AI, Machine Learning, Deep Learning — What are the differences?

Source: Deep Learning on Medium

BEGINNER GUIDE

AI, Machine Learning, Deep Learning — What are the differences?

To whom new to data science or confused about definition of Artificial Intelligence, Machine Learning, Deep Learning.

Artificial Intelligence

AI is the abbreviation for Artificial Intelligence. Throughout the history of AI, the researchers has divided into 4 distinct approaches:

  • Acting humanly
  • Thinking humanly
  • Thinking rationally
  • Acting rationally

In particular, telling computers mimic human behaviors is the most difficult challenge and this is the goal that scientists are aiming at. Moreover, researchers developing AI also target to deducible ability of the computer with the following features:

  • Reasoning: The ability to solve problems in a logical way.
  • Knowledge: The ability to digest information in a context (like analyzing how many objects, events, obvious situations in real world and classifying based on the characteristics of corresponding objects).
  • Planning: The ability to create and pursue a goal based on the knowledge.
  • Communication: The ability to understand human written and verbal communication.
  • Perception: The ability to understand the context from images, audios and other kinds of input.

Turing Test

Picture by SearchEnterpriseAI.

A famous mathematician named Alan Turing, in 1950, has proposed an approach to creating AI acting like human. The proposal was known as the “Turing Test”.

According to Turing, the purpose was examining how smart the system acted. The test were performed in discussion-like way: one person asks a question, another person answers the question and the computer is expected to respond to the question. During the discussion, if the questioner cannot differentiate whether the answers come from human or machine, the machine can be considered “smart”.

This approach has been bringing remarkable achievements:

  • Natural Language Processing: The computer is able to understand text or voice and communicate with human
  • Knowledge Representation: The ability to memorize knowledge via vision, audition or document.
  • Automated Reasoning: The machine is capable if using its knowledge to answer a question or make a good decision.
  • Machine Learning: The computer is able to adapting to a context, then conclude new principles from acquired knowledge and use them for decision making purpose.
  • Computer Vision: The capability of observing and determining the objects.
  • Robotics: The machine is able to interact with nearby objects as well as moving in an area.

Cognitive Science

Thinking like human is another approach. Researchers will need a mega library to work on this field, because, to “teach” a computer thinking in human way, we first understand how humans think.

There are 3 ways to examine:

  • Observe the thinking process
  • Observe a person’s actions
  • Observe the activity of brain

If we give the computer a human similar input and receive a human similar output, we may conclude it is thinking like humans at some point. There is belief that if we are able to combine the results from Turing test and cognitive science, we can achieve the AI goal faster.

Logic

However, it’s not easy, even very hard, to let the computer think in human way so we can narrow the problem to: thinking rationally is enough. Thus, the logicism at that time could play its role.

In 19th century, the logicians invented mathematical notations in order to describe a statement or relationship among objects in the world, which were useful in training computer later. However, this approach was a rough path. The inventors has faced 2 challenges:

  • First, the computer is incapable of digesting informal knowledge and expressing it in the formal terms, especially the knowledge was not assured of 100% certainty.
  • Second, there is a big difference between solving a problem in theory and solving it in the real world. We need a large set of data as the input, which can exhaust the computer resources of any computers, even the input is a problem with few hundreds facts.

Agent

And finally, a robot, a proper approach to acts. I usually call agent as robot, “agent” comes from Latin word “agere”, means to do.

Obviously, computer programs must do something, but computer agents are expected to satisfy more requirements. They must operate autonomously, collect information from the surrounding environment, survive in a long time period, adapt to any changes, make plans and try to achieve them.

A typical example is the NASA robots, which were launched to planets with unpleasant condition, received few or no commands from human. How can they collect the soil, prolong in the stormy, windy climate, make its own plan for charging the solar battery in order to continue the mission?

Machine Learning (ML) and Deep Learning (DL)

And, ML is just a part of acts-like-human approach, and this approach is just one of 4 ways to research AI.

