Typically, there are two main types of Artificial Intelligence (AI), in the process of their training and the technologies used: Traditional and Generative AI.
Artificial Intelligence. Overview
Traditional AI
Traditional AI, often referred to as narrow or weak AI, is focused on performing a specific task intelligently. It refers to systems that are designed to respond to a specific set of inputs. Such systems are capable of learning from data and making decisions or predictions based on it.
Imagine that you are playing computer chess. The computer knows all the rules, it can predict your moves and make its own based on a predetermined strategy. He does not invent new ways of playing chess, but chooses a strategy from those already programmed. This is traditional AI.
Traditional artificial intelligence is like a master strategist, capable of making intelligent decisions within a specific set of rules. Other examples of traditional AI are voice assistants such as Siri or Alexa, Netflix or Amazon recommendation systems, and Google’s search algorithm. They have been trained to follow certain rules, to do certain work, but they do not create anything new.
Generative AI
Generative AI can be considered the next generation of artificial intelligence. This is a form of AI that is capable of creating something new. Let’s say you have a friend who loves to tell stories. But instead of a human friend, you have artificial intelligence. You give the AI a starting line, say, “A long time ago in a galaxy far, far away…”, then the AI takes it and creates a whole space adventure story, with characters, plot twists, and a thrilling ending.
Unlike traditional AI, generative AI creates something new based on the information it is given. It can create original, creative content, be it text, images, music or computer code. Generative AI models are trained on a data set and, by learning underlying patterns, generate new data that replicates the training set.
A prime example of generative AI is GPT-4 [Wiki], the language prediction model of OpenAI [Wiki]. Trained on vast amounts of Internet data, it is capable of creating human-like texts that are virtually indistinguishable from texts that a human would write.
AI Comparison Summary Table
The main difference between traditional and generative AI lies in their capabilities and applications. Traditional AI systems are used primarily to analyze data and make predictions, while generative AI goes further and creates entirely new content.
| # |
Traditional AI |
Generative AI |
|
Main characteristics |
Specific tasks |
Ability to create new data |
|
Examines data and makes decisions or predictions |
Using raw data to create new, original content |
|
|
Works within a predefined set of rules |
Can generate text, images, music and code |
|
|
Teaching method |
Supervised learning |
Unsupervised learning |
|
Required to label training data |
Does not require labeled training data |
|
|
Restrictions |
Limited to specific tasks |
Requires large amounts of training data |
|
Inability to create new, original content |
Generated content may not be consistent or accurate enough |
|
|
Training may require a large amount of labeled data |
It is difficult to control the specifics of the generated content |
|
|
Examples |
AI playing chess |
|
|
Spam filters in email |
AI to convey DeepArt style |
|
|
Content creation (stories, art, music) |
||
|
Creation of fakes |
||
|
Search algorithms (Google) |
Personalized AI responses |
Artificial Intelligence Training Process
Traditional AI
- Data collection. The data set must be relevant to the task for which the AI is intended. For example, if you are training an AI to recognize spam emails, you will need a dataset of emails labeled as “spam” or “not spam.”
- Pre-processing of data. Data often needs to be cleaned and formatted before it can be used for training. This may include removing unnecessary information, handling missing data, or converting text data into a numeric format.
- Data separation. The data set is usually divided into a training set and a test set. The training set is used to train the model, and the test set is used to evaluate the performance of the model on unseen data.
- Model training. The model is trained based on the training data. In supervised learning, models show inputs and correct outputs. The model makes a prediction based on the input, and the difference between the prediction and the correct output (error) is calculated. The model parameters are then adjusted to minimize this error.
- Model evaluation. Once the model is trained, it is evaluated using test data. This gives an idea of how well the model can perform on unseen data in the future.
- Setup and optimization. Depending on the model’s performance on the test data, you may decide to go back and adjust the model parameters, use a different model, or collect more data.
Generative AI
- Data collection. As with traditional AI, the first step is collecting a data set. For GPT models, this dataset consists of a large corpus of texts. For example, GPT-4 was trained on hundreds of gigabytes of text from the Internet.
- Pre-processing of data. The data is then pre-processed. This involves cleaning the data and converting it into a format that the model can understand. For GPT, this includes tokenization, where text is broken down into smaller pieces or tokens.
- Model training. The model is trained using transformer-based language modeling. Models are shown a sequence of tokens and are asked to predict the next token in the sequence. For example, when entering “The cat was sitting on…” the model might be required to predict the missing word. The model makes a prediction and the difference between the prediction and the actual word is calculated. This difference (loss) is used to update the model’s weights and improve its predictions.
- Backpropagation and optimization. The loss is used to perform backpropagation, a process that calculates the gradient of the loss function relative to the model parameters. An optimization algorithm such as Adam or stochastic gradient descent is then used to adjust the parameters to minimize the loss.
- FRepetition. Steps 3 and 4 are repeated many times until the model’s performance no longer improves. Each pass through the training data is called an epoch.
- Fine tuning. After initial training, the model can be further tuned to a specific task. This involves further training the model on a smaller, task-specific dataset. For example, if you wanted to use GPT to generate medical text, you could further customize it on a dataset of medical articles.
