AI Model Families

What is an AI Model Family?

A model family is a group of related machine-learning (ML) models. Their commonality: They share the same underlying structure and learning principle. Each individual model then is adapted to different tasks through different objectives and outputs.

Models within a family do

• have similar training logic,
• share same core structure or learning principle,
• differ mainly in...
  1. the task they solve (classification, regression, etc.)
  2. the loss / objective function
  3. the type of output

Think of it as one algorithmic idea, adapted to different problem types.

Examples of Model Families

There is many variants of model families, the following ones are most important:

• Embedding Models (EM): represent data (text, audio, images or other dataformats) as vectors / lists of numbers. They have a dimension property which is the amount of numbers the model outputs in one embedding run. Embeddings can be used to compare objects semantically, so by meaning. For example you can calculate the distance between the words "Good" and "Nice" and notice that it's a shorter distance then "Bad" to "Nice". The same can be done for files and other data too.
• Decision Trees (DT): same tree structure, splitting logic, pruning; different leaf output, loss function, evaluation metrics e.g. DecisionTreeClassifier, DecisionTreeRegressor
• Linear Models (LM): same linear combination of features; different link function, loss, output interpretation e.g. Linear regression, Logistic Regression, Poisson Regression
• Support Vector Machines (SVM): same margin maximization, kernel trick; different loss, outputs e.g. SVC, SVR
• Tree Ensembles: e.g. RandomForestClassifier, GradientBoostingRegressor
• Recommendation Systems: various approaches (collaborative filtering, content-based, matrix factorization) that share the goal of ranking items by relevance to users
• Neural Networks (NN): same layered neural architecture, backpropagation; different loss functions, output heads e.g. CNN, RNN/LSTM, MLPRegressor, MLPClassifier
• (Generative) Transformer Models (as in GPT), an instance of a concrete model type within this family would be large language models (LLMs).

As you see these Model Families greatly vary in their complexity and goals, architectures.