Over the last decade Deep Neural Networks have produced unprecedented performance on a number of tasks, given sufficient data. They have been demonstrated in variety of domains (Gu et al., 2015) spanning from image classification (Krizhevsky et al., 2012; He et al., 2015a, b, 2016; Huang et al., 2016), machine translation (Wu et al., 2016)2015), speech recognition (Hinton et al., 2012a), and synthesis (Wang et al., 2017), learning from human play (Clark & Storkey, 2015)2015; Silver et al., 2016; Foerster et al., 2016; Van Hasselt et al., 2016; Gu et al., 2016) among others. In all cases, very large datasets have been utilized, or in the case of reinforcement learning, extensive play. In many realistic settings we need to achieve goals with limited datasets; in those case deep neural networks seem to fall short, overfitting on the training set and producing poor generalisation on the test set.
Techniques have been developed over the years to help combat overfitting such as dropout (Hinton et al., 2012b)2015), batch renormalisation (Ioffe, 2017) or layer normalization (Ba et al., 2016). However in low data regimes, even these techniques fall short, since the the flexibility of the network is so high. These methods are not able to capitalise on known input invariances that might form good prior knowledge for informing the parameter learning.
It is also possible to generate more data from existing data by applying various transformations (Krizhevsky et al., 2012)
to the original dataset. These transformations include random translations, rotations and flips as well as addition of Gaussian noise. Such methods capitalize on transformations that we know should not affect the class. This technique seems to be vital, not only for the low-data cases but for any size of dataset, in fact even models trained on some of the largest datasets such as Imagenet(Deng et al., 2009) can benefit from this practice.
Typical data augmentation techniques use a very limited set of known invariances that are easy to invoke. In this paper we recognize that we can learn a model of a much larger invariance space, through training a form of conditional generative adversarial network (GAN) in a different domain, typically called the source domain. This can then be applied in the low-data domain of interest, the target domain. We show that such a Data Augmentation Generative Adversarial Network (DAGAN) enables effective neural network training even in low-data target domains. As the DAGAN does not depend on the classes themselves, it captures the cross-class transformations, moving data-points to other points of equivalent class. As a result it can be applied to novel unseen classes. This is illustrated in Figure 2.
We also demonstrate that a learnt DAGAN can be used to substantially and efficiently improve Matching Networks (Vinyals et al., 2016) for few-shot target domains. It does this by augmenting the data in matching networks and related models (Vinyals et al., 2016; Snell et al., 2017) with the most relevant comparator points for each class generated from the DAGAN. This relates to the idea of tangent distances (Simard et al., 1991), but involves targeting the distances between manifolds defined by a learnt DAGAN. See Figure 2.
In this paper we train a DAGAN and then evaluate its performance on low-data target domains using (a) standard stochastic-gradient neural network training, and (b) specific one-shot meta-learning111By meta-learning, we mean methods that learn a particular approach from a source domain and then use that approach in a different target domain. methods. We use 3 datasets, the Omniglot dataset, the EMNIST dataset and the more complex VGG-Face dataset. The DAGAN trained from Omniglot was used for the unseen Omniglot terget domain, but was also for the EMNIST domain to demonstrate benefit even when transferring between substantially different source and target domains. The VGG-Face dataset provide a substantially more challenging test. We present generated samples which are of good quality, and also show substantial improvement on classification tasks via vanilla networks and matching networks.
The primary contributions of this paper are:
Using a novel Generative Adversarial Network to learn a representation and process for a data augmentation.
Demonstrate realistic data-augmentation samples from a single novel data point.
The application of DAGAN to augment a standard classifier in the low-data regime, demonstrating significant improvements in the generalization performance on all tasks.
The application of DAGAN in the meta-learning space, showing performance better than all previous general meta-learning models. The results showed performance beyond state-of-the art by 0.5% in the Omniglot case and 1.8% in the EMNIST case.
Efficient one shot augmentation of matching networks by learning a network to generate only the most salient augmentation examples for any given test case.
To our knowledge, this is the first paper to demonstrate state-of-the-art performance on meta-learning via novel data augmentation strategies.
