One of the fastest-growing areas of study in artificial intelligence and chatbot development is machine learning, but what is machine learning exactly? Well, in this post, we’ll try and answer that question and get you familiar with the basics of machine learning.
We’ll explain what it is, how it works, its applications, and the different types of Machine Learning. Now that I have your attention, let’s get into the details.
What Is Machine Learning?
Machine Learning is a branch of artificial intelligence that involves learning by recognizing patterns, instead of following a specific set of rules.
Software developed using Machine Learning has the capacity to learn from their mistakes. Meaning they evolve with experience.
Here are the three steps process of Machine Learning:
- It takes in the data
- Recognizes the patterns within the data
- Then adjust accordingly to those patterns
Applications of Machine Learning
In this section, we’ve listed all the applications of Machine Learning. There are a total of 11 applications and we’ve described all of them in detail.
1. Speech Recognition
Speech recognition is the act of understanding a set of instructions given using voice. Converting voice into text comes under the category of speech recognition as well.
Most of us use this application on a daily basis. A good example is when we give voice commands to our smartphones. Alexa, Siri, and Cortana, all these famous bots use speech recognition to understand commands.
2. Automatic Language Translation
As you can tell by the name, this feature is used for translating words from one language to another. GNMT (Google Neural Machine Translation) uses this to help hundreds of thousands of people every year.
This technology is especially great for tourists that are traveling in a foreign country. No matter what they need to say, they can open GNMT, type, or speak the words in their language, and the rest will be taken care of.
3. Traffic Prediction
It’s also used for predicting traffic. If you want an example of machine learning being used for predicting traffic, Google Maps is a good one.
It can tell us the shortest route to our destination and it can tell us whether the traffic is high, medium, or low. It uses these two metrics to make the prediction.
- Average time it took vehicles to complete the distance at the given route at the same time.
- Location of the user’s vehicle.
Everyone who uses Google Maps can help improve the services by uploading and updating information about locations.
4. Image Recognition
Image Recognition is perhaps the most used application of machine learning. It’s used to identify objects, people, places, etc. One of the most popular features that uses Image Recognition is Facebook’s Automatic Friend Tagging Suggestion.
If you’re not familiar with this feature of Facebook, here is how it works. You upload an image of you and your friends, and Facebook’s Automatic Friend Tagging Suggestion gives you tagging suggestions with the name of the people who are in the photo.
5. Self-Driving Cars
Machine Learning plays a huge part in self-driving cars. Every company that is making these cars uses this technology. Tesla, Cruise, Voyage, Nauto, and every other company that makes self-driving cars uses unsupervised learning methods for detecting people and objects while moving.
6. Product Recommendation
Various companies use Machine Learning for Product Recommendations. Companies that use it are in both industries: E-Commerce and Entertainment.
You might have noticed that whenever you open a product page a few times, you’ll start getting ads about similar products whenever you’re at the Amazon website.
When you search for a keyword multiple times on Google Search Engine, you’ll start seeing articles of similar topics on Google’s Feed.
On Netflix, you get recommendations of Movies and TV Series that are like the ones that you usually watch. All these companies are using Machine Learning for recommending their products.
7. Online Fraud Detection
This technology has also helped a lot in preventing online fraud. There are many ways in which fraud can take place in an online transaction. Fake Ids and fake accounts are just two of many ways.
Machine Learning has a part called FFNW (Feed Forward Neural Network). The FFNW helps detect whether a transaction is genuine or fake.
Here is how it does this. Each transaction involves two parts. In the first part, the machine creates a pattern of hash values. In the second part, the machine checks whether or not that pattern is the same. If the pattern hasn’t changed, it means the transaction is genuine. If there is even a slight difference in the pattern, the FFNW understands that the transaction is fake and it stops the transaction at that very moment.
8. Stock Market Trading
Machine Learning is used very often in stock market trading. If you know anything about the stock market, it would be the fact that shares can go very up or down at any minute.
This technology gives a good idea about whether the stocks will go up or down in a given timeframe.
9. Virtual Personal Assistant
As you can tell by the name, a virtual personal assistant can help us by listening to our voice. It can help us with various tasks such as locking doors, playing music, turning accessories such as an oven on or off, making the room cold or warm by controlling the HVAC (Heating Ventilation Air Conditioning system).
You can simply tell the VPA to turn off lights, increase the volume, switch to cool, and it’ll perform those tasks. Without the ML algorithm, the VPA wouldn’t be able to perform these tasks.
Here is how the process looks like. You give a voice instruction to a VPA, it records the instruction, sends it to a server on the cloud, decodes it using ML algorithm, and acts on the instruction.
WIthout ML algorithms a VPA does not decode the instructions therefore it can’t act on them. Some of the well-known VPAs are Alexa, Cortana, Siri, and Google Assistant.
10. Medical Diagnosis
ML is doing wonders for Medical Diagnosis. Before this technology, it was pretty hard to find and diagnose various illnesses and disorders. This technology has made the process much easier.
