that they can make decisions and predictions based on new data. The better the algorithm, the more accurate the output decisions and predictions.

Predicting machine learning can include answering questions such as recognizing a fruit in an image, recognizing people crossing the street, recognizing precise speech to generate YouTube video captions, or separating email and spam.

The key difference between machine learning and older computer software is that the human developer does not write machine learning code. Rather, the machine learning model learns how to distinguish between elements in a large body of data. The key to machine learning is the sheer amount of data that makes it possible to learn.

The difference between artificial intelligence and machine learning

Machine learning has been very successful recently; But it is only one branch of artificial intelligence. In the early days of the advent of artificial intelligence in the 1950s, the definition was given to it: any machine that is able to perform tasks and needs human intelligence. Each AI system has one or more of these features:

• planning
• Learn
• Conclusion
• Problem Solving
• Provide information
• Understand
• Move
• Information manipulation
• Social intelligence and creativity

In addition to machine learning, there are several other ways to build artificial intelligence systems, including evolutionary computations in which algorithms undergo random mutations and evolve for optimal solutions. Another type of artificial intelligence is expert systems.

In expert systems, computers are programmed with rules that allow them to mimic human behavior in specific areas, such as automated aircraft guidance systems.

^{Machine learning can be used to improve data centers or data centers}

Learning how the machine works

There are four main steps to building an application or model of machine learning. Data scientists are working closely with business experts to develop these steps:

Step 1: Selecting and preparing the educational database

Educational data is a set of data that represents the machine learning model and is used to solve specific problems. In some instances, the learning data is labeled to call features and classifications for the model. Other data are unlabeled. In this case, the model must extract some properties to assign the classifications.

In both cases the learning data must be well prepared. The preparation process involves randomizing and examining deviations that affect learning. The data at this stage is divided into two subsets: the learning subset that is used to teach the program and the evolutionary subset that is used to test and modify it.

Step 2: Select an algorithm to execute the training dataset

The algorithm consists of a set of statistical processing steps. The type of algorithm depends on the type of data (labeled or unlabeled) or the amount of data in the training dataset as well as the type of problem.

 Common types of machine learning algorithms that can be used with tagged data are:

• Regression Algorithms : Linear and logical regression are examples of regression algorithms used to understand the relationships between data. Linear regression is used to predict the value of a dependent variable based on the value of an independent variable. Logical regression is also used when the dependent variable is binary in nature.
•   Tree decision: Decision trees classified data to make recommendations based on their decision rules. For example, the decision tree that recommends betting on a horse’s board is like this.
• Sample-based algorithms : A good example of a sample-based algorithm is the nearest neighbor algorithm K or the k-nn algorithm. This algorithm uses classification to estimate the probability of having data points in a group.
• The algorithm of the categories of : categories imagine the same group. Categorization focuses on identifying groups with similar records and tagging records based on the group to which they belong. This is done without prior knowledge of the groups or their characteristics. Types of classification algorithms include: K-Mean, TwoStep Classification, and Cohennon Classification
• Algorithm Relational : the algorithms look for patterns and relationships in data interfaces and relationships are repeating “if it ‘or if-then identify. These relationships are similar to the rules used in data mining.
• Neural Networks : A neural network is an algorithm that defines layered networks of computations including the input layer. There is at least one hidden layer whose calculations have different results on the input, and there is an output layer in which each conclusion is assigned to a probability. A deep neural network is a network with multiple hidden layers, each of which refines the results of the previous layer.

Step Three: Teach the algorithm to build the model

Algorithm training is an iterative process. This step involves executing the variables in the algorithm, comparing the output with the expected results, adjusting the weights and deviations within the algorithm, and restarting the variables until the algorithm returns the correct result. The resulting trained algorithm is a kind of machine learning model.

Step 4: Use the model and improve it

The final step is to use the model with new data and, at best, improve its accuracy and performance over time. The source of new data also depends on the type of problem. For example, a machine learning model designed to detect spam receives email messages as input, while a machine learning model that drives a vacuum cleaner robot receives data about actual interaction with home furniture or new objects in the room as input. he does.