Transfer Learning and Dataset Shift:
the one shot learning problem is a particular case of dataset shift, summarised by the graphical model in Figure 2. The term dataset shift (Storkey, 2009) generalises the concept of covariate shift (Shimodaira, 2000; Storkey & Sugiyama, 2007; Sugiyama & Müller, 2005)
to multiple cases of changes between domains. For one shot learning there is an extreme shift in the class distributions - the two distributions share no support: the old classes have zero probability and the new ones move from zero to non-zero probability. However there is an assumption that the class conditional distributions share some commonality and so information can be transferred from the source domain to the one-shot target domain.
Generative Adversarial Networks
(GAN) (Goodfellow et al., 2014), and specifically Deep Convolutional GANs (DCGAN) (Radford et al., 2015) use of the ability to discriminate between true and generated examples as an objective, GAN approaches can learn complex joint densities. Recent improvements in the optimization process (Arjovsky et al., 2017; Gulrajani et al., 2017) have reduced some of the failure modes of the GAN learning process.
(Krizhevsky et al., 2012) is routinely used in classification problems. Often it is non-trivial to encode known invariances in a model. It can be easier to encode those invariances in the data instead by generating additional data items through transformations from existing data items. For example the labels of handwritten characters should be invariant to small shifts in location, small rotations or shears, changes in intensity, changes in stroke thickness, changes in size etc. Almost all cases of data augmentation are from a priori known invariance. Prior to this paper we know of few works that attempt to learn data augmentation strategies. One paper that is worthy of note is the work of (Hauberg et al., 2016), where the authors learn augmentation strategies on a class by class basis. This approach does not transfer to the one-shot setting where completely new classes are considered.
Few-Shot Learning and Meta-Learning:
there have been a number of approaches to few shot learning, from (Salakhutdinov et al., 2012)2016)
, hierarchical variational autoencoders in(Edwards & Storkey, 2017) and most recently a GAN based one-shot generative model (Mehrotra & Dukkipati, 2017). One early but effective approach to one-shot learning involved the use of Siamese networks (Koch, 2015). Others have used nonparametric Bayesian approaches (Salakhutdinov et al., 2012), and conditional variational autoencoders (Rezende et al., 2016). With few examples a nearest neighbour classifier or kernel classifier is an obvious choice. Hence meta-learning distance metrics or weighting kernels has clear potential (Vinyals et al., 2016; Snell et al., 2017). Skip-residual pairwise networks have also proven particularly effective (Mehrotra & Dukkipati, 2017). Various forms of memory augmented networks have also been used to collate the critical sparse information in an incremental way (Santoro et al., 2016; Vinyals et al., 2016). None of these approaches consider an augmentation model as the basis for meta-learning.
3 Models for Data Augmentation
If we know that a class label should be invariant to a particular transformation then we can apply that transformation to generate additional data. If we do not know what transformations might be valid, but we have other data from related problems, we can attempt to learn valid transformations from those related problems that we can apply to our setting (Figure 2). This is an example of meta-learning; we learn on other problems how to improve learning for our target problem. This is the subject of this paper.
Generative Adversarial Methods (Goodfellow et al., 2014) are one approach for learning to generate examples from density matching the density of a training dataset of size denoted by . They learn by minimizing a distribution discrepancy measure between the generated data and the true data. The generative model learnt by a Generative Adversarial Network (GAN) takes the form222The assumption of a Gaussian is without much loss of generality - all the other usual generating distribution are just known nonlinear transformations from a Gaussian; transformations that are fairly straightforward to implement via a neural network.
where is implemented via a neural network. Here,
are the vectors being generated (that, in distribution, should match the data), and are the latent Gaussian variables that provide the variation in what is generated.
A generative adversarial network can be thought of as learning transformations that map out a data manifold: gives one point on the manifold, and changing in each different direction continuously maps out the data manifold.
3.1 Data Augmentation Generative Adversarial Network
A generative adversarial network could also be used to map out a data augmentation manifold. Given a data point we could learn a representation of the input , along with a generative model as before. The combined model would take the form
where the neural network now takes the representation and the random as inputs. Now given any new we can
Obtain a generatively meaningful representation of the data point , that encapsulates the information needed to generate other related data.
Generate extra augmentation data that supplements the original , which can be used in a classifier. This can be done y sampling
from the standard Gaussian distribution and then using the generative network to generate an augmentation example.
