The Dev Process

The Framework

When we develop Artificial Intelligence, we first have to decide how it will become intelligent in the first place. Remember that we defined intelligence as the ability of something to reason, learn, solve, perceive, and communicate. Of all these components, learning is the most difficult to implement. As we have said before, humans learn in many different ways, such as observation, trial and error, and pattern recognition. That last bit serves as the basis for machine learning, which, as the name implies, is a method of making machines learn. It helps computers learn on their own.

After that, we then determine the dataset the AI will use to train itself. The dataset must be carefully selected because, as we have discussed previously, an AI will acquire any bias in the dataset used for training. Finally, we have to test the AI one way or another. The testing method will heavily depend on the machine learning algorithm used to train the AI.

Supervision

A machine learning algorithm is a method of implementing machine learning. There are two general types of machine learning algorithms. One of them is known as supervised algorithms. Simply put, supervised algorithms have a "correct answer" on paper. Say, if we give an AI trained with a supervised algorithm a number x, there is an expected value y for the AI to spit out. AIs trained using this algorithm are best suited to solve classification and regression (i.e., estimating the relationship between two things) problems.

The other general type of machine learning algorithm is known as unsupervised algorithms. In simple terms, unsupervised algorithms do not have a "correct answer" on paper. If we give an AI trained with this type of algorithm x, it can spit out y, z, or even a greek letter. AIs trained using this algorithm are best suited to solve clustering (i.e., spotting groups in a pool of data) and association (i.e., finding connections between two things) problems.

Hand-picking Data

After creating our machine learning algorithm, we must find and prepare the AI's training dataset. Again, remember that an AI will acquire any bias in the dataset used for training. Therefore, choose your data wisely and make sure that:

1. The dataset you will choose aligns with the purpose of the AI.
2. The dataset you will take is replicable to some extent (to ensure that testing is possible).
3. The dataset you will select is unbiased (unless you intend to simulate human decision-making).

Preprocessing

After finding a dataset to work with, you must process the data before use (i.e., preprocess the data). The way you will process the data will depend on its function in training the AI. Let's say your AI only needs to look into the data as percentages of the range (i.e., the difference between the maximum and minimum values). In other words, 0% is the lowest value, 100% is the highest value, and 50% is halfway between the two. Here, you must perform a preprocessing technique known as min-max scaling, which expresses all the values in the dataset as a percentage of the range.

We also have to remember that, at this stage, our AI can only understand numerical data; it cannot comprehend words yet. Therefore, if we want the AI to analyze quantitative (descriptive) data, we must somehow convert the descriptions to numbers before passing it to the AI. This process is known as label encoding. Most of the time, this process occurs in two steps: alphabetization of descriptions (or labels), followed by assignment of numbers to each item.

Testing and Deployment

Once we have trained our AI, we then have to test it. Again, the testing methods we will use depend on the machine learning algorithm used by the AI:

• For AIs based on supervised algorithms, we must first find another dataset similar to the AI's training dataset. Then, we shall pass some of the values in the new dataset through the system and see if the results make sense.
• Meanwhile, for AIs based on unsupervised algorithms, we must get more creative in testing them as they do not have an expected output. The most straightforward way to test these AIs is the Turing test.

After testing, the AI may need further preparations depending on its deployment environment. For example, an AI built into a robot must learn to control the robot's parts first before deployment. AIs inside artificial environments must learn to navigate the artificial environment first (i.e., know where all the input and output fields are) before roll-out. Once those further preparations are complete, the AI can finally function in its environment.