Types of machine learning

Machine learning is divided into two main categories: supervised and unsupervised learning.

Supervised learning

In the supervised learning process, systems are exposed to large amounts of labeled data, such as images of handwritten digits. The supervised learning system, despite having a sufficient number of instances, can identify groups of pixels and shapes associated with each number and recognize handwriting numbers.

However, machine learning training requires a large number of labels, so that some systems require millions of labels to master a task. As a result, the datasets used to teach these systems are vast.

For example, Google’s Open Images Database contains approximately 9 million images, or YouTube-8M is a source of tagged videos that link to seven million tagged videos, and ImageNet is one such database with 14 million images categorized.

The size of educational datasets is increasing as Facebook compiles approximately 3.5 billion public Instagram images using hashtags attached to each image as a tag. With one billion of these images for image recognition training, the accuracy level was 85.4 based on the ImageNet index.

The tedious process of tagging the datasets used in machine learning training is often done using mass-service services such as Amazon Mechanical Turk, which allow access to a large source of low-cost labor around the world. ImageNet, for example, recruited 50,000 people over two years through Amazon Mechanical Turk.

However, Facebook’s method of using public data to train systems can be an alternative way to train systems using strong datasets, and does not require manual labeling.

Learning without supervision

In contrast, unsupervised learning algorithms try to find similarities between them by identifying data patterns. An example of this is the Airbnb or Google News categorization system, which allows you to group stories with a similar theme. Unsupervised learning algorithms are not designed to separate specific types of data but merely follow similar data.

Semi-supervised learning

The importance of large sets of tagging data for machine learning training decreases over time due to the advent of part-time learning. As the name implies, this is a combination of supervised and unsupervised learning. Semi-supervised learning depends on a small set of labeled data and a large set of unlabeled data for systems training.

Labeled data is used to partially teach the machine learning model, then this model is used to label unlabeled data. This process is called quasi-labeling. The model is then taught based on a combination of quasi-labeled data and labeled data.

The duration of semi-supervised learning in GANs has been improved. These machine learning systems can use labeled data to generate completely new data and then reuse this data for the machine learning model.

If semi-supervised learning is as optimal as supervised learning, access to large amounts of computational power will be more important to the success of machine learning systems than access to large labeled datasets.

What is reinforcement learning?

Reinforcement learning is a type of machine learning model similar to supervised learning, but its algorithm is not trained using sample data. This model is trained during the implementation process through trial and error. A sequence of successful outputs is reinforced to develop the best suggestions or policies for a particular issue.

To better understand reinforcement learning, suppose someone wants to run an old computer game for the first time. One is not familiar with any of the rules or methods of controlling the game until one gradually realizes the way the game is played and improves one’s performance by observing the relationship between the buttons and what happens on the screen.

An example of reinforcement learning is Google’s Deep Q network from Deep Mind, which has defeated humans in a series of old video games. The system feeds on the pixels of each game and displays various information about the status of the game, such as the distance between objects on the screen.

It then describes how the game is run and how it works and how it relates to scores. During the process of a large number of game cycles, the system eventually builds a model according to which some actions maximize the score under certain conditions.

Evaluation of machine learning models

The model is evaluated after completion using residual data that were not used during the training. When teaching a machine learning model, approximately 60% of the data set is used for training.

Another 20% is used to evaluate model predictions and adjust additional parameters to optimize model output. Model settings are designed to improve the accuracy of model predictions when dealing with new data.

Neural network and how to train them

Neural networks are one of the most important algorithms in both supervised and unsupervised machine learning. These networks form the basis of a large part of machine learning. On the other hand, simple models such as linear regression can be used to make predictions based on a small amount of data.

Neural networks are useful for working with big data that has many properties. Inspired by the human brain, neural networks are a collection of interconnected algorithmic layers called neurons that input data so that the output of the previous layer acts as the input of the next layer.

Each layer can be considered a set of different characteristics of the general data. For example, consider using machine learning to recognize handwritten numbers between 0 and 9.

The first layer in the neural network measures the density of individual pixels in the image, the second image is able to recognize shapes such as lines and curves, and the final layer classifies the handwriting digit into groups 0 to 9.

