An embarrassingly simple approach to zero-shot learning

By Bernardino Romera-Paredes, Philip H. S. Torr. 2015

1,342 citations recorded by the source corpus.

Zero-shot learning consists in learning how to recognise new concepts by just having a description of them. Many sophisticated approaches have been proposed to address the challenges this problem comprises. In this paper we describe a zero-shot learning approach that can be implemented in just one line of code, yet it is able to outperform state of the art approaches on standard datasets. The approach is based on a more general framework which models the relationships between features, attributes, and classes as a two linear layers network, where the weights of the top layer are not learned but are given by the environment. We further provide a learning bound on the generalisation error of this kind of approaches, by casting them as domain adaptation methods. In experiments carried out on three standard real datasets, we found that our approach is able to perform significantly better than the state of art on all of them, obtaining a ratio of improvement up to 17%.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Radiology, Nuclear Medicine and Imaging, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Radiology, Nuclear Medicine and Imaging

In Representation & Self-Supervised Learning

Citations in the corpus

Massively Multilingual Neural Machine Translation in the Wild: Findings and Challenges (2019, background): “Other problems closely related to multitask learning include zero or few-shot learning (Lake et al., 2011; Romera-Paredes and Torr, 2015; Vinyals et al., 2016; Pan and Yang, 2010), meta-learning and life-long learning (Thrun and Mitchell, 1995; Silver et al.”
Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing (2021, background): “This is traditionally called the zero-shot setting [111], as there is zero training data for the task of interest.”
Learning to Compare: Relation Network for Few-Shot Learning (2017, background): “Motivated by the failure of conventional deep learning methods to work well on one or few examples per class, and inspired by the few- and zero-shot learning ability of humans, there has been a recent resurgence of interest in machine one/few-shot [8, 39, 32, 18, 20, 10, 27, 36, 29] and zero-shot [1”
Novel Class Discovery without Forgetting (2022, background): “Zero-shot learning methods [54,42] require prior knowledge of extra semantic attribute information about the unlabeled classes.”
Subspace Regularizers for Few-Shot Class Incremental Learning (2021, background): “…information, a number of previous studies in vision and language have also investigated the effectiveness of directly learning a mapping from the space of semantic embeddings to the space of class weights (Das & Lee, 2019; Socher et al., 2013; Pourpanah et al., 2020; Romera-Paredes & Torr, 2015).”
Zero-Shot Learning—A Comprehensive Evaluation of the Good, the Bad and the Ugly (2017, methodology): “ESZSL [10] applies a square loss to the ranking formulation and adds the following implicit regularization term to the unregularized risk minimization formulation:”
Zero-Shot Learning — The Good, the Bad and the Ugly (2017, methodology): “ESZSL has some variance because its al-gorithm randomly picks a validation set during each run, which leads to different hyperparameters.”
Latent Embeddings for Zero-Shot Classification (2016, methodology): “The main idea of structured embedding frameworks [1, 2, 12, 32] is to first represent both the images and the classes in some multi-dimensional vector spaces.”
An Empirical Study and Analysis of Generalized Zero-Shot Learning for Object Recognition in the Wild (2016, methodology): “Existing research on ZSL has been almost entirely focusing on this setting [4,10,11,12,13,14,15,8,16,17,18,19,20,21,22,23,24,25,26,27,28].”
Feature Generating Networks for Zero-Shot Learning (2017, methodology): “It can also be deployed with various classiﬁers, e.g. ALE, SJE, DEVISE, LATEM, ESZSL that constitute the state of the art for ZSL but also the GZSL accuracy improvements obtained with softmax is important as it is a simple classiﬁer that could not be used for GZSL before this work.”
Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders (2018, methodology): “On the other hand, the classic ZSL methods DeViSE [6], SJE [2], ALE [1], EZSL [24] and LATEM [34] use a linear compatibility function or other similarity metrics to compare embedded visual and semantic features; CMT [27] and LATEM [34] utilize multiple neural networks to learn a non-linear embedding”
Exploring the Challenges towards Lifelong Fact Learning (2018, methodology): “dding architecture proposed in [7] as our backbone architecture to compare the evaluated methods. We chose this architecture due its superior performance compared to other joint-embedding models like [9,13,24] and its competitive performance to multi-class cross-entropy. The main difference between ”
Generalized Zero-Shot Learning via Synthesized Examples (2017, methodology): “, using a class-attribute based similarity measures), a number of ZSL methods have been proposed that are based on representing the parameters representing each unseen class as a similarity-weighted combination of the parameters representing the seen classes [22, 24, 4].”
Zero-Shot Recognition via Semantic Embeddings and Knowledge Graphs (2018, background): “The core idea is to represent each category with learned vector representations that can be mapped to visual classifiers [48, 44, 13, 41, 25, 15, 14, 18, 23, 4, 5, 55, 54].”
Zero-Shot Learning via Semantic Similarity Embedding (2015, methodology): “[30] proposed a simple method by introducing a better regularizer.”
Generating Visual Representations for Zero-Shot Classification (2017, methodology): “ 66.7 50.1 -/- Changpinyo et al. [8] - 72.9 54.7 90.0/62.8* Net GoogLe [36] Xian et al. [41] - 71.9 45.5 - Ours. 55.34 77.12 60.10 85.50/56.41 Lampert et al. [22] 38.16 57.23 - 72.00/- Romera-Paredes [32] 24.22 75.32 - 82.10/- Zhang et al. [43] 46.23 76.33 30.41 82.50/- Zhang et al. [44] 50.35 80.46”