Written by Allison Mejia | 2020
What Is Considered AI and What Is Not?
In science fiction, AI is often portrayed in a physical form. They can be portrayed as robots performing multiple human-like tasks such as having a natural conversation with a human or rebelling against the human race. This type of AI would be considered as “strong AI”.
Strong AI has yet to be achieved in 2020. When you hear AI, this is what we expect it to be; an intelligence that thinks and acts like a human. As humans, we may not have the computing power of an ordinary computer, but we are able to accomplish tasks that are highly difficult to replicate by machines. Thanks to our vision, we are able to see and quickly analyze our environment, recognize people or objects and even interpret human emotions.
Thus, given the complexity of human nature, it is no surprise that weak AI is what we can only achieve to this day. Most AI’s are highly specialized programs that can accomplish a task that is human-like. In video games, this includes NPC programs; a non-player character in a game which is not controlled by a player. NPCs are interesting to look at because they are extremely simple programs yet are categorized as AI alongside more complex algorithms. NPC software is simply programmed with pre-defined states. When an event occurs, the NPC will change from one state to the other. For anyone who has programmed before, you could think of it as changing a state attribute when a “if” statement is triggered. Pretty basic but because the NPCs provide an experience that feels like intelligence, it is still labelled as AI. However, the AI that most companies rave about has to be machine learning.
Machine Learning
Machine learning is a set of mathematical-heavy algorithms that simulate the learning process that a normal human goes through. They often use data to finetune parameters and make decisions.
One popular method to do machine learning is to feed tons and tons of examples labelled to an AI until it recognizes patterns and refines a mathematical formula so that when it receives new unseen scenarios, it is able to classify them with the correct label. This process is called supervised learning. An example of such an application is image recognition. Suppose we want to make an AI capable of identifying humans from non-humans. You can use pictures of your family as input to your AI. A “human” label is associated with each image, so the AI will start seeing patterns in the pixels of each image (such as recognizing where eyes are usually located on a human face). Then it will fine-tune a mathematical formula that takes as input a series of pixels and outputs a classification label. You will also want to feed examples labeled as “non-human”. To test your AI, you’ll have to give it an image of someone new and the AI will classify the image as either “human” or “non-human”. The correct identification of the person will depend on the extent of your training data set, i.e,. The amount and diversity of the images that are used to train the algorithm.
Choosing the training data is no trivial matter and directly affects the performance of the AI. There have been controversies in the field of human face recognition where bias to certain skin types could be observed. Face recognition softwares from big companies such as IBM and Microsoft were tested by the Gender Shades project to measure their ability to properly identify 1270 images of very diversified individuals. The results of the study showed that the AI programs would perform better on lighter skin type people than on darker skin [2]. Thus supervised learning can be greatly influenced by human input. Although supervised learning is a widely used method of machine learning, other branches of machine learning exist. For example unsupervised learning is perfect in cases where we want the AI to learn about labels on its own, without human input. This type of AI is often exploited in marketing to discover types of customers among other things. Another branch of AI known as reinforcement learning takes a reward system approach in order to teach a desired behavior.
While decision or prediction making is a common application of machine learning, there is another interesting way of using machine learning programs: generate new data. These algorithms can create pictures, music, text and so on. These generative programs usually use artificial neural networks, a set of algorithms that mimic the inner working of the neural network in a human brain. Neural networks are also behind the most promising branch of machine learning known as deep learning. This is the technology behind the historical defeat of world GO champion Lee Sedol by AlphaGo, an AI capable of playing GO, a game believed to have an infinite number of moves. Deep learning allows a machine to learn exclusively from data without the need to for a human to label data into categories.
A Look Back (Old Fashioned AI)
The venue of machine learning was an important milestone for the field of AI, although one might ask how AI programs were achieved prior to this. And the answer still resides in mathematics. But a different kind.
Let’s take a look at Deep Blue, a chess-playing computer developed by IBM that was able to defeat the world champion of chess in 1997. Deep Blue managed to defeat Russian Garry Kasparov with pure computational power. The AI behind this machine used a simple but effective set of algorithms in order to calculate the optimal moves during a game. To find the optimal solution, the AI created a list of all possible game configurations in the form of a tree. A tree in computer science is a data representation with a root node connected to multiple children. And these children can be connected to their own set of children, and the children of the children can also have their own. In short, the tree can grow exponentially. This was a very good representation for a chess game, because one move generates several other possible moves as the game progresses.The goal of Deep Blue was to search the tree for all possible configurations and evaluate a score for each branch of the tree before selecting the branch with the optimal score. For a long time, this was what AI programs could do best: search a tree of possibilities and output the best outcome.
With machine learning now, it is possible for AIs to make decent predictions without the need to go through the painstakingly tedious process of calculating every single combination of outcomes.
AI is evolving at an interesting speed and who knows when we will reach a true strong AI, capable of mirroring humans more accurately. At the moment, we are beginning to see the rise of many applications of AI in a wide variety of fields. Because of the ability of an AI to make predictions based on patterns, many are seeing opportunities to have prediction tools enhancing decision making processes such as an application for buying stocks on the stock market, detecting cancerous cells from medical images and so on. Many people are concerned about the place that AI will occupy in our society, perhaps even to replace people in similar occupations. But as AI is not an infallible tool and can still make mistakes, it is often necessary to have a second human input. Thus, as a society, we should aim for symbiotic coexistence with AI. As for finding ways to implement strong AI, we could eventually achieve it given the speed at which our technology has progressed.But before we do so we need to ask ourselves how far we want to go with these advanced technologies.
