What problems have we already solved on the road to Artificial General Intelligence?
In the previous article in this series on Artificial General Intelligence (AGI), we learned about The Levels of AI. We covered Artificial Narrow Intelligence (ANI), Artificial General Intelligence (AGI), and Artificial Super Intelligence (ASI).
In this article, we’ll explore the problems in AI research that we’ve solved over the past few decades. This includes topics like neural networks, machine learning, and deep learning.
Since the first AI conference in 1956, there have been tremendous breakthroughs in AI. However, in this article, we’ll focus on only the most important concepts for our journey from ANI to AGI.
A neural network consists of a set of input neurons and a set of output neurons, with one or more hidden layers of neurons in between. The neurons are connected by a set of edges that propagate information from the input to the output of the neural network.
Data goes into the input layer of the neural network and a prediction (of some kind) comes out the other end. The input neurons take a set of numeric values as input. The hidden layers perform a bit of math. And, the output layer produces a prediction (represented as a probability).
Neural networks allow us to make decisions and predictions similar to an organic brain. Essentially, they provide us with an “artificial brain” for AI systems.
Machine Learning (ML) is a type of AI that learns how to solve problems without being explicitly programmed to do so. It does this by detecting statistical patterns that exist in data.
With machine learning, we use existing data and a training algorithm to learn a model of the data. We can then feed new data into that model that it has never seen before, and it will make predictions about the new data.
For example, if we want to create an image recognition model, we label a set of images based on their contents. We feed this labeled data into an ML training algorithm. The algorithm learns a model of the data. Then, we provide the model with a new image as input, and it provides us with a label indicating the type of object and the probability the image contains that object.
An ML model is essentially a function. It is simply a mapping from an input to an output. In this case, it takes data as input and produces a prediction as output. ML allows AI to learn on its own.
A Deep Neural Network (DNN) is a neural network with more than one hidden layer. Adding more hidden layers allows the network to model progressively more complex functions.
For example, imagine we want to teach a DNN how to detect human faces. First, we would feed a set of labeled images of human faces into the input layer of the DNN. Then we use a technique called backpropagation with gradient descent to teach the model what each person’s face look like.
The first layer of the DNN would learn to detect horizontal, vertical, and diagonal lines. The middle hidden layers would learn to detect facial features like eyes, noses, and mouths. The later hidden layers would learn to detect the general pattern for entire faces. And, the output layer would detect the most abstract representation of a person (i.e., the name of the person being recognized).
With Deep Learning (DL), we can use complex data as input and produce complex data as output. This includes text, images, audio, video, and more. DL provides AI with the ability to perceive its environment and make predictions about future (possible) states of its environment.
Neural Networks, ML, and DL are the fundamental building blocks of modern AI. However, they have significant challenges that must be overcome to create AGI. We’ll cover these problems in the next article in this series: AI of Today.