Because of ML medical science can build 3D models that can predict the exact position of lesions in the brain. ML is helping medical diagnosis grow at a rapid pace and with time it’s expected to not only make it easier to diagnose the diseases that are hard to diagnose at the moment, but to help with diagnosing diseases that are almost impossible to do with current technology.
11. Email Spam and Malware Detection
You have probably noticed that Gmail automatically marked the email as spam, normal, and important. The normal and important emails are placed in the inbox section. The normal email has no sign on them and the important emails have the important sign on them. As for the Spam ones, they are placed in a different section called Spam.
This function is done with the help of Machine Learning. Below, we’ve mentioned the primary spam filters used by Gmail.
- Header filter
- Permission filter
- Rules-based filter
- General Black List filer
- Content filter
The three Machine Learning algorithms used for filtering email and detecting malware are the Naïve Bayes classifier, Multi-layer Perception, and Decision tree.
How Does Machine Learning Work?
Although the process of machine learning is quite complicated, we’ve described it in a simple and easy-to-understand manner.
Machines learn by finding patterns in data. You can think of data as information that we gain in the real world. The more information a person has the smarter they become, similarly the more data a machine has the smarter it becomes.
Another common thing that data has with information is that not all data is the same. Here is what we mean by it. If your goal is to become better at persuasive writing but you’re reading tips and doing exercises for narrative writing, the information you gain will be useless. To become a persuasive writer, you must acquire the right information.
Similarly, if you want your machine to get the job done, then you’ll have to give it the right data. This quote by Riley Newman explains the importance of good data very well “More data beats better models. Better data beats more data”.
While better data is important than more data, machines can only predict a future when they’re given a sufficient amount of data. This is because machines make new predictions based on the data they were given in the past.
In order for machines to make the right prediction, the new data has to be similar to the past data. If both have little to no similarities then the predictions generated by the machine won’t be useful.
Types of Machine Learning
There are a total of 14 types of Machine Learning. Although 3 of them do the most of the work, the rest 11 when combined together are a lot of value.
We’ve divided this section into two parts. The First will cover the 3 main types of machine learning and the second will cover the rest.
Let’s look at the name of all 14 types of machine learning first and then we’ll dive into the details. All types are divided into four categories.
The first part includes the Learning Problem category:
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
The second part includes the other three categories. Hybrid Learning Problems, Learning Techniques, and Statistical Interface.
Hybrid Learning Problems:
- Self-Supervised Learning
- Semi-Supervised Learning
- Multi-Instance Learning
- Active Learning
- Online Learning
- Transfer Learning
- Ensemble Learning
- Multi-Task Learning
- Deductive Interface
- Transductive Interface
- Inductive Learning
Category: Learning Problems
1. Supervised Learning
It’s easier for machines to learn from labeled data than from unlabeled data, which is why Supervised learning is the most common type of learning. Supervised Learning can solve two types of problems: Classification and Regression.
Classification: This type is used for predicting discrete values. An example is “Does he own a Rolex?” the answer will fall in one of these two categories Yes or No. This process is also known as a binary classification problem.
Regression: It involves predicting continuous values. Unlike discrete values, continuous values aren’t restricted to 1,2, or 3 values. The range of continuous value is unlimited.
This type of problem involves telling the size of something, telling the temperature, etc.
2. Unsupervised Learning
There is no labeled data in unsupervised learning. Machines are given a set of data and their job is to find patterns and group them. In other words, an unsupervised machine’s job is to learn on its own.
There are two methods of unsupervised learning for grouping data together: Association and Clustering.
Association: This method involves finding relations between two objects. A good example is when a store found out that people who buy diapers are very likely to buy bears. They found that most males who bought diapers for their babies also bought a pack of bears.
Clustering: This method involves clustering similar data. When the algorithm finds a similarity between two sets of data, it groups the similar parts together. The difference between this method and the Association method is that data gathered couldn’t be from two completely different things (like beer and diaper).
This method involves teaching the machine what to do by using the reward/penalty method. If the machine performs the right task, give it a reward. If the machine performs the wrong task, penalize it.
This type is the subset of artificial intelligence. Here are a few examples of Reinforcement learning. Training machine how to play Chess, Training machine how to play King of Fighters, Training machines to drive cars on their own.
This is all we had to discuss in this section. These were the 3 main types of Machine Learning. Now let’s discuss the rest.
Hybrid Learning Problems:
The difference between supervised learning and unsupervised learning can be subtle. In this section, we’re going to discuss the hybrid types of machine learning.
1. Self-Supervised Learning
Self-Supervised Learning is basically unsupervised learning presented in the form of supervised learning in order to provide supervised solutions for problems.
Supervised Learning is generally used for solving pretext or alternate tasks. This results in a representation or a model that can be used in the actual modeling problems.
A common example of Self-Supervised learning is computer vision where a bunch of unlabeled images are available for training the supervised model. The Self-Supervised Learning machine can use these images to make predictions such as colors of the image, removing blocks, missing parts, whether the image is grayscale, and more.
2. Semi-Supervised Learning
In Semi-Supervised Learning, there is a very little amount of labeled data and huge amounts of unlabeled data. Unlike Supervised Learning, the goal of Semi-Supervised Learning is to use all the data available, not just the labeled data.
