Learning a Deep Embedding Model for Zero-Shot Learning
Citations
2,496 citations
Cites background or methods from "Learning a Deep Embedding Model for..."
...This DNN is pre-trained on ILSVRC 2012 1K classification without fine-tuning, as in recent deep ZSL works [25, 30, 45]....
[...]
...We add weight decay (L2 regularisation) in FC1 & 2 as there is a hubness problem [45] in cross-modal mapping for ZSL which can be best solved by mapping the semantic feature vector to the visual feature space with regularisation....
[...]
...The 10 shallow models results are from [42] and the result of the state-of-the-art method DEM [45] is from the authors’ GitHub page1....
[...]
...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 [11, 3, 24, 45, 25, 31] learning....
[...]
...While on AwA1, our method is only outperformed by DEM [45]....
[...]
785 citations
Cites background from "Learning a Deep Embedding Model for..."
...larly, [35] argues that the visual feature space is more discriminative than the semantic space, thus it proposes an end-to-end deep embedding model which maps semantic...
[...]
783 citations
747 citations
570 citations
References
111,197 citations
"Learning a Deep Embedding Model for..." refers methods in this paper
...Adam [20] is used to optimise our model with a learning rate of 0....
[...]
73,978 citations
72,897 citations
"Learning a Deep Embedding Model for..." refers methods in this paper
...We omit the formulation of the bidirectional LSTM here and refer the readers to [15, 14] for details....
[...]
...We use the BasicLSTMCell in Tensorflow as our RNN cell and employ ReLU as activation function....
[...]
...The word embedding size and the number of LSTM unit are both 512....
[...]
...When the semantic representation was encoded from descriptions for the CUB dataset, a bidirectional LSTM encoding subnet is employed (see Fig....
[...]
...It is worth pointing out that this result was obtained using a word-CNN-RNN neural language model, whilst our model uses a bidirectional LSTM subnet, which is easier to train end-to-end with the rest of the network....
[...]
55,235 citations
49,639 citations
"Learning a Deep Embedding Model for..." refers background or methods in this paper
...Extensive experiments carried out on four benchmarks including AwA [22], CUB [45] and large scale ILSVRC 2010 and ILSVRC 2012 [6] show that our model beats all the stateof-the-art models presented to date, often by a large margin....
[...]
...However, most existing recognition models are based on supervised learning and require a large amount (at least 100s) of training samples to be collected and annotated for each object class to capture its intra-class appearance variations [6]....
[...]