The network is able to learn the pixels of the shape of numbers during the training process by gradually changing the priority of the data flowing between the layers of the network.

Each connection between layers has a weight whose value decreases or increases to change the importance of the connection.

At the end of the learning cycle, the system checks how far the final output of the neural network is from the target.

For example, has the network improved in recognizing script number 6?

To close the gap between the actual output and the desired output, the system works inversely on the neural network and changes the weights attached to the entire connection between the layers as well as the associated value called the deflection.

This process is called reverse replication.

Finally, this process is performed on values ​​based on weights and deviations, and allows the network to perform specific tasks such as recognizing handwritten numbers, and it can be said that the network has learned how to perform specific tasks.

Deep learning and deep neural networks

The machine learning subset is called deep learning. In deep learning, neural networks develop into scattered networks with a large number of layers that include a large number of units; These units are taught using large amounts of data.

Deep neural networks can improve the ability of computers to perform tasks such as machine vision and speech recognition.

There are different types of neural networks with different strengths and weaknesses. Recursive neural networks are suitable for tasks such as language processing and speech recognition, while torsional neural networks are commonly used in image recognition.

Neural network design is also evolving, and researchers have come up with a more optimal design for an effective type of neural network called short-term memory, or LSTM, which is fast for demand-based systems such as Google Translate.

The artificial intelligence technique of evolutionary algorithms and a process called neural evolution is used to optimize neural networks. The method was unveiled by Uber AI Labs, which published articles on the use of genetic algorithms to teach neural networks for reinforcement learning problems.

Is machine learning implemented only with neural networks?

The answer to the above question is no. There is a set of mathematical models that can be used to teach systems and predictions. A simple model involves logical regression, while its name can be used to categorize data, for example, to separate spam from non-spam. When implementing simple binary classification, logical regression can be easily implemented and can be used to label more than two sets of data.

Another common model is support vector machines (SVMs), which are widely used for data categorization and regression prediction. SVMs can categorize data even if the mapped data is irregular and difficult to classify. To achieve this, SVMs perform a mathematical operation called a kernel trick that depicts data at new values ​​so that they can be broken down into separate categories.

The choice of machine learning model depends on several criteria such as the size and number of features in the database, and each model has advantages and disadvantages.

The reason for the success of machine learning

Although machine learning is not a new method, interest in the field has increased dramatically in recent years. The outburst followed a series of developments: in-depth learning set new records in areas such as speech and language recognition as well as machine vision.

These successes depended on two factors: the large number of images, speech, video, and text available to teach machine learning systems.

But more importantly, the emergence of large amounts of parallel processing power, especially in the field of modern graphics processing units (GPUs) that can be used for machine learning power plants.

Today, anyone with an Internet connection can use cloud services such as Amazon, Google, and Microsoft to teach machine learning models. With the increasing use of machine learning, companies are looking to build special hardware to implement and teach machine learning models.

An example of these custom chips is Google’s Tensor Processing Unit (TPU), which speeds up machine learning, at which speed machine learning models are built using Google’s TensorFlow libraries and can separate information from data.

The chips are used not only to train models such as Google DeepMind and Google Brain, but also for features such as Google Translate and image recognition in Google Photo, as well as services that allow you to build machine learning models using the TensorFlow Research Cloud feature.

 The third generation of these chips was unveiled in May 2018 at the Google I / O conference.

In 2020, Google announced the fourth generation of TPUs at 2.7 times faster than the previous generation TPUs in MLPerf. Based on this index, the system is able to infer at high speed through a trained machine learning model. Continuous TPU enhancements allow Google to improve services related to premium machine learning models.

For example, models can be taught to Google Translate.

With the increasing advancement of hardware and the refinement of machine learning software frameworks, it will become common to perform machine learning tasks for smartphones and computers and use them instead of cloud data centers.

In the summer of 2018, Google went a step further by offering offline automatic translation for 59 languages ​​in the Google Translate app for Android and iOS.

Alphago

Perhaps the most famous symbol of the efficiency of machine learning systems can be considered Alphafa’s artificial intelligence following Google’s deputation of Grand Master Go, which does not seem to be predictable until 2026.

