Authored by Abhinav K. & Team
Can you ever stop thinking? Well, thinking of nothing is basically thinking of something. Hundreds of thoughts go through our heads every day. Even our dreams are representations of our feelings and emotions. Converting all this information into data and utilizing it is a huge feat for humanity and a big leap into the future. This is the gist of AI (Artificial Intelligence).
Contrary to the popular belief, AI started to gain popularity in the 1950s. In 1956, John McCarthy defined AI as ‘The science and engineering of making intelligent machines’. In other words, AI is the development of computer systems that are capable of having human-like intelligence. This computer system should be capable of decision-making, solving complex problems, etc.
From the first computer program that could learn to play checkers created in 1951, to today’s self-driving cars, the development of AI has branched multifold, its tentacles reaching every corner of the world. Today, almost every form of technology used by any person is embedded with AI.
Because AI is an integral part of most people’s lives, it is imperative to understand it. To understand how AI is so versatile, we have to look at the different types of AI. The AI that is used in playing games is not the same as what is used in self-driving cars.
As stated earlier, AI has evolved and branched over the last 70-80 years - sort of like an evolution. There are two types of AI:
Based on Capability
Based on Functionality
We will be looking at a few types of AI and talking in a bit of detail about what it is, and what it is used for. You will also realize how these types of AI interconnect with each other. Let us first take a look at AI based on capabilities.
Based on Capability
Artificial Narrow Intelligence (ANI) also known as Weak AI is the type of AI that exists in our world today. ANI systems can perform in real-time, but they cannot perform outside of what they are programmed for. If a system is programmed to analyze raw data it can only do that. It cannot perform tasks outside of analyzing data.
Everything from Siri to website chatbots are all ANI because they perform a specific task and can only perform what they are programmed for. Siri is programmed to answer questions if it has the answers in its database. Chatbots are also programmed to answer questions related to the website they are hosted on. They cannot answer questions related to topics that are outside their database.
Artificial General Intelligence (AGI) also known as strong AI refers to machines that are on par with human intelligence. AGI can perform any task at the level of human intelligence. These types of machines are conscious, have sentience, and are driven by emotions. These are the typical machines you see in sci-fi movies.
The AI we have today can analyze data at an unimaginable speed. But humans are able to think strategically, and are able to tap into our emotions to aid in the thinking process. In other words, we have a conscience, whereas computers do not. This is what makes human intelligence more superior to computer intelligence. It is difficult to replicate human conscience because even we don’t completely understand what conscience is, let alone replicate it.
Artificial Super Intelligence (ASI) refers to the machines that surpass human intelligence. This means that - machines will be exhibiting intelligence that is beyond even the smartest of humans. Hence, researchers are afraid for the same reason. People are concerned that ASI can overtake humanity and become the next conqueror of earth. ASI might be able to evolve into beings that can advance at an even greater rate than humans can, ultimately leading to a new species. Humanity has a long way to go before reaching this stage.
We briefly talked about AI-based on capabilities. But what about AI-based on functionality?
Based on Functionality
This is the most basic type of AI. This type of AI does not have the ability to form memories or use past experiences to learn from. This type of AI reacts to the current circumstances. For example, playing a chess game requires one to know the rules and thinking several moves ahead of the opponent. This is what a reactive machine does. In chess, the AI can assess the current situation and make predictions of its next moves and its opponents.
We can see that Reactive machines can only function under specific circumstances. They cannot form memories or learn from past experiences. We can classify this machine as an ANI because reactive machines have a narrow range of capabilities. They can only function in the range of what they were programmed for. For example, IBM’s Deep Blue machine can only play chess and cannot function outside the “chess realm”. To learn more about how IBM’s Deep Blue machine defeated grandmaster, Garry Kasparov, watch this video:
At first glance, reactive machines may not pose much usefulness in today’s world. However, reactive machines are being used extensively in many fields. For example, software development uses reactive machines to test and implement a specific software. Since reactive machines focus on a specific task - implementing, testing, and deploying can all be done independently. This improves both output and workflow.
Unlike reactive machines, this type II class of AI can store past experiences or data, for a very short time. Hence, they can use stored data for a limited time. Because this type of AI can use data from past experiences, it is able to make predictions of what might happen based on current circumstances. To have a better understanding of how this AI makes predictions, let’s take a self-driving car as an example. While on the road, the car stores information such as distance between vehicles, traffic signals, stop signs, lanes, etc. The AI in the car can then utilize these stored data and make a prediction about what to do next.
But how does it use the data to make an accurate prediction? There are a few ways to teach a computer to make accurate predictions. Let’s go over one method of “teaching” machines.
Like a toddler learning how to walk, the robot will find different ways or methods to achieve its goal. It will get feedback on how successful these methods are. Depending on the feedback, it will find the most optimal and successful method of achieving its goal. One practical example of reinforcement learning in industrial automation.
Google was able to reduce its energy consumption in its data centers because of reinforcement learning. AI found the best possible way to reduce the energy consumed. Google DeepMind created DQN (Deep Q-networks). DQN plays games and finds the best possible way to play by using reinforcement learning. DQN store’s different players’ experiences and replays them to provide more diverse training data. DQN played multiple Atari games, and the results were astonishing
The chart below shows the results of the performance of DQN agents after playing multiple Atari games. These DQN agents did not have any previous knowledge of the rules of the games before the play. To learn more about DQN, read this article.
This is where we differentiate today’s AI from future AI. This type of AI will not only have representations of the world, but also about other beings. The understanding of thoughts and emotions of others can affect how AI acts. Solving problems and working together as a team required one to understand the motives and emotions of others. In other words, this AI should have a Theory of Mind. What is the theory of mind?
Theory of Mind : Theory of mind is a skill that allows you to think about mental states, both your own and others. Theory of mind is not only about understanding mental states, but it is also about being able to form predictions on emotion based on body language, surroundings, what the other person is talking, etc. For example, if your friend is talking in a very depressing tone, you could probably tell that he/she is having a bad time. If AI is able to do this, robots can have social interactions and can be on par with human intelligence. They would be able to read the emotions of different people and respond in the correct manner.
As this type of AI is still under research and development, we can only fathom its real-world applications. Having a theory of mind means understanding other’s emotions and being able to handle social interactions. Hence, AI can then replace professions that require social interactions on a daily basis such as - clerks, interviewers, hotel staff, etc. This can save companies millions of dollars as they would not have to pay these robots.
The last step of AI development is to build machines that have consciousness. This is an extension of the Theory of Mind. “I want burgers” is very different from “I know I want those burgers”. In the second example, we can tell that the machine has a conscience because it knows what it wants. In the first example, it is just executing code based on the environment it is in. If a burger is necessary, the machine will get a burger.
Researchers do not fully understand conscience, let alone creating one. We are far from building self-aware machines. That is why it is important to do what can be done in this present moment. Researchers need to fully understand how to enable machines to learn from past experiences. This alone is a huge step towards human intelligence.
AI is a technology that will be in every aspect of our day to day lives very soon. In the present, AI is spread across many fields, and is even being used by most of the people around the world. Because AI is becoming so integrated, it is imperative to acknowledge and understand it in detail; and if possible, speak its language. Today’s artificial intelligence is just the start of something great. Who knows, humans and AI may even co-exist as two different species!