For using the unlabeled data effectively, Semi-Supervised Learning machines sometimes use the methods of Unsupervised Learning, such as clustering and regression. Once some pattern has been formed, the methods from Supervised Learning methods such as regression and classification may be used to label the unlabeled data for making better predictions.
3. Multi-Instance Learning
Multi-Instance Learning is a part of Supervised Learning where individual examples are unlabeled and only groups of samples are labeled.
A big part of modeling is to know how many instances in a bag are associated with a target label and to predict the label for new bags in the future by recognizing the composition of multiple unlabeled examples.
Many techniques can be described as types of learning. In this section we’ll take a look at the 5 most common techniques.
1. Multi-Task Learning
Multi-Task Learning includes only one dataset, but it deals with multiple problems. The performance of the model is improved by working on various tasks rather than working on a single task.
Multi-Task learning can be a very useful approach for solving problems if there is an abundance of labeled data for one task that can be shared with another task that possesses little data.
A common example of Multi-Task learning is when the same word embedding is used for learning the distributed representation of words in text that is shared across multiple natural language processing.
2. Transfer Learning
Here is how Transfer Learning works. The machine is first trained on one task, then anywhere from some of the data acquired to all of the data acquired is transferred to a new task.
This approach is useful when the task that is trained first has plenty of similarities with the main task. The main difference between Transfer Learning and Multi-Task Learning is that tasks are learned sequentially in Transfer Learning, whereas in Multi-Task Learning the machine works on all the tasks at the same time parallel.
An example of Transfer Learning is image classification. In this scenario, a predictive model can be trained on a large group of general images. The weight of the model is used as a starting point when working on a smaller dataset, such as a chair or dog. The features already learned by working on the broader task will be helpful for the smaller task.
3. Online Learning
This is how Online Learning works. The machine uses the data that is available and uploads it when a prediction is required or after the last observation is made.
Online learning is useful when observations for a problem are provided over time and where the probability distribution of observation is expected to change with time. In order to harness and capture those changes, the Online Learning model changes itself as well.
The algorithms also use this approach when the numbers of observations are too much to fit in memory. The algorithm performed learning incrementally over observations such as a stream of data.
Online Learning reduces the element of regret in learning. By this, we mean people don’t have to compare the results they’ve got with the results they might have gotten if all the data was available as a batch.
Here is one good example of Online Learning, in Online Learning the stochastic gradient descent minimizes the generalization errors where examples of mini-batches are drawn from the stream of data.
4. Active Learning
In Active Learning, the model helps the human operator by giving queries, so the ambiguity is reduced from the learning process.
Active Learning is the form of Supervised Learning, and it helps to achieve the same or better results than Passive Supervised Learning. The way active learning works is by being very efficient about the data that is collected and used.
It’s not unreasonable to see Active Learning as something that solves Semi-Supervised Learning problems. Active Learning works best when there isn’t much data available and the data that is available is quite expensive to label and collect. The main thing about Active Learning is it minimizes the number of samples and maximizes the effectiveness of the model.
5. Ensemble Learning
As you can guess by the name Ensemble Learning, it’s about joining multiple parts together. Here is how it works. Multiple mods fit on the same data and predictions come from combining the answers of all mods.
Ensemble Learning provides better prediction than most models. It is able to do it because when multiple models are giving the same prediction their worth is better than just one model.
Ensemble Learning is good for improving predictive skills on a problem domain and reducing the stochastic learning algorithms. Some common examples of Ensemble Learning include bootstrap aggregation, weighted average, and stacked generalization.
1. Deductive Interface
This type of learning involves using general rules to determine specific outcomes. In contrast to Inductive Learning which we’ll cover soon, Deductive learning works from general to specific. It works from top to bottom. First, it’ll see the data from top (general perspective) and step-by-step it’ll move towards bottom (specific). This deduction method is pretty good for making predictions.
2. Transductive Interface
Transductive Learning is used for predicting specific outcomes by using a very specific set of data. It doesn’t involve any generalization, it uses specific examples directly to make predictions. It is perhaps the simplest method to make predictions.
A fantastic example of a Transductive Interface is the K-Nearest Neighbor algorithm. The algorithm does not model the training data, it uses the data directly each time it is making a prediction.
3. Inductive Learning
As we’ve mentioned in the above section, Inductive learning works from specific to general. It uses specific data to determine general outcomes.
Most machine learning models use an inductive interface or inductive reasoning where general rules are learned from historical examples. Fitting a machine learning model is also a process of induction. The model is the generalization of the specific examples in the training dataset.
The fundamental assumption of this type of learning is that the best hypothesis regarding unseen data is that it should fit the observed training data.
By reading this post you’ve become very familiar with what is machine learning, how it works, its various types, and some of the applications of it. Machine Learning is a growing area of Artificial Intelligence and as the algorithms used for Machine Learning grow more complex, the uses for it should move well beyond what we’ve shared here.
If you like what you’ve read and believe it’ll help people learn about machine learning, feel free to share it.