Artificial Intelligence vs Machine Learning vs Deep Learning

Feature image

The above is a hot topic, yet at the same time, for some, an excruciatingly confusing one. It isn’t always a case of buzz words having precise meanings, and sometimes they may have an element of smudging in them – perhaps due to the overlap of the terms at play. Now, such is the case when it comes to Artificial Intelligence (AI), Machine Learning (ML), and of course, the seemingly newer one, Deep Learning, which ironically enough is not known as DL mostly, unlike its counterparts. Maybe, just maybe, it is too deep for it.

People think AI, ML, and Deep Learning are separate entities, and they may be justified in feeling that way. However, there is massive interlinking between the three. Knowing the differences between them is quite a necessity in this ever-technological world of today. To fully understand their differences and what some would want to know, which is superior if there is one, let us dive deep into what the three mean.

Here is a simplistic overview of exactly where the three stand next to one another.

AI vs ML vs DL

AI vs ML vs DL


Artificial Intelligence - AI

As can be seen above, AI is a superset containing Machine Learning, ML, and Deep Learning within. There is a good reason for that as well. AI is a concept and methodology of incorporating the level of intelligence a human being possesses into a machine. The above is evident through the term ‘Artificial,’ which stands for something not natural or something a human makes. The term ‘Intelligence’ stands for a level of thinking with which something can decide on its own after observing its surroundings and the things that are at play in terms of the events that have transpired and may transpire.

Some have even gone so far as to say that AI is a mechanism through which machines are trained and taught to mimic or impersonate an actual brain – its thinking capability, decision-making capability, and understanding capability. Well, these people aren’t wrong about that one bit, as we shall see once we delve into ML and Deep Learning, which happen to be subsets of AI, and their working and what they represent. Keeping all this in mind, it shouldn’t come as a surprise that AI focuses mainly on the skills of learning, reasoning and, of course, self-correction!

AI has three types as well:

  • ANI (Artificial Narrow Intelligence): This also has another name, right in line with ‘Narrow Intelligence’ - ‘Weak AI.’ It is pretty easy to decipher what these terms entail; simple, goal-oriented, non-complex, single task-based machine programming that includes things like chatbots, self-driving cars, Siri and so on. Now, machines that exhibit ANI do not have human sentiments, nor are they conscious to that degree. Instead, they use the data that is given to them only to perform their designated task.

  • AGI (Artificial General Intelligence): As you would expect, it also has another name, ‘Strong AI.’ Now here, as the name implies, machines exhibiting this form of AI will have a certain level of human intelligence. However, of course, such machines have not yet been created, at least not openly. However, such is the goal with AGI. Machines that would entail AGI would be able to think, understand, strategize, and perform various tasks simultaneously just as a human would, given situations involving problems that a human must solve and resolve each day. In achieving this, such machines would have to keep the context of what happened in the past, be innovative, think outside the box and come up with perhaps some unwarranted solutions!

  • ASI (Artificial Super Intelligence): which, well, is undoubtedly relatively superficial at this point. Machines employing ASI will have exceeded human capabilities in all aspects of intelligence. The above seems believable to many people, but they perhaps forget that humans make machines, not the other way around. This type of AI will pretty much deem us as human beings useless – yes, probably, and indeed not happening any time…ever?

The Fujitsu built K, one of the fastest supercomputers on the planet, right now took 40 minutes to simulate a neural activity that takes us, humans, 1 second. AGI itself seems pretty far-fetched as of now. Hence, ASI is quite definitely out of the picture, whether we like it or not.

AI Stages

AI Stages, source: https://twitter.com/niomaticinc/status/1061589358921289728


Machine Learning

Now, we come onto Machine Learning. For starters, it is a subset of AI. So, ML is the process or mechanism which enables a machine to learn automatically without human intervention and without being explicitly programmed to do so. It is expected to work in such a way so that as we keep feeding the machine with more and more data, its performance continues to get better and better. Now, albeit the fact that this does not happen all the time, yet it is the intent or desire of the developer to achieve such a working mechanism.

An ML model works based on the data provided to find links and dependencies of the output on the features that are at play and the relationship of the features amongst themselves. Just like AI, ML does not shy away from having types either, so, therefore, let’s have a look at them:

  • Supervised Learning:

    This is a form of ML where a machine trains itself based on data that has labels. A set of known input variables and an output variable/variables is given to the machine. Now since the machine has seen so many cases for which it knows the outputs, once it is given some new data, it can quickly work out the outcome based on its training quite accurately. The above also makes supervised learning the go-to option for tasks such as classification and regression.

    Some of its applications include search, computer vision, language detection and spam filtering, amongst a pool of others.

    • A few of its algorithms are listed below:
    • K-Nearest Neighbours
    • Logistic Regression
    • Linear and Polynomial regressions
    • Naive Bayes
    • Support Vector Machine
    • Decision Tree
  • Unsupervised Learning:

    This, as you certainly would have expected by now, is a form of ML where the machine trains itself on data that is not labeled. Here, the machine is responsible for inferring patterns in the data given to it, albeit that, as stated before, does not have known outputs. There is no human interference, and the machine has does this on its own. Therefore, usually, such a form of ML is used to find clusters within data and for dimensionality reduction. The programmer probably does not know what to see in the data. However, obviously must be some sort of a pattern being followed. For example, the size of different shirts can be clustered in small, medium, and large segments based on some relationship that a model may help infer. The inference involves exploring the similarities, differences, and patterns in data and, as such, partakes in exploratory data analysis for data preparation or insights in the early phases of an ML project.