The precise form of data augmentation model we use in this paper is given in Figure 3. An outline of the structure of the model is elaborated in the figure caption.
The data augmentation model (3) can be learnt in the source domain using an adversarial approach. Consider a source domain consisting of data and corresponding target values . We use an improved WGAN (Gulrajani et al., 2017) critic that either takes
[leftmargin=*, nolistsep, itemsep=2pt]
some input data point and a second data point from the same class: such that .
some input data point and the output of the current generator which takes as an input.
The critic tries to discriminate the generated points (b) from the real points (a). The generator is trained to minimize this discriminative capability as measured by the Wasserstein distance (Arjovsky et al., 2017).
The importance of providing the original to the discriminator should be emphasised. We want to ensure the generator is capable of generating different data that is related to, but different from, the current data point. By providing information about the current data point to the discriminator we prevent the GAN from simply autoencoding the current data point. At the same time we do not provide class information, so it has to learn to generalise in ways that are consistent across all classes.
In the main experiments, we used a DAGAN generator that was a combination of a UNet and ResNet, which we henceforth call a UResNet. The UResNet generator has a total of 8 blocks, each block having 4 convolutional layers (with leaky rectified linear (relu) activations and batch renormalisation (batchrenorm)(Ioffe, 2017)
) followed by one downscaling or upscaling layer. Downscaling layers (in blocks 1-4) were convolutions with stride 2 followed by leaky relu, batch normalisation and dropout. Upscaling layers were stride 1/2 replicators, followed by a convolution, leaky relu, batch renormalisation and dropout. For Omniglot and EMNIST experiments, all layers had 64 filters. For the VGG-Faces the first 2 blocks of the encoder and the last 2 blocks of the decoder had 64 filters and the last 2 blocks of the encoder and the first 2 blocks of the decoder 128 filters.
In addition each block of the UResNet generator had skip connections. As with a standard ResNet, a strided 1x1 convolution also passes information between blocks, bypassing the between block non-linearity to help gradient flow. Finally skip connections were introduced between equivalent sized filters at each end of the network (as with UNet). A figure of the architecture and a pseudocode definition is given in Appendix A.
We used a DenseNet (Huang et al., 2016) discriminator, using layer normalization instead of batch normalization; the latter would break the assumptions of the WGAN objective function. The DenseNet was composed of 4 Dense Blocks and 4 Transition Layers, as they were defined in (Huang et al., 2016). We used a growth rate of and each Dense Block had 4 convolutional layers within it. For the discriminator we also used dropout at the last convolutional layer of each Dense Block as we found that this improves sample quality.
We trained each DAGAN for 500 epochs, using a learning rate of 0.0001, and an Adam optimizer with Adam parameters ofand .
For each classification experiment we used a DenseNet classifier composed of 4 Dense Blocks and 4 Transition Layers with a
, each Dense Block had 3 convolutional layers within it. The classifiers were a total of 17 layers (i.e. 16 layers and 1 softmax layer). Furthermore we applied a dropout of 0.5 on the last convolutional layer in each Dense Block. The classifier was trained with standard augmentation: random Gaussian noise was added to images (with 50% probability), random shifts along x and y axis (with 50% probability), and random 90 degree rotations (all with equal probability of being chosen). Classifiers were trained for 200 epochs, a learning rate of 0.001, and an Adam optmizer withand .
We tested the DAGAN augmentation on 3 datasets: Omniglot, EMNIST, and VGG-Faces. All datasets were split randomly into source domain sets, validation domain sets and test domain sets.
For classifier networks, all data for each character (handwritten or person) was further split into 2 test cases (for all datasets), 3 validation cases and a varying number of training cases depending on the experiment. Classifier training was done on the training cases for all examples in all domains, with hyperparameter choice made on the validation cases. Finally test performance was reported only on the test cases for the target domain set. Case splits were randomized across each test run.
For one-shot networks, DAGAN training was done on the source domain, and the meta learning done on the source domain, and validated on the validation domain. Results were presented on the target domain data. Again in the target domain a varying number of training cases were provided and results were presented on the test cases (2 cases for each target domain class in all datasets).
