Artificial intelligence, also known as AI, is a form of intelligence created by humans. It takes a long time to build an AI, from training to polishing. For example, the most popular artificial intelligence, ChatGPT, took about 6 years to develop. Several stages are required to create a working intelligence, around 13 to be more precise. From data gathering to publishing, there is a lot to learn and comprehend. There are slight variations from one AI to another, but here are the main steps.
At the start, we have a problem definition. It is the most essential part of the process because it is where a human manually defines a problem or problems that needs to be solved. It puts the entire project in the right direction. It will help in later stages, such as data collection, another crucial step, and creating a solution to the problem. It can last from a few hours to a few months. It could be anything from detecting a spam email to identifying an object in a picture. It could also be translating languages or creating videos like Sora AI.
Secondly, AI developers need to gather various pieces of data and information from the World Wide Web. Depending on the artificial intelligence, engineers need to feed their creations raw text, such as: “Just finished a great workout! Feeling energized and ready for the day!” Or even pieces of coding, like: { “rawtext”: [{ “text”: “Hello everyone!” }] }. You can also give it audio files and various images of dogs, school objects, people, and buildings. There is also data cleaning, where you remove useless data and add more of what you want.
The next step is optional, mostly used when an AI is supervised. It is human annotation, it consists of manually labelling or classifying data to make a higher-quality data library; a ”place filled” with information for the AI. A library is a software used by an AI that the creator fills with information.
The next step is called Train, Validation and Test split. It is used to ensure that the created AI doesn’t just memorize but actually learns the given information. You take a set of information (a group) and divide it into three groups. In the first set, called the training set, you take anywhere from 70 to 80% of the data you collected and give it to the machine. The goal is to make the AI’s guesses and predictions more precise. Then we have validation, you decide when to stop training and choose the best version of the model. For the final part of this stage, the test split, you test your final model on brand new data to see if it’s good or not.
Finally, for part one, we have: model architecture design. It is a bit tricky, so I will do my best to make it simple. In this step, you choose what kind of architecture(a sort of structure) you want to use. The structure depends on the neural network. A neural network, simply put, is a ”machine” that learns patterns. There are many networks to choose from, like image, which learns to understand shapes and patterns, text, which obviously learns text sequence and meaning, and lastly sound, which comprehends audio waves(vibration of sound).
To recap, the first six steps are: Defining a problem, collecting data depending on the problem, removing useless data, making a high quality library, dividing data to train, tune and evaluate and choosing your architecture design.
