Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders

Abstract

Many approaches in generalized zero-shot learning rely on cross-modal mapping between the image feature space and the class embedding space. As labeled images are expensive, one direction is to augment the dataset by generating either images or image features. However, the former misses fine-grained details and the latter requires learning a mapping associated with class embeddings. In this work, we take feature generation one step further and propose a model where a shared latent space of image features and class embeddings is learned by modality-specific aligned variational autoencoders. This leaves us with the required discriminative information about the image and classes in the latent features, on which we train a softmax classifier.The key to our approach is that we align the distributions learned from images and from side-information to construct latent features that contain the essential multi-modal information associated with unseen classes. We evaluate our learned latent features on several benchmark datasets, i.e. CUB, SUN, AWA1 and AWA2, and establish a new state of the art on generalized zero-shot as well as on few-shot learning. Moreover, our results on ImageNet with various zero-shot splits show that our latent features generalize well in large-scale settings.

Generalised Zero-Shot Learning with Domain Classification in a Joint Semantic and Visual Space

“…We consider the simplified models above to describe CADA-VAE [4] and cycle-WGAN [3] as the latent space learning models.…”

“…Encoder x for x), then the decoder of a different modality is used (e.g.Decoder a from z x -see Fig. 2), which constraints the visual and semantic projections to be in the same region of the latent space represented by the mean µ and variance Σ of the samples produced by the encoder [4]. Figure 2: Depiction of the method CADA-VAE [4].…”

“…in (2). The training of CADA-VAE aims to produce a latent space that satisfies (4). In terms of the model architecture and hyper-parameters (e.g.…”

“…In terms of the model architecture and hyper-parameters (e.g. the number of epochs, batch size, the number of layers, learning rate, and, weight decay), we followed the specifications provided by [4]. The encoder for visual representation is parameterised with 1560 hidden neurons, and the encoder for the semantic representation is parameterised with 1450 hidden neurons.…”

“…The model is optimised with Adam for 100 epochs [20]. We use an adaptive scheduling rate for the hyper-parameters γ, δ, by (0.044, 0.0026), with respective epochs (21−75, 0−90) [4]. We also extended cycle-WGAN [3], as explained in Sec.…”