The Omniglot data (Lake et al., 2015) was split into source domain and target domain similarly to the split in (Vinyals et al., 2016). The order of the classes was shuffled such that the source and target domains contain diverse samples (i.e. from various alphabets). The first 1200 were used as a source domain set, 1201-1412 as a validation domain set and 1412-1623 as a target domain test set.
The EMNIST data was split into a source domain that included classes 0-34 (after random shuffling of the classes), the validation domain set included classes 35-42 and the test domain set included classes 42-47. Since the EMNIST dataset has thousands of samples per class we chose only a subset of 100 for each class, so that we could make our task a low-data one.
In the VGG-Face dataset case, we randomly chose 100 samples from each class that had 100 uncorrupted images, resulting in 2396 of the full 2622 classes available in the dataset. After shuffling, we split the resulting dataset into a source domain that included the first 1802 classes. The test domain set included classes 1803-2300 and the validation domain set included classes 2300-2396.
6 DAGAN Training and Generation
6.1 Training of DAGAN in source domain
A DAGAN was trained on Source Omniglot domains using a variety of architectures: standard VGG, U-nets, and ResNet inspired architectures. Increasingly powerful networks proved better generators, with the UResNet described in Section 4 generator being our model of choice. Figure 4 shows the improved variability in generations that is achieved with more powerful architectures. The DAGAN was also trained on the VGG-Faces source domains. Examples of the generated faces are given in Figure 5
6.2 Vanilla Classifiers
The first test is how well the DAGAN can augment vanilla classifiers trained on each target domain. A DenseNet classifier (as described in Section 4) was trained first on just real data (with standard data augmentation) with 5, 10 or 15 examples per class. In the second case, the classifier was also passed DAGAN generated augmented data. The real or fake label was also passed to the network, to enable the network to learn how best to emphasise true over generated data. This last step proved crucial to maximizing the potential of the DAGAN augmentations. In each training cycle, varying numbers of augmented samples were provided for each real example (ranging from 1-10); the best annotation rate was selected via performance on the validation domain. The results on the held out test cases from the target domain is given in Table 1. In every case the augmentation improves the classification.
|Omniglot DAGAN Augmented Classification|
|Experiment ID||Samples Per Class||Test Accuracy|
|EMNIST DAGAN Augmented Classification|
|Experiment ID||Samples Per Class||Test Accuracy|
|Face DAGAN Augmented Classification|
|Experiment ID||Samples Per Class||Test Accuracy|
7 One Shot Learning Using Data Augmentation Networks and Matching Networks
A standard approach to one shot learning involves learning an appropriate distance between representations that can be used with a nearest neighbour classifier under some metric. We focus on the use of Matching Networks to learn a representation space, where distances in that representation space produce good classifiers. The same networks can then be used in a target domain for nearest neighbour classification.
Matching networks were introduced by (Vinyals et al., 2016). A matching network creates a predictor from a support set in the target domain by using an attention memory network to generate an appropriate comparison space for comparing a test example with each of the training examples. This is achieved by simulating the process of having small support sets from the source domain and learning to learn a good mapping.
However matching networks can only learn to match based on individual examples. By augmenting the support sets and then learning to learn from that augmented data, we can enhance the classification power of matching networks, and apply that to the augmented domain. Beyond that we can learn to choose good examples from the DAGAN space that can best related the test and training manifolds. Precisely, we train the DAGAN on the source domain, then train the matching networks on the source domain, along with a sample-selector neural network that selects the best representative input used to create an additional datum that augments each case. As all these quantities are differentiable, this whole process can be trained as a full pipeline. Networks are tested on the validation domain, and the best performing matching network and sample-selector network combination are selected for use in the test domain.
When training matching networks with DAGAN augmentation, augmentation was used during every matching network training episode to simulate the data augmentation process. We used matching networks without full-context embedding version and stacked K GAN generated (augmented) images along with the original image. We ran experiments for 20 classes and one sample per class per episode, i.e. the one shot learning setting. The ratio of generated to real data was varied from 0 to 2. Once again standard augmentations were applied to all data in both DAGAN augmented and standard settings. In Table 2, showing the Omniglot one-shot results, we see that the DAGAN was enhancing even simple pixel distance with an improvement of 33.815%. Furthermore in our matching network experiments we saw an improvement of 0.5% over the state of the art (96.9% to 97.4%), pushes the matching network performance to the level of the Conv Arc (Shyam et al., 2017) technique which used substantial prior knowledge of the data form (that the data was built from pen strokes) to achieve those results.
