Transfer learning using the example of image analysis

from | 2 April 2019 | Basics

Transfer learning is one of several learning methods used in the field of machine learning to equip algorithms with the skills they need for a specific task. The best known and most widespread learning methods are certainly Unsupervised Learning and Supervised Learning. This article gives an overview of the most important aspects of transfer learning.

Transfer learning is one of the Machine learning methods and has been in the course of the Upgrading image data analyses has become increasingly important in recent years. When you give a machine Image data it does not see a picture, but the data representing individual pixels. Whether it is a picture by Picasso, a photograph of a traffic light or a computer-generated image is irrelevant for a machine.

What is Transfer Learning?

If we first consider in a very simplified way only the beginning and the end of a Learning processthen in the case of image analysis it looks like this: The Data input consists of the data X images. The Output consists mostly of two or more Categories. These categories are presented in a Artificial neural network with two nodes in the output layer.

For example, certain images can be Categorised according to whether on them Cats or dogs are to be seen. In the case of security systems based on biometric analysis, the category is whether or not the data represents a particular face. In the case of Transfer Learning In principle, it is now a question of Learning processeswhich are necessary for the solution of these questions, to shorten. Instead of starting from scratch each time, a pre-trained model can be taken as a starting point. A model that already knows the basic structures of a face can be trained much more quickly to recognise a specific face. Thanks to transfer learning, a pre-trained model only needs to be trained to recognise certain features.

How Transfer Learning works

At the end of a learning process, in the case of Deep Learning-models a Output layer. This is virtually cut off during transfer learning. First of all, it must be said that a Machine LearningAlgorithm, who has been trained to recognise faces in pictures, sees pictures quite differently than we humans do. For him, images consist of individual data points that have certain values. For example, each pixel has information about its value on a greyscale or information about a colour value such as green, blue or yellow. At the next level of observation, smaller shapes and patterns may become visible, such as circles, lines or other geometric patterns.

For a model that is trained to recognise faces, a layer of nodes is interesting that provides information about which geometric patterns and colour information make up a face. In transfer learning, a model is "cut off" at this point, so to speak, and another learning process is continued at this interface. In face recognition, an algorithm thus does not have to learn from scratch what a face is in the first place.

Model Transfer Learning
Transfer learning illustrated: A new "prediction layer" is set up on the basis of the pre-trained layers (blue).

In cases like these, transfer learning makes sense

For one of our clients in the insurance industry, the challenge was, Damage casesthat occur after natural disasters, assessable. In this specific case, the task was to check roofs of houses for damage caused by wind. The task - limited to the analysis of the image data - was to recognise houses or their roof surfaces isolated from their surroundings and subsequently to distinguish between intact and damaged roofs.

Link tip: in this article we look at the Image recognition.

Transfer Learning offers the possibility of a model that is suitable for the Recognition of everyday objects (trees, tables, vehicles, cats, flowers, etc.) was developed and to use it on the basis of a relatively small number of new training images on recognising of damaged roofs. This is possible because roofs or damage to roofs essentially consist of edges, corners and colour contrasts - just like most everyday objects. A model that can already recognise such shapes therefore only needs to learn which configurations correspond to the respective shapes of intact or damaged houses.

The limits of transfer learning

In certain use cases, transfer learning is invaluable as it comes with an enormous Time saving is associated with it. However, it remains Not a direct substitute for high quality Data.

For example, for the example mentioned, it is of crucial importance whether the images with which the model was further trained were taken in a densely or sparsely populated landscape. Because in order to recognise the difference between "roof of a house" and "landscape / nature", the model must be able to recognise, for example, a flat roof, which is typical for a city. In rural regions, on the other hand, gable roofs are characteristic. If the model was further trained only on images of urban houses, fitting to rural houses would fail.

The potential of transfer learning

With Transfer Learning, the area of Machine Learning provides a method that allows certain Optimise processes in modelling. This can save an enormous amount of time, so that applications can be run economically accordingly. In particular, the challenge of obtaining labelled training data during training can be resource-intensive.

Transfer learning can of course not only be used in the field of image analysis. In particular also in the area of Language analysis and Language Comprehension Transfer Learning is an ideal way to shorten learning processes. The reason is simple: thanks to transfer learning, the general basics of the language do not have to be trained for every specific application. In this respect, transfer learning is representative of an ability that we humans have anyway: The Ability to transfer learned knowledge to other areas.



Our AT editorial team consists of various employees who prepare the corresponding blog articles with the greatest care and to the best of their knowledge and belief. Our experts from the respective fields regularly provide you with current contributions from the data science and AI sector. We hope you enjoy reading.

0 Kommentare

Submit a Comment

Your email address will not be published. Required fields are marked *