The process of training a model through repeated exposure to images and metadata, so that it develops pattern recognition, and learns to improve automatically through experience
A type of algorithm that uses neural networks to learn how to categories. The output of a training process. When done and trained properly, allows you to input an image -> get a prediction of something about the contents of the image out
A type of 'cheat' in ML where the model is not exposed to a wide enough variety of images & data - it starts to remember the images / data set, rather than developing an understanding of the patterns, otherwise known as "generalising" to the problem
Another 'cheat', where the model detects a pattern that isn't related to the focal point of the image and uses that to generate its output, rather than the patterns. E.g. 'R' & 'L' in the COVID set
An AI system whose inputs and operations are not visible to the user / developer. It is generally difficult for data scientists, programmers and users to interpret or understand what and why something has happened.
If you pass an image in, and get a result out, the channels and pathways this result went through to reach this conclusion are not clear. Asking "why did you come to this conclusion", and trying to shed light on more of the internal processes, is often called 'unblackboxing'. In science treating something as a 'black box' typically means sticking in an input, getting an output, and rather stupidly/unresponsibly trusting the process that you don't understand in the middle
Training data is images and data that are shown to the model as examples of certain features of an image. During the training process the model looks at these example to understand the essence of the subject that it's learning. Validation data is then new images and data that the model hasn't been trained on. This is used to test the model to see if it truly has learned something.
The ability to identify the factors that influenced the machine learning model's decision and how the output was achieved; knowing how the ML's decision was made
Allows humans to explain and communicate whether models have been thoroughly tested for a sufficient and fair subject distribution and/or demographic
Comprehend the ML model and present the basis for decision-making in a way that we, as humans, can understand. ML models are a complicated network of connections, we need a way to interpret what's really going on.
Saying 'where' something is in an image. Annotating helps with telling the machine the features that is should and shouldn't focus on
Highlighting areas (or segments) within an image through 'colouring in' pixels and identifying these as a certain item / object. If you segment something, you are saying "this is an instance of this type of thing"
Take the segmented areas and apply a "class" to them - e.g. "elephant", "not elephant". Imagine someone asked you to paint all the parts of an image with an elephant on it white, and all the other parts black. The result would be a mask, where white = True and black = False.
A technology that supports viewing very large images in the browser by loading progressively more detailed portions of the image as the user zooms in. Was originally part of the now-deprecated Microsoft Silverlight platform, but now used in various applications which support viewing large images. Zegami has extended this approach to also allow zooming out, allowing many images to be viewed in parallel. We use the term 'Deepzoom' to refer to collections that use this technology, as opposed to 'Dynamic' ones. We've discussed the possibility of adjusting the term to reflect that our version does more than this.
A view option, and another way to display data. These are also known as flexible data cards, and are rendered when needed. Mainly used for Dynamic collections, combining image and meta data
A member of the Zegami team 😀