AI vs ML vs Deep Learning

By | December 5, 2020
AI vs ML vs Deep Learning

Introduction

The world is slowly depending more on data science, mainly because of a machine’s capabilities to think and act like a human. To enhance its skills and reach the level of a human is still a far away dream for a machine; however, we humans are leaving no stone unturned to help machines get there.

AI, ML and deep learning (DL) are three technologies that are inter-related and use the machine’s knowledge to solve various business problems. While AI is the bigger picture, ML and DL fall under its umbrella. Further, DL is a specialized form of ML. Here is how the three are interrelated:

Vamware

AI, ML and deep learning

AI has been around since 1956. It has come a long way since then; however, the goal remains the same: get a computer to perform intelligent tasks that only humans were capable of doing earlier. The simplest example is a game of chess with a computer. AI is just a term that comes up because human intelligence is considered natural, but since intelligence is induced in machines, it is considered artificial. AI has many techniques and systems like expert systems, rule-based systems etc. that solve the problems. The most common technique or category of AI is machine learning.

Machine learning not only mimics human behaviour but also mimics how humans learn. Machine learning has many algorithms which are fed with data, and the algorithm trains through the data or by itself. The machine creates a generic model, the accuracy of which is improved through multiple iterations until the desired accuracy is achieved.

Deep learning is a subset of machine learning that uses neural networks with multiple hidden layers to solve complex problems. For example, if we have to identify the image of a dog, it doesn’t matter whether the dog is in a crowd of other animals, or it is in a different position or pose, the machine will recognize the image. Chatbots work on the same principle – the machine can understand the response that should be given based on a user’s question.

Definitions

With the above knowledge, let us try to define each term in short words, to understand the core difference.

  • Artificial Intelligence (AI): AI is the intelligence shown by machines or ‘intelligent agents’ so that machines can mimic human behaviour and thoughts. A machine that can learn and solve problems is said to be artificially intelligent.
  • Machine Learning (ML): ML is the most common type or category of AI where machines can predict outcomes of various business problems with accuracy without being programmed explicitly. ML uses a lot of past data to build, train and test a model that does the predictions.
  • Deep Learning (DL): Deep learning, also called neural learning, is a subset of ML which deals with artificial neural networks with representation learning. Just like machine learning, deep learning can be supervised, unsupervised and semi-supervised.

Categories of AI

Let us now dive into details of each of the terms before we do head to head comparison.

Since the human brain is complex, it is difficult to understand, unlike what the scientists thought initially. They were of the impression that to digitize the human brain wouldn’t be so hard! However, they understood that the key to imitating human behaviour was learning, along with natural language processing and creativity. Well, we are still far from imitating the human brain, but there are improvements every day. There are three categories of AI:

  • Weak AI (Artificial Narrow Intelligence or ANI): Weak AI is used in many fields of science, business and healthcare. There are many computer games like Chess and AlphaGo, that are an example of weak AI. Weak AI systems have powerful analytical tools to analyze, recommend and provide business insights. They are useful for finding trends and determining patterns.
  • General AI (Artificial General Intelligence or AGI): This is the next step for making machines more human-like. Machines can make their own decisions and can learn from past mistakes. These systems also have some emotions that are produced by programs that respond to particular stimuli. Chatbots and virtual assistants are examples of machines performing human-like interactions. Even robots are being taught on how to respond (with emotions) in certain situations.
  • Strong AI (Artificial Super Intelligence or ASI): Well, this is a state where machines will completely overpower humans – they will be smarter, more creative and have better social skills. This does sound a bit dangerous, and with smartphones having more and more features, humans are moving towards that very direction, there is still a long way for this to happen completely. Robotics and Machine reasoning and learning are two important methods of creating super AI.

The achievements we have so far are mostly with Weak and General AI, and the future goal is ASI.