To reach the ML goal, the researchers has published many algorithms as well as different ways to deal with problems:

  • Supervised learning: decision tree, k-NN, Naive-Bayes, SVM, neural network, deep learning, …
  • Unsupervised learning: k-means, hierarchical clustering
  • Reinforcement learning: passive/active/generalization

Once again, we can see DL is one the methodologies to solve problems using supervised learning of ML. So why does DL make such a big storm in technology industry? Everyone, every organization, every engineers put a lot of effort into researching Deep learning.

Strength: Deep learning is a technique to implement ML in order to achieve state-of-the-art results of computer vision and speech recognition. Deep neural networks perform effectively with images, audios and text data, capable of updating the model with new data through batch propagation. The architecture of the model (quantity and structure of layers) can be applied to different problems, particularly, the hidden layers play a role in reduce the cost of feature engineering.

Weakness: Deep learning is not a Swiss army knife, it needs a large dataset of certain area to be trained. Actually, deep learning has been defeated by Tree ensembles method in some basic machine learning tasks. In addition, training a deep learning model is expensive, it requires powerful resources as well as high quality specialists who perform tuning hyperparameters (number of layers, number of nodes per layer, learning rate).

From my point of view

The deep learning trend

The reason why people are focusing crazily in DL, because of its efficiency in mining images data or digital signal, which is up to 95% accurate. Moreover, that is a way to raise fund, when it’s hot trend, investors tend to play gamble on those DL companies. Besides that, to apply DL, there are a lot of tasks to be done like data preprocessing, finding a good number of nodes, layers, designing CNN, RNN, LSTM model, using various kind of activate functions to compare efficiency, … Throughout the process, it results in enormous numbers of DL model and can be used for solving different problems.

But there are disadvantages

It does not mean DL is the key to unlock all issues. In some performance experiments, DL cannot beat some common algorithms like Random forest, SVM or Monte Carlo estimation. We also face issue when using DL with categorized data or text data.

In Fintech, credit scoring in particular, DL doesn’t show how it “thinks” because it’s a black box method, we only know the input and output, nothing else. Let’s imagine a customer credit card application is rejected, we don’t know why the heck the computer made that decision and the bank clerks will have hard time to explain the reason. In other hands, training DL needs huge amount of specific data, and for credit scoring, besides basic information like name, address, … there is no more information to let the DL model understand the context. Also, keep in mind, there are people in specific roles are allow to access credit check data, and in this case, we have to hand over it to engineers which may result in privacy violation.

And Skynet conspiracy

There are countless number of people suppose that AI, at some level of training, it will take control of human being like in Terminator, i-Robot, … Or recently, the fear of job loss from AI. In my opinion, it is absolutely impossible:

  • Machines perform more efficiently than human, but not smart. They only work under human command, by that way, we are free from boring repetitive tasks. The old kind of jobs will be replaced by the new one.
  • The decisions given by machines are not 100% precise. And humans always do the last step, testing then making final decision.
  • Machines help humans in creative works but they cannot do such works. Some typical applications are: Deep Dream by Google, Brain Storming by Adobe, some mobile app generators, Wix web generators, …
  • They need more flexible mathematical system or language in order to simulate human thoughts, perception or actions. It is not different to create a new human species (new math system, new language), which costs billion years.
  • We, humans, always think about back up plans for everything. There are manual ways in any cases. That’s why the Esc button was born.

And there may not be related, but I still would like to say: instead of worrying about AI domination, we should worry about our privacy has been used every day. There is no robots control humans but some humans control other humans.

In Big Hero 6, a famous cartoon about a robot named Baymax, he can chat with us, scan the body and tell about our illness, then he gather new knowledge to perform treatment. He can distinguish right-wrong or dangerous actions and help human avoid injury. What I would like to say here is, Baymax, although just a fictional character, but he is a helpful AI product which we should target to, instead of imagining about the apocalypse made by those robots. Actually, the past catastrophes mostly were created by humans, no robots involved in those events.

My closed words

A long article, but I hope my readers have a clear picture about the definition of AI, ML and DL. If you want to step into AI world for your career, ask yourself which approach you aim at (logical thinking or actions) and the level of human-like (just enough or near-human). Those AI systems require a lot of effort to complete but also high risk of failure. However, if they are successful in the mission of serving human kind, you have spent worthy time and effort. Good luck!