Go is a traditional Chinese game whose high complexity has puzzled computers for decades. The player has 200 possible moves per turn, while the number of moves for playing chess reaches twenty.

There are many possible moves throughout the gameplay, each of which is very costly to search to identify the best game performance from a computational point of view.

But Alfago had trained more than 30 million players based on the movements of skilled human players, and the data entered the game’s deep learning neural networks.

Deep learning network training requires a lot of time and huge data sets that must be gradually injected into the system to modify the model and achieve the best output.

However, Google recently improved the training process in the AlphaGo Zero version.

This system is able to run completely random games for itself and learn from their results. Demis Hasabis, CEO of Google DeMand, unveiled Alfazero at the 2017 Neural Information Processing Systems Conference (NIPS), a generalized version of the Alfago Zero that specializes in chess and chess games.

DePaymand advances in machine learning continue. In July 2018, the company unveiled artificial intelligence agents who have learned how to play the Quake III Arena 3D first-person shooter and can easily defeat human players.

These agents used the same information that human players use, except that they use their performance feedback during the game.

The most influential application of DePaymand research emerged in late 2020 with the unveiling of AlphaFold 2.

The capabilities of this system are an indicator of progress for medical science.

AlphaFold 2 is a neural network focused on attention that has the potential to accelerate disease modeling and drug development. This system can plot the three-dimensional structure of proteins by analyzing the building blocks of their amino acids.

In protein structure prediction competition assessments, AlphaFold 2 is able to determine the three-dimensional structure of the protein with an accuracy comparable to crystallography and can model precise protein structures in a matter of hours.

Are machine learning systems purposeful?

The selection and breadth of data used to train systems affect the tasks associated with them. There are also many concerns about how to deviate from social deviations and inequalities in machine learning systems.

For example, Rachel Tatman, a researcher at the Natural Sciences Foundation and the Department of Linguistics at the University of Washington, found that Google’s speech recognition system for automatically captioning YouTube videos to male voices performed better than female voices.

According to the findings, face recognition systems also encountered problems in identifying women and people of color. Racial errors in such systems led manufacturers to stop selling some face recognition systems to the police.

In 2018, Amazon abolished the hiring machine learning tool, which was considered desirable by a large number of male applicants.

As machine learning systems move into new areas, such as assisted medical diagnosis, the possibility of systems deviating from offering better services or fair treatment to a specific group of people has become a major concern. Current research is moving towards neutralizing deviations in self-learning systems.

Environmental effects of machine learning

The environmental effects of strengthening and cooling computing farms used to train and set up machine learning models were the subject of a 2018 World Economic Forum article. According to one 2019 estimate, the power required by machine learning systems doubles every 3.4 months.

On the other hand, the size of the models and datasets used for day-to-day training is increasing, for example, the GPT-3 language prediction model is a kind of distributed neural network with 175 billion parameters. As a result, concerns about the carbon footprint of machine learning increase.

The energy required for learning models is also increasing day by day, but the cost of setting up trained models is increasing with the increasing demand for machine learning-based services.

On the other hand, predictable machine learning capabilities have a significant and positive impact on many key areas from the environment to healthcare.

Machine learning courses

One of the recommended courses for beginners in the field of machine learning is the free Stanford University kits  offered by Andrew Angie, an artificial intelligence expert and founder of Google Brain.

Angie has recently published an in- depth learning course focusing on a wide range of machine learning topics and topics, as well as neural network architectures.

If you are looking for top-down learning, you can start by implementing machine learning instructional models and then addressing minor topics. An in-depth learning course for coders from fast.ai is also recommended for developers with experience working with Python.

Both courses have strengths: The Angie course begins with an overview of machine learning theory topics, while the fast.ai proposal revolves around the Python axis. When machine learning data is widely used among engineers and scientists.

Another free course with high quality instruction and extensive coverage is a introduction to machine learning from Columbia University and Edx . Of course, this course requires a basic knowledge of university-level mathematics.

Available machine learning services

All major cloud platforms, including Amazon Web Services, Microsoft Azure, and Google Cloud, provide access to the hardware needed to train and deploy machine learning models.