Applications of AI

Some of the most popular applications of AI are:

  • Chatbots: AI-powered chatbots have been helping in solving common user queries, thereby reducing the load on human support.
  • eCommerce: Through AI-powered systems, users can search for an item by giving its image, rather than typing it on Google search box.
  • Healthcare: By gathering data of patients, AI systems can give a better diagnosis and suggest appropriate treatment for them. This will help reduce the burden on doctors and other medical staff.
  • Logistics and supply chain: Businesses can mine customer data and make improvements in the overall process from order placement to fulfilment. Also, the delivery time is less with autonomous trucks and robots picking the items from their source to destination, thus making the process go on 24×7.
  • Sports: Through AI technologies, betting can be done with more accuracy, as huge amounts of data can be analyzed and processed in a jiffy, and the future outcome can be predicted.

Components of Machine learning

Machine learning, as we have learnt, is a branch of AI that deals with algorithms that can find patterns and make predictions. Machines can learn and train themselves with the data being fed. Machine learning is of three types:

  • Supervised learning: In this learning type, humans act as a supervisor (teacher), and feed the machine with training data, along with the correct outputs (labels). The computer learns and finds patterns based on this data. Based on this learning, new inputs are given to the computer for which the output is unknown, and the computer tries to predict the outcome. Supervised learning is of two types – regression and classification.
  • Semi-supervised learning: Semi-supervised learning falls between supervised and unsupervised, where there are labels for some of the data, but not for all. The cost for labelling is practically quite high, so for model building, semi-supervised methods work out to be more efficient.
  • Unsupervised learning: In this type, the computer trains itself with no labelled data. If humans are unable to decipher what to infer from a dataset, unsupervised learning algorithms can be applied to get some interesting patterns and results. These algorithms can mine the data, detect patterns, and group the data, to derive useful insights and predictions. Clustering and association rules are some popular types of unsupervised learning algorithms.
  • Reinforcement learning: This the most advanced type of learning where machines and agents (the learning algorithm) collectively and automatically determine the ideal behaviour based on a particular state (or condition) to maximize performance. If the performance is maximized, rewards are given through feedback. The agent observes the input state, makes decisions and receives the reward or punishment based on the action performed. Some common algorithms are Q-learning, Temporal difference. Most computer games are based on the concept of reinforcement learning.

ML algorithms:

Several algorithms are used to solve machine learning problems of different types. We will not get into the details of the algorithms, but only mention the most popular algorithms and the use cases where the algorithms are used. Details of algorithms will be covered separately.

  1. Linear regression: use cases like whether a patient has a particular disease or not based on a set of symptoms, whether a promotion material would be suitable for a particular customer or not based on their past preferences, analyzing the type of campaign that works best for a customer, ice-cream sales based on weather, etc.
  2. Naïve Bayes: Text document classification, spam filtering, classify news articles into categories
  3. SVM: handwriting recognition, classification of genes, speech recognition, categorizing facial expressions, Geo-spatial data analysis, cancer diagnosis
  4. Decision tree/Random forest: which medicine to recommend to a patient, which insurance plan is most suitable for a client, whether to play a game of cricket or not
  5. k-nearest neighbours: economic forecasting, genetics, data compression, predicting house rent, recommender systems
  6. Neural networks: stock exchange prediction, handwriting recognition, image classification, autonomous systems like self-driving cars (computer vision), flight control, NLP
  7. Apriori: market-basket analysis, healthcare systems, flood area prediction, analysis of patient’s records
  8. k-means clustering: image segmentation, image compression, document clustering, market segmentation, fraud detection, analyzing student result data.
  9. Q learning: traffic light control, robotics, recommendation systems, supply chain management, stock market trading, electric power systems security

Applications of ML:

There are many applications of machine learning; some of the most important ones are:

