AI concepts

Before you dive into the world of deep learning, there are a few concepts we think you should have a grip on.

The most important concepts are listed below, but if you want more, we also have a huge glossary explaining even more concepts.

AI - ML - DL

AI, ML and DL are often used interchangeably, but they do not mean the same thing.

  • Artificial intelligence (AI) is the science and engineering of building intelligent machines. This includes many techniques, including traditional programming.

  • Machine learning (ML) is a subset of AI. Machine learning algorithms allow computers to solve problems using data as examples instead of coding an explicit set of rules, as in traditional software development.

  • Deep learning (DL) is a type of machine learning capable of working with complex, unstructured data like text or images. It also works great for many use cases based on structured tabular data. DL learns to both represent data and make predictions.
    Deep learning is what you do on the Peltarion Platform. Read more about what deep learning is here.

History of artificial intelligence
Figure 1. History of artificial intelligence. This graph also shows that DL is a subset of ML that is a subset of AI.

Data concepts

  • Datasets. All data on the platform are grouped into datasets.

The datasets can contain of different kinds of data:

  • Text data. For example: chat messages, tweets, customer feedback, lyrics, books, etc.

  • Tabular data. For example: sales data, customer info, sensor data, etc.

  • Image data. For example: photos, satellite images, heat maps, x-ray, etc.
    And Audio, which can be converted to spectrogram images.

This data is organized in:

  • Features. The various variables that are included in a dataset. Each feature has a feature encoding which determines the way its data is turned into numbers that a model can do calculations with. For example: Categorical, numeric, text, or image.

  • Subsets. The examples contained in a dataset are split into subsets on the platform. By default, one subset is used for training your model, and one is used to validate your model.

Problem types

The platform can help you solve the following problem types:

  • Classification. Put labels on your examples.

    • Single-label classification. When an example belongs to only one class and can have only one label. For example: Cat or dog.

    • Multi-label. When an example can belong to many classes and can have many labels. For example: A day can be both rainy and cloudy.

  • Regression. When you want to predict a number. For example: The revenue of a shop.

  • Similarity. Find similar images or pieces of text. For example: Similar images of shoes.

Experiments and models

  • Experiment. The basic hypothesis that you want to try, i.e., “I think I might get good accuracy if I train this model, on this data, in this way.” An experiment contains all the information needed to reproduce the experiment; dataset, model, and parameters used to run the experiment.

  • Model. A sequence of blocks that have been strung together.

  • Block. The basic components of a neural network and/or the actions that can be carried out on them.

    • The model starts with one or several Input blocks. This is where your data comes into the model.

    • At the end of the model, there is a Target block that represents the output that we are trying to learn with our model.

  • Loss function. The loss function is set in the Target block. It calculates how much to correct the model, during training, based on the difference between its current predictions and the target feature from your dataset.

Run your model

  • Run. When you’ve created your experiment and set up a model, you’re ready to run the experiment. That is, to start learning from the examples in the training subset and validate the results against the validation subset.

  • Training. When you run the experiment, you train the model.

Next step - Problem types

Was this page helpful?