In addition to Omniglot, we experimented on EMNIST and VGG-Face. We used the Omniglot DAGAN to generate augmentation samples for EMNIST to stress test the power of the network to generalise over dissimilar domains. The one-shot results showed an improvement of 1.8% over the baseline matching network that was training on EMNIST (i.e. from 59.5% to 61.3%, again averaged over 3 runs). In the VGG-Face one shot matching networks experiments we saw that the the DAGAN augmented system performed with same performance as the baseline one. However it is worth noting that the non augmented matching network did not overfit the dataset as it happened in Omniglot and EMNIST experiments. This indicates that the embedding network architecture we used perhaps did not have enough capacity to learn a complex task such as discriminating between faces. The embedding network architecture used was the one described in (Vinyals et al., 2016) with 4 convolutional layers with 64 filters each. Further experiments with larger embedding functions are required to better explore and evaluate the DAGAN performance on VGG-Face dataset.
|Technique Name||Test Accuracy|
|Pixel Distance + DAGAN Augmentations||0.60515|
|GR + Siam-I||0.936|
|GR + Siam-II||0.912|
|Matching Nets (Local Reproduction)||0.969|
|Matching Nets + DAGAN Augmentations||0.974|
Data augmentation is a widely applicable approach to improving performance in low-data setting, and a DAGAN is a flexible model to automatic learn to augment data. However beyond that, We demonstrate that DAGANS improve performance of classifiers even after standard data-augmentation. Furthermore by meta-learning the best choice of augmentation in a one-shot setting it leads to better performance than other state of the art meta learning methods. The generality of data augmentation across all models and methods means that a DAGAN could be a valuable addition to any low data setting.
- Arjovsky et al. (2017) M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein GAN. ArXiv e-prints, January 2017.
- Arjovsky et al. (2017) Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein gan. arXiv preprint arXiv:1701.07875, 2017.
- Ba et al. (2016) Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.
Clark & Storkey (2015)
C. Clark and A.J. Storkey.
Training deep convolutional networks to play go.
Proceedings of 32nd International Conference on Machine Learning (ICML2015), 2015. (arxiv 2014).
- Deng et al. (2009) Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pp. 248–255. IEEE, 2009.
- Edwards & Storkey (2017) Harrison Edwards and Amos Storkey. Towards a neural statistician. ICLR, 2017.
- Foerster et al. (2016) Jakob Foerster, Yannis M Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems, pp. 2137–2145, 2016.
- Goodfellow et al. (2014) I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative Adversarial Networks. ArXiv e-prints, June 2014.
- Goodfellow et al. (2014) Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pp. 2672–2680, 2014. URL http://papers.nips.cc/paper/5423-generative-adversarial-nets.
- Gu et al. (2015) Jiuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting Liu, Xingxing Wang, and Gang Wang. Recent Advances in Convolutional Neural Networks. dec 2015. URL http://arxiv.org/abs/1512.07108.
- Gu et al. (2016) Shixiang Gu, Timothy Lillicrap, Ilya Sutskever, and Sergey Levine. Continuous deep q-learning with model-based acceleration. In International Conference on Machine Learning, pp. 2829–2838, 2016.
- Gulrajani et al. (2017) I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved Training of Wasserstein GANs. ArXiv e-prints, March 2017.
- Gulrajani et al. (2017) Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron Courville. Improved training of wasserstein gans. arXiv preprint arXiv:1704.00028, 2017.
- Hauberg et al. (2016) Søren Hauberg, Oren Freifeld, Anders Boesen Lindbo Larsen, John Fisher, and Lars Hansen. Dreaming more data: Class-dependent distributions over diffeomorphisms for learned data augmentation. In Artificial Intelligence and Statistics, pp. 342–350, 2016.
- He et al. (2015a) Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. feb 2015a. URL http://arxiv.org/abs/1502.01852.
- He et al. (2015b) Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. dec 2015b. URL http://arxiv.org/abs/1512.03385.