  • Making traffic predictions and giving real-time traffic updates, suggesting alternate routes to a user
  • Voice assistants like Alexa, Siri, Cortana that use Natural language processing techniques
  • Social media services like face tagging, face recognition, sentiment analysis
  • Spam detection and malware filtering
  • Online fraud detection
  • medical diagnosis
  • Search Engine Optimization to offer the best search results based on the context
  • Recommendation systems to recommend the right products and services based on user’s preferences
  • Image/face recognition, image compression, text recognition and analysis
  • Video surveillance, detection of unusual activity, providing alerts, crowd analysis, violence detection

More about deep learning

Deep learning is a specialized case of machine learning, which tries to mimic the human brain to solve complex problems. It consists of multi-layered neural networks, the same like a human brain. The level of abstraction is gradually increased through the non-linear transformation of input data. The multiple neural network layers are hidden, i.e. what happens inside is unknown to us.

deep learning

The hidden layer need not be a single layer but is a series of connected channels, through which information is relayed. These channels are called weighted channels. Neural networks have a feedback system, wherein the weighted sum of inputs is applied to an activation function, and the result decides which neurons are to be activated and passed to the next layer. This goes on until the second last layer (before the output layer) is reached.

Neural networks need a huge amount of data to train. Deep learning is more accurate than machine learning and is closer to artificial intelligence.

How deep learning network works:

In deep learning, both supervised or unsupervised techniques can be used for training. Let us take an example of predicting the value of a house, using supervised deep learning technique. The price will depend on various factors like – square feet measurement, age of the house, floor, furnishings, etc.. All these inputs are called neurons. These are passed to the first hidden layer.

The most important input will have the maximum weight, i.e. the weights decide the importance of the input value. In our case, the square feet area of the house is the most important factor in determining the price.

The number of hidden layers should be decided optimally – the more the number of hidden layers, the more time it takes for training. The number of hidden layers can be decided only with trial and error, and there is no formula for it, although it does depend on the number of input and output neurons.

Each neuron (input) has an activation function, which defines the output based on the input or set of inputs.

For a supervised learning approach, we give the data for the input neurons and then compare the results with the outputs from the dataset. The first time, since the system is untrained, we will get incorrect results. For this, we create a cost function that determines the deviation of the received output from the correct output. Our target is to make the cost function as a minimum (almost zero) as possible. This needs many iterations and hence high computational power.

Deep learning applications

There are several interesting applications of deep learning in various industries:

  • Personalized content display for a particular user on their login, on sites like Amazon, Netflix etc.
  • Automatic gameplay
  • Handwriting generation
  • Pixel restoration – detecting a face with a low-resolution image
  • Describing photos using the English language
  • Translating images into text of particular language by reading the image
  • Image colourization
  • Automatic sorting of images based on location, event date, group of people, faces etc.

Differences in a nutshell:

Now that we know about AI, ML and DL and how each of them works, we can compare them head-to-head. This will help us appreciate the differences in a handy manner.

AI ML DL
Great performance on huge datasets Performance is high with medium or small datasets Good performance on huge datasets
A system that encapsulates machine learning and deep learning encapsulates deep learning a specialized type of machine learning
Requires high power systems built for handling AI workloads, like high-power GPU, RAM etc. Doesn’t require high computational systems – CPU, RAM with good performance is sufficient. high-computing power is required, however, if images are not involved, no need for GPU, having a high capacity CPU should suffice
It takes time to understand the problem, apply the solution and obtain an accurate result Takes a few minutes or hours to solve the problem Can take up to weeks to compute the weights
Many algorithms are available and are easily understood You can choose from multiple algorithms to solve a problem only a few algorithms are available

Summary

Through this article, we have understood that AI, ML and DL are closely related, and while AI is the more generalized form, DL is specialized and comes under the AI umbrella. Deep learning is said to be the next level of machine learning. Where ML involves some human intervention, deep learning doesn’t require even that much. For example, for a machine learning problem, the features used for input are determined by humans, whereas deep learning decides the features on its own and weighs them accordingly. To summarize, AI encompasses both ML and DL, and ML further encompasses DL.

You might also be interested in:

Leave a Reply

Your email address will not be published. Required fields are marked *