edfas.org 21 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 2 (a) Classification EXPERIENCE Depending on how machine learning algorithms experience a dataset, they can roughly be split into three types: supervised, unsupervised, and reinforcement learning (Fig. 6). The dataset is a collection of multiple examples, where every example is a collection of features, whose values are quantitatively measured from some artifact, e.g., pixels of an image taken for a sample, a text description of findings, values measured by a curve tracer, or a sequence of tasks executed for a job. Supervised algorithms experience datasets comprising examples associated with one or more labels. Such algorithms use training examples to create a model able to assign a correct label(s) to any previously unseen example. Depending on a learning task, different label types can be used to annotate an example. For instance, one can annotate an image as a whole, only its part, or even assign a label to each pixel of an image, as depicted in Fig. 7a, b, and c, respectively. In the case of unsupervised algorithms, examples are not labeled and, therefore, these algorithms focus on learning the probability distribution that generated the dataset or properties of this distribution. For instance, clustering methods often aim to learn properties that can be used to group given examples similar with respect to these properties. For example, a well-known K-means algorithm[31] estimates k means (centroids) of the underlying distribution and divides examples into k clusters by assigning each of them to the closest mean according to the Euclidian distance. Note, both experience types discussed above do not have a formal definition allowing one to distinctly assign an algorithm to one of them. Using conditional probability and chain rule we can represent an unsupervised learning problem by n supervised ones and vice versa. Self-supervised approaches can construct labeled examples from unlabeled ones, e.g., some words in a given text can be hidden and an algorithm is asked to predict them. In semi-supervised learning some of the examples have labels and some do not, whereas in multi-instance learning a set of examples is labeled, but not members of the set. Finally, reinforcement learning algorithms do not experience a fixed dataset. Instead, these algorithms can interact with their environment and collect the experience resulting in this interaction. In most cases, reinforcement learning is applied to learn policies allowing an agent to make decisions in different situations occurring in the environment. Applications of these methods in FA are a topic of current research focusing on optimizing lab or tool operations. For instance, fast scheduling methods based on reinforcement learning developed for production,[32] can be applied to schedule resources of a lab. PERFORMANCE In practice, machine learning algorithms experience a dataset with a limited number of examples. As a result, it is often split into Fig. 6 Machine learning tasks. (b) Object detection (c) Segmentation Fig. 7 Different label types of an example.
RkJQdWJsaXNoZXIy MTMyMzg5NA==