Users on the Google Cloud Platform can test TPUs. TPUs are custom chips optimized to train and drive machine learning models.

Cloud infrastructure includes the data centers needed to store large amounts of training data, data analysis preparation services, and visualization tools to display results clearly.

New services also make it possible to build custom machine learning models. Google, for example, offers a service called Cloud AutoML to automate the construction of smart models.

This Drag and Drop service allows you to build custom image recognition models and the user does not need any expertise in machine learning to work with it.

Similarly, Amazon has its own AWS services designed to speed up the learning process of machine learning models.

The Google Cloud AI platform for data scientists manages the machine learning service, enabling the training, development, and output of custom machine learning models based on the TensorFlow Google Open Source Learning Framework, or Keras, an open neural network framework.

Database admins with no background in data science can use Google’s beta service called BigQueryML, which allows admins to contact trained machine learning models with SQL statements and predict the database; This process is simpler than outputting data to separate machine learning and analytical environments.

For organizations that are not looking to build their own machine learning models, cloud platforms offer artificial intelligence demand-based services such as voice, visual and language recognition.

At the same time, IBM is trying to sell specialized artificial intelligence services such as healthcare to retail, alongside demand-based and more general offerings.

These offers are managed by the IBM Watson Group.

In early 2018, Google developed machine-based services in the world of advertising and provided a set of tools needed for optimal physical and digital advertising.

On the other hand, while Apple is not as credible as Google and Amazon in natural language processing, speech recognition, and machine vision, it is investing in improving its artificial intelligence services.

The former head of Google Machine Learning is now in charge of Apple’s AI strategy division, which is dedicated to developing Siri Assistant and the Core ML machine learning service.

In September 2018, Nvidia launched a hybrid software and hardware platform that will be installed on data centers.

This platform can speed up the learning of machine learning models for audio, video and video recognition as well as other machine learning related services.

Software libraries

There is a wide range of software frameworks for getting started and learning machine learning models, especially for the Python, R, C ++, Java and MATLAB programming languages. Python and R are the most widely used and popular languages ​​in the field of machine learning.

Popular examples of machine learning frameworks and libraries include Google TensorFlow, the Keras open source library, the Python scikit learn library, the CAFFE deep learning framework, and the Torch machine learning library.

Applications of machine learning in everyday life

The interest in machine learning is growing in various fields as the amount of data available increases over time. Machine learning has many techniques for extracting information from data and can use this information purposefully.

Machine learning algorithms can enhance field information and automated tuning and optimization functions. In addition, machine learning along with machine learning can be useful in many areas such as medical diagnosis, statistical data analysis and algorithms, scientific research and many other areas.

Today, machine learning can be found in smartphone applications, computer devices, online websites, cybersecurity and many more. The following are the most common uses of machine learning in everyday life.

^{Applications of machine learning in everyday life}

Displacement estimates

In general, the time of a trip may be longer than the average time; Because multiple modes of transportation, including traffic scheduling, affect the time to reach the destination.

Reducing travel time is not an easy task, but here’s how machine learning can help:

•  Google Maps : Google Maps uses location data to track traffic speeds at any given time. In addition, the map can organize user-reported traffic, including the likelihood of construction, vehicle traffic, and accidents. Google Maps can reduce travel time by displaying the fastest route by accessing relevant data and appropriate algorithms.
•   Riding apps : Things like paying for a ride and reducing waiting times are some of the problems with urban transportation; But machine learning helps solve this problem and helps companies estimate transportation costs, calculate the optimal riding position, and ensure the shortest route and detect deception. Uber, for example, uses machine learning to optimize its services.
•   Commercial flights and the use of autopilots: With the help of artificial intelligence technology, autopilots control current flights. According to the New York Times, pilots control the flight manually for only seven minutes during takeoff and landing, and other flights continue as autopilots or autopilots.

