• Corporate Communications

Understanding Artificial Intelligence and Machine Learning

What is AI?

Artificial intelligence (AI) refers to the studies of computer and robot development that may both emulate and exceed human intelligence. While there have been many definitions of AI coined throughout the previous few decades, computer and cognitive scientist John McCarthy, whom many consider a towering figure in the relative field over his time at Stanford, expressed that AI is the engineering and science behind the creation of intelligent machines, particularly intelligent computer programmes. It is akin to the same challenge of utilising computers to comprehend human intellect, but AI does not have to limit itself to biologically observable ways, as mentioned in his 2007 work titled “What is Artificial Intelligence”.

In its most basic form, AI integrates large datasets and computer science to solve problems. It includes the subfields of machine learning (ML) and deep learning (DL), which are referenced in the context of AI. Algorithms are used in these areas to develop expert systems that make classifications or predictions based on input data. AI-enabled applications may contextualise and evaluate data to trigger activities or offer information without the need for human intervention.

The classification of AI has 4 distinct types, which are roughly based on Maslov's hierarchy of needs, with the simplest level requiring only basic functioning versus the highest level requiring all-seeing and all-knowing awareness, in which two of these have been achieved and the other remain theoretical. The four types are reactive machines, limited memory, theory of mind and self-awareness.

4 Types of AI

  1. Reactive machines, being the simplest level of AI, can conduct simple operations like producing output in response to some type of input given. This is the first stage of any AI system, a basic reactive machine that takes a human face as input and produces a box around the face to recognise it as a face. At this level, there is no ‘learning’ occurring as the system is trained to execute a certain job or task and will not deviates from that goal. They are reactive machines that cannot retain inputs, function outside of a certain context, or grow over time. IBM’s chess-playing supercomputer Deep Blue and Google’s AlphaGo are some examples of reactive machines present.

  2. Limited memory is the most often utilised type of AI nowadays, which through monitoring behaviours or data, learns from the past and gains experiential knowledge. This type of AI makes predictions and performs sophisticated categorisation of tasks by combining historical, observational data with pre-programmed knowledge. Autonomous vehicles, for example, employ limited memory AI to observe the speed and direction of other cars, allowing them to “read the road” and modify as needed. This knowledge and interpretation of incoming data keep them safer on the road. However, as the name implies, is still restricted in the sense that the data that autonomous vehicles use is transient.

  3. The concept that individuals, creatures, and things in the environment can possess emotions and thoughts that impact their behaviour is known as the "theory of mind". Machines will be able to obtain full decision-making abilities comparable to humans with this sort of AI. Machines with the theory of mind AI will be able to perceive and retain emotions, and then modify their behaviour in response to those feelings when they interact with humans. However, due to the process of adjusting behaviour based on fast fluctuating emotions being so dynamic in human interaction, there are still a lot of obstacles to developing the theory of mind AI.

  4. The last stage of AI development is self-awareness, which is to create systems capable of forming depictions of themselves. In certain ways, this is an extension of the third level's theory of mind AI. Having the capability to differentiate between wanting and needing an item is an ability of self-awareness that constitutes consciousness. Conscious beings are mindful of themselves, aware of their thoughts and feelings, and can foresee the reactions of others. While the development of a fully self-aware AI is still a long journey away, developers are concentrating their efforts on comprehending memory, learning, and the capacity to make judgments based on prior encounters.

What is ML?

Machine learning is a subset of AI and computer science that utilises data and algorithms to mimic how people learn, progressively enhancing its accuracy. Over the previous several decades, breakthroughs in processing and storage capacity have supported several creative machine learning-based technologies that include algorithms that learn by using their historical data. ML is currently utilised in various everyday applications such as in online recommender systems for Google search algorithms, Netflix recommendations, email spam filters, Facebook’s auto friend-tagging and even autonomous vehicles.

ML is a major element of the expanding discipline in data science. Using statistical approaches, algorithms are trained to produce categorisation or projections, and also to find critical insights in data mining operations. These insights then influence decision-making within software and enterprises, ideally influencing key growth indicators. As big data expands and grows, so will the market growth for data scientists.

Here, we will be looking at the three primary types of ML algorithms, supervised learning, unsupervised learning and reinforcement learning.

3 different types of ML algorithms

These many forms of ML, like the various types of AI, encompass a variety of intricacies. While there are various types of ML algorithms, the majority are a mix of — or built on — these three basic categories.

  1. The most basic of these is supervised learning, which occurs when an AI is monitored during the learning process. Data scientists or researchers will feed the machine a large amount of data to analyse and learn from, in addition to some sample results of what that data should create, often known as “inputs” and “desired outputs”. Supervised learning produces an entity that can anticipate outcomes based on fresh input data. The machine's learning may be refined further by storing and constantly re-analysing these estimations, boosting its accuracy throughout time. Image identification, media recommendation engines, predictive modelling, and spam filtering are examples of supervised ML applications. Subsequently, neural networks, naïve Bayes classifiers, logistic regression, random forest, linear regression and support vector machine (SVM) are some of the approaches used in supervised learning.

  2. Unsupervised learning, also known as unsupervised ML, analyses and clusters unstructured datasets using ML techniques. Without the use of human interaction, these algorithms uncover hidden trends or data classifications. Because of its capacity to detect similarities and contrasts in data, this approach is perfect for consumer segmentation, exploratory data analysis, cross-selling tactics, and picture and pattern recognition. It is also used to decrease the number of elements in a model via the dimensionality reduction process. Two popular methodologies are principal component analysis (PCA) and singular value decomposition (SVD). Other unsupervised learning algorithms include k-means clustering and probabilistic clustering approaches. Unsupervised ML, like supervised ML, may develop and evolve.

  3. Reinforcement learning is the most complicated of all three algorithms because no data set is given to help the machine train. Alternatively, the algorithm learns through interaction with the setting in which it is situated. The AI is put in a game-like environment in reinforcement learning where it uses trial and error to find a solution to the issue. To ensure that the computer accomplishes what the programmer intends, the AI is rewarded or punished for the choices it takes, as its purpose is to increase overall return. Even though the developer establishes a reward policy in the sense of it being the game rule, there are no tips or recommendations provided to the model on how to complete the game. It is up to the model to discover ways to accomplish the job to maximise the reward, beginning with completely random trials and progressing to complex tactics and superhuman abilities. Reinforcement learning is presently the most effective technique to hint at machine creativity by utilising the potential of search and repeated trials. Unlike humans, they can benefit and learn from thousands of simultaneous gameplays if a reinforcement learning algorithm is performed on sufficiently adequate computer infrastructure.