- He et al. (2016) Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European Conference on Computer Vision, pp. 630–645. Springer, 2016.
- Hinton et al. (2012a) Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal Processing Magazine, 29(6):82–97, 2012a.
- Hinton et al. (2012b) Geoffrey E. Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R. Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. jul 2012b. URL http://arxiv.org/abs/1207.0580.
- Huang et al. (2016) Gao Huang, Zhuang Liu, Kilian Q Weinberger, and Laurens van der Maaten. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993, 2016.
- Ioffe (2017) Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models, 2017.
- Ioffe & Szegedy (2015) Sergey Ioffe and Christian Szegedy. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. feb 2015. URL http://arxiv.org/abs/1502.03167.
- Koch (2015) Gregory Koch. Siamese neural networks for one-shot image recognition. Doctoral dissertation, University of Toronto, 2015.
Krizhevsky et al. (2012)
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton.
Imagenet classification with deep convolutional neural networks.In Advances in neural information processing systems, pp. 1097–1105, 2012.
- Lake et al. (2015) Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332–1338, 2015.
- Mehrotra & Dukkipati (2017) Akshay Mehrotra and Ambedkar Dukkipati. Generative adversarial residual pairwise networks for one shot learning. arXiv preprint arXiv:1704.00028, 2017. URL http://arxiv.org/abs/1703.08033.
- Mnih et al. (2015) Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
- Radford et al. (2015) Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In Proceedings of ICLR 2016, volume abs/1511.06434, 2015. URL http://arxiv.org/abs/1511.06434.
- Rezende et al. (2016) Danilo Rezende, Ivo Danihelka, Karol Gregor, Daan Wierstra, et al. One-shot generalization in deep generative models. In Proceedings of The 33rd International Conference on Machine Learning, pp. 1521–1529, 2016.
Salakhutdinov et al. (2012)
Ruslan Salakhutdinov, Joshua B Tenenbaum, and Antonio Torralba.
One-shot learning with a hierarchical nonparametric bayesian model.
ICML Unsupervised and Transfer Learning, pp. 195–206, 2012.
- Santoro et al. (2016) Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. One-shot learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016.
- Shimodaira (2000) H. Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90:227–244, 2000.
- Shyam et al. (2017) Pranav Shyam, Shubham Gupta, and Ambedkar Dukkipati. Attentive recurrent comparators. CoRR, abs/1703.00767, 2017. URL http://arxiv.org/abs/1703.00767.
- Silver et al. (2016) David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.
- Simard et al. (1991) Patrice Y. Simard, Bernard Victorri, Yann LeCun, and John S. Denker. Tangent prop - a formalism for specifying selected invariances in an adaptive network. In NIPS, pp. 895–903, 1991.
- Snell et al. (2017) Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. CoRR, abs/1703.05175, 2017. URL http://arxiv.org/abs/1703.05175.
- Storkey (2009) A.J. Storkey. When training and test sets are different: Characterising learning transfer. In Candela Sugiyama Schwaighofer Lawrence (ed.), Dataset Shift in Machine Learning, chapter 1, pp. 3–28. MIT Press, 2009. URL http://mitpress.mit.edu/catalog/item/default.asp?ttype=2&tid=11755.
- Storkey & Sugiyama (2007) A.J. Storkey and M. Sugiyama. Mixture regression for covariate shift. In Advances in Neural Information Processing Systems 19 (NIPS2006), 2007.
Sugiyama & Müller (2005)
M. Sugiyama and K. -R. Müller.
Input-dependent estimation of generalisation error under covariate shift.Statistics and Decisions, 23:249–279, 2005.
- Van Hasselt et al. (2016) Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In AAAI, pp. 2094–2100, 2016.
- Vinyals et al. (2016) Oriol Vinyals, Charles Blundell, Tim Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, pp. 3630–3638, 2016.
- Wang et al. (2017) Yuxuan Wang, R. J. Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc V. Le, Yannis Agiomyrgiannakis, Rob Clark, and Rif A. Saurous. Tacotron: A fully end-to-end text-to-speech synthesis model. CoRR, abs/1703.10135, 2017. URL http://arxiv.org/abs/1703.10135.
- White (2016) T. White. Sampling Generative Networks. ArXiv e-prints, September 2016.
- Wu et al. (2016) Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016.