 Email intelligence

• Spam filters: Some rule-based filters do not play an active role in the inbox. For example, when you receive a message with the words “online consultation”, “online pharmacy” or “anonymous address”. Machine learning provides the powerful ability to filter a variety of email cues, including words in a message, message metadata. Gmail uses machine learning to filter 99.9% of spam messages.
• Class of email : Gmail, email groups, Primary, Promotions, Social and Update classifies and categorizes emails by importance.
• Smart Answers : Gmail can send simple phrases like “thank you” in response to the reader. These responses are customized based on each email.

 Banking and Finance

• Deception prevention: In most cases, the volume of daily transaction data is very high and manually checking each transaction is a complex process. To solve this problem, artificial intelligence systems have been designed to detect misleading transactions. Banks use this feature of artificial intelligence. Companies also use neural networks to identify fraudulent transactions based on criteria such as last transaction repetition, transaction size, and type of retailer.
• Credit decisions : Financial institutions use machine learning algorithms to make credit decisions and assess each user’s risk.
• Check checks on mobile: In addition, artificial intelligence technology has made banking easy and customized for people who do not have time to visit the bank. Banks, for example, allow checks to be verified through a smartphone app, and the user no longer needs to physically deliver the check to the bank. Most banks use Mitek technology to interpret and translate checkscripts into text through character recognition.

 Assessment

• Plagiarism Investigation : Machine learning can be used to build a plagiarism detector. Many schools and universities require plagiarism assessors and students’ writing skills analysis. The plagiarism algorithm uses similarity functions that examine the equality of two documents.
• Automated readers : In the past, dissertation ranking was a very complex task, but now researchers and organizations are looking to build artificial intelligence systems for scoring and ranking a dissertation. The GRE test ranks articles through a human reader and an automated reader called an e-Rater. If the scores differ significantly, a second human reader is considered to estimate the difference.

In the near future, current classes will be replaced by flexible and customized learning that takes into account each student’s strengths and weaknesses. Machine learning also helps identify at-risk students, so schools can pay more attention to these students by offering extracurricular learning courses. For example, artificial intelligence in education helps with custom learning, voice assistants, and managerial tasks.

Social Networks

• Facebook: When uploading an image to Facebook, faces are automatically detected and tagged friends. Facebook uses artificial intelligence and machine learning to identify faces. Other uses for artificial intelligence on Facebook include:

•  It uses the ANN algorithm, which mimics the human brain and enhances facial recognition software.
•    Facebook users use artificial intelligence to customize the newsletter and ensure that posts are reflected.
• Displays ads related to specific interests and businesses.

• Pinterest : Pinterest uses machine vision to automate the detection of objects in images or to “pin” them, and then recommends similar pins. It also uses other AI and machine learning capabilities to search, discover, email marketing and improve ad performance.
• SnapChat : Offers face filters (called lenses) that filter and navigate facial activity, allowing users to tag moving images or digital masks that move when the face moves.
• Instagram: With the help of machine learning algorithms, the purpose of emojis can be identified. Instagram offers automatic emojis and emoji hashtags. Emoji can also be widely discovered by translating emojis into text.

Health and medical diagnosis

Machine learning uses a set of tools and methods to diagnose and predict problems in various fields of medicine. Machine learning algorithms in medicine are used for the following:

• Analyze medical data to discover patterns in the data
• Inadequate data control
• Description of data generated by medical units
• Optimal monitoring of patients

Machine learning helps to estimate the progression of the disease, control medical information to research outputs, plan and assist in the treatment and overall management of the patient. In addition to machine learning, artificial intelligence is also used for effective monitoring.

Smart personal assistant

From Siri to Cortana to Google Assistant, Amazon Alexa and Google Home, personal assistants have many benefits and capabilities. With the implementation of artificial intelligence, these assistants are subject to commands such as setting reminders, searching for information online, controlling lights, and much more.

Personal assistants include machine learning chat bots based on machine learning algorithms that gather information, understand priorities, and improve the user experience based on interaction with individuals.

Concludin Remarks

Machine learning by processing and analyzing massive data has simplified many human tasks. This technology is currently used in many applications and devices in everyday life, including smartphones, home appliances and social networks.

In the field of science, machine learning has made many heavy and tedious tasks easy for researchers.

Due to the growing trend of this technology in the future we will see wider applications of machine learning in all areas of human life.