China
Africa
Explore chapters and articles related to this topic
The Future
Published in James Luke, David Porter, Padmanabhan Santhanam, Beyond Algorithms, 2022
James Luke, David Porter, Padmanabhan Santhanam
Zero-shot learning (ZSL) is about creating ML algorithms that need no labelled training data for the task. Consider language translation. The idea is that you give the textual input in the source language (say, English), and the output from the application is in the target language (say, Spanish) and vice versa. Languages are complicated concepts with a lot of underlying structure, ambiguities, nuances, etc. Scope of your application can be something basic, such as assistance with elementary phrases for Londoners visiting Madrid for a few days, or a lot more demanding (i.e. translating newspaper articles, automatic translation of English language books, etc.). Typically to train the system, depending on the scope, you need to give examples of sentences in English and the corresponding proper translations in Spanish. This is what is called “labelled data” or “the ground truth”. Bigger the scope, more data you need. Clearly, the same labelled data can be used to build a model in both directions, i.e. English to Spanish and Spanish to English.
Computer Vision Methodologies for Automated Processing of Camera Trap Data
Published in Yuhong He, Qihao Weng, High Spatial Resolution Remote Sensing, 2018
Joshua Seltzer, Michael Guerzhoy, Monika Havelka
One concerning issue with this methodology is that endangered and elusive species will be underrepresented in the data sets, and thus the networks will be less tuned to recognize their features. This is known as one-shot or few-shot learning, wherein the target categories (which are to be matched with labels) are represented by sparse data—or even zero-shot learning, where only a description of a category, but no example of images from the category, is available. This might often be the case in ecological studies, where species lacking any preexisting photographic data might be uniquely recognized by a set of features a network can recognize. One attempt to eliminate biases in the data set by Hoffman et al. (2014) focused on adapting the CNN's internal architecture to better identify categories with sparse representation, of which endangered species are a prime example. They were able to increase performance on a subset of ImageNet from 66% accuracy to 77% using these techniques. Evidently, while identifying underrepresented species is a significant challenge, further research in network architectures could mitigate the problem.
Generating visual representations for zero-shot learning via adversarial learning and variational autoencoders
Published in International Journal of General Systems, 2023
Therefore, in recent years Zero-Shot Learning (ZSL) has gained much attention, where the main task is to classify an image for which no labeled training data are available during training (Akata et al. 2013; Lampert, Nickisch, and Harmeling 2009). The task of ZSL is to classify images from unseen classes. The information about how unseen classes are semantically related to seen classes relies on side information. This side information is based on semantic attributes (Akata et al. 2013; Farhadi et al. 2009; WordNet Miller 1995) and word2vec (Mikolov et al. 2013). ZSL methods are evaluated in two different settings i.e. conventional ZSL (Akata et al. 2015; Xian et al. 2018a, 2018b) and Generalized ZSL (GZSL) (Chao et al. 2016; Gulrajani et al.2017; Xian et al. 2019). In conventional ZSL, the assumption is that test samples come solely from unseen classes as the search space is restricted to unseen classes only. On the other hand, in GZSL, both seen and unseen classes are disjoint, but the test examples can be from seen or unseen classes. This paper focuses on both settings. In recent works by Chao et al. (2016); Gulrajani et al. (2017); Xian et al. (2019); and Schonfeld et al. (2019) it is observed that the classification accuracies of ZSL decrease significantly in the generalized setting because it is biased toward seen classes (Xian et al. 2018a).
Semiotically adaptive cognition: toward the realization of remotely-operated service robots for the new normal symbiotic society
Published in Advanced Robotics, 2021
Tadahiro Taniguchi, Lotfi El Hafi, Yoshinobu Hagiwara, Akira Taniguchi, Nobutaka Shimada, Takanobu Nishiura
For rapid installation, service robots that are newly introduced into the local environment should be capable of inferring categories or classes that appear only in small numbers in the training data. Hence, few-shot or even zero-shot learning is expected. Few-shot learning is an area related to transfer learning that addresses models that can predict categories or classes that appear only in a small number of training data [34]. Zero-shot learning addresses models that can predict classes that have never appeared in the training data [35].