    It is used in various applications and scenarios such as risk management, analysis of fake images, data segmentation, and anomaly detection, amongst many others.

    Here are some of its algorithms:

    • Singular Value Decomposition (SVD)
    • Principal Component Analysis (PCA)
    • Latent Dirichlet allocation (LDA)
    • Latent Semantic Analysis
    • FP (Frequent Pattern) Growth
    • K-means Clustering
    • DBSCAN (Density-Based Spatial Clustering of Applications with Noise)
    • Mean-Shift

    To find out more on exploratory data analysis in ML and more, refer to our 7 Steps of Machine Learning .

  • Reinforcement Learning:

    This is a type of Machine Learning that works on rewards for correctly performed actions and penalties for incorrectly performed ones. Here, an agent interacts with the environment in a trial and error fashion and continuously learns from it by utilizing the feedback it has gotten from its past actions and experiences. With such a form of learning, the fun part is that Reinforcement Learning can learn in noisy, dynamic, and unusual environments such as game worlds or the real world instead of the usual and expected static and immobile datasets. Using this mechanism, the machine tries to evaluate the optimal solutions to the given problem. This type of ML is usually found in robot navigations, skills acquisition and Game AI.

    Its applications are resource management, robot navigations, skills acquisition, Game AI and self-driving cars and aircrafts.

    A few of its algorithms include the following:

    • SARSA (State Action Reward State Action)
    • DQN (Deep-Network)
    • A3C (Asynchronous Advantage Actor-Critic)
    • Q-Learning (Quality Learning)
    • Genetic algorithm

Here is an excellent illustration of the various types of Machine Learning with their applications/use cases.

Machine learning classification

Machine learning classification


Deep Learning

Finally, we come to Deep Learning, a subset of Machine Learning, and therefore a subset of Artificial Intelligence. Being a subset of ML, Deep Learning can be considered a class of algorithms, mechanisms, or models of Machine Learning that take their inspiration from a human brain since they can have simple to highly complex ‘neural’ networks. Having these layers helps a machine perform computations that would not be possible using just the standard ML algorithms.

The layering allows a level of complexity to be exhibited by a model where it can reach an acceptable level of accuracy in things that were otherwise done only by human beings. For example, Deep Learning is the driving force behind automatically driven cars. A car can observe its surroundings so well that it knows when to stop by reading a stop sign, pedestrian, or even signal. Consequently, this requires the use of Computer Vision, an excellent advancement in the domain of AI. You can find much more in our First Steps with OpenCV for Python and Generating images with Deep Learning .

Here is a great yet simple depiction of what Deep Learning is.

Deep Learning heart

Deep Learning heart

Now, the way a Deep Learning algorithm works is quite simple to define, however much more challenging to grasp for a particular application once it is completed since there can be massive shifts in the way things work. The layers can get quite mind-boggling, to say the least.

Taking the example of an unstructured images dataset, the initial layers of the neural network will be incorporated to figure out the low-level features like edges and contours, whilst the deeper layers will be employed to bring out a whole representation of images by combining features from the previous layers. Such an approach, going from a relatively simple task to such a complicated task, with the ability to handle enormous datasets in comparison to the ML algorithms (albeit the fact that they are now catching up too through newer innovations, for example, through the random forests and gradient boosting built atop decision trees) causes the usage of Deep Learning to become a necessity.

So, therefore, the applications of Deep Learning include a vast majority of practical use cases, such as:

  • Robots and Self-Driving Cars
  • Natural Language Processing
  • Speech Recognition
  • Image Recognition
  • Portfolio Management
  • Drug Discovery
  • Prediction of Stock Price Movements
  • Language Translation
  • Virtual Assistants
  • Fraud Detection
  • Healthcare (Diagnosis of Diseases)
  • Demographic and Electron Predictions
  • Photo Descriptions

Now, the usage of Deep Learning is criticized at times by doctors and medical practitioners because they are challenging to understand and lack the level of Explainability and Causality. However, this is becoming a rather hot topic due to the advent of Explainable AI. As a result, the justification of Deep Learning in the medical domain is now losing its slack and gaining tautness.

Some of the top Deep Learning Algorithms and Models include the following:

  • Radial Basis Function Networks (RBFNs)
  • Multilayer Perceptrons (MLPs)
  • Self Organizing Maps (SOMs)
  • Deep Belief Networks (DBNs)
  • Restricted Boltzmann Machines (RBMs)
  • Autoencoders
  • Convolutional Neural Networks (CNNs)
  • Long Short-Term Memory Networks (LSTMs)
  • Recurrent Neural Networks (RNNs)
  • Generative Adversarial Networks (GANs)

Conclusion

What a long journey it has been then, has it not? Well, let us finish off by having a look at an immaculate illustration of the differences between AI, ML and DL.

They all relate to one another directly, and one cannot exist without the other. The advancement of Deep Learning means the advancement of Machine Learning since Deep Learning is the subset of ML, and the same can be said of the dependence of Artificial Intelligence on ML.

Hence, rather than arguing about which of the three is better, it would benefit the entire world to make each better. So, if things had to be epitomized in a tidbit fashion, then here it is – ML and Deep Learning are the mechanisms on top of which AI is built!

Thanks for reading!