Adversarial Feature Learning

Jeff DonahuePhilipp KrähenbühlTrevor Darrell

article2016ICLR1,922 citations

Proposes Bidirectional Generative Adversarial Networks (BiGANs) to jointly learn data generation and latent feature inference, establishing generative adversarial modeling as an effective framework for unsupervised visual representation learning.

Listen

Modern computer vision and machine learning systems rely heavily on massive datasets annotated by humans. Hand-labeling data is expensive, time-consuming, and fails to exploit the rich structural information naturally present in raw, unlabeled data. Generative Adversarial Networks (GANs) have demonstrated an impressive ability to capture complex data distributions and semantic variations within their internal representations. However, standard generative models only map random codes to synthesized data, lacking a mechanism to project real data back into that feature space for downstream recognition tasks.

To address this limitation, the article introduces and evaluates Bidirectional Generative Adversarial Networks (BiGANs). The objective is to demonstrate that an adversarial framework can simultaneously learn a generative model and an inverse mapping encoder without supervision, yielding rich feature representations useful for transfer learning.

The authors designed a framework where an encoder maps real data into a latent feature space, a generator maps latent variables into data space, and a discriminator is trained to jointly distinguish true data-feature pairs from generated data-feature pairs. Theoretically, the authors prove that optimizing this minimax objective forces the generator and encoder to become exact mathematical inverses at the global optimum. Empirically, the framework was evaluated across unstructured data using the permutation-invariant MNIST digit benchmark and complex visual data using the 1.2-million-image ImageNet dataset. Learned representations were transferred and tested on standard computer vision benchmarks, including ImageNet classification, PASCAL VOC object classification, Fast R-CNN object detection, and Fully Convolutional Network semantic segmentation.

The findings confirm that BiGAN successfully learns meaningful semantic representations completely unsupervised. First, on high-resolution ImageNet transfer tasks, BiGAN outperformed standard GAN baselines—such as training a feature regressor or extracting features from a standard discriminator—by substantial margins (e.g., achieving 58.9% mean average precision on PASCAL VOC classification versus 56.4%–57.1% for GAN baselines). Second, BiGAN proved competitive with domain-specific self-supervised methods on visual transfer benchmarks (reaching 46.9% detection mAP and 35.2% segmentation mean intersection-over-union when using a higher-resolution encoder), despite making no domain-specific assumptions about image structure. Third, on permutation-invariant MNIST, BiGAN achieved 97.39% accuracy, performing on par with standard autoencoders while operating on completely unstructured data vectors.

These results imply that generative adversarial modeling can serve as a powerful, general-purpose foundation for unsupervised feature learning. Because BiGAN operates directly on uncorrupted data and is domain-agnostic, it avoids the task-specific corruptions, domain shifts, and video-tracking requirements that limit many self-supervised approaches. This lowers the cost and complexity of pretraining deep networks for perceptual tasks where labeled data is scarce.

Organizations developing machine learning pipelines should consider BiGAN-style joint adversarial training as an effective strategy for representation learning from unlabeled data. When dealing with images, teams should adopt the generalized BiGAN formulation that inputs higher-resolution images to the encoder than the generator synthesizes, as this yields measurable downstream performance gains. Future efforts should focus on integrating advanced generative network architectures to further close the performance gap with fully supervised models.

Key limitations include the non-convex optimization inherent to adversarial training, which means the encoder and generator operate as approximate rather than exact theoretical inverses in practice. Furthermore, while BiGAN outperforms other purely unsupervised baselines, fully supervised pretraining still achieves superior downstream accuracy (e.g., 78.3% vs. 60.3% on VOC classification). Readers can be confident in the theoretical derivations and empirical transfer gains, though operational deployment must account for standard adversarial training stability considerations.

University of California Berkeley University of Texas at Austin generative adversarial networks generative model latent factor Adversarial feature learning BiGAN encoder Unsupervised training

arXiv: 1605.09782 jeffdonahue/bigan

Paper: Generative Adversarial Networks, Ian J. Goodfellow et al. (2014). It introduces the foundational generative adversarial network (GAN) framework that BiGAN directly adapts and inverts for feature learning.
Paper: Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, Alec Radford et al. (2016). It establishes standard convolutional GAN architectures and evaluates discriminator representations for unsupervised downstream tasks, providing a baseline comparison for BiGAN's encoder.
Paper: Adversarial Autoencoders, Alireza Makhzani et al. (2015). It introduces adversarial autoencoders, establishing early techniques for matching latent distributions and performing representation learning with adversarial objectives.
Paper: Autoencoding beyond pixels using a learned similarity metric, Anders Boesen Lindbo Larsen et al. (2015). It pioneered the combination of autoencoders and GAN discriminators to extract high-level feature representations beyond pixel-space reconstruction.
Paper: Tutorial on Variational Autoencoders, Carl Doersch (2016). It provides the essential conceptual foundation for probabilistic latent variable models and amortized inference networks.
Paper: Semi-supervised Learning with Deep Generative Models, Diederik P. Kingma et al. (2014). It formulates deep generative semi-supervised learning by leveraging learned latent feature representations for downstream classification tasks.

Paper: Adversarially Learned Inference, Vincent Dumoulin et al. (2017). It develops Adversarially Learned Inference (ALI), proposing a concurrent and complementary framework for bidirectional joint-distribution adversarial feature learning.
Paper: Generative Adversarial Networks: An Overview, Antonia Creswell et al. (2017). It provides a comprehensive overview contextualizing bidirectional inference architectures like BiGAN within the broader landscape of generative adversarial models.
Paper: Wasserstein Auto-Encoders, Ilya Tolstikhin et al. (2018). It generalizes adversarial autoencoding and latent distribution matching by utilizing optimal transport and Wasserstein distances.
Paper: Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks, Jun-Yan Zhu et al. (2017). It builds upon bidirectional adversarial mappings by introducing cycle-consistency constraints to translate between unpaired domains.
Paper: Large Scale GAN Training for High Fidelity Natural Image Synthesis, Andrew Brock et al. (2019). It scales adversarial generative modeling to massive datasets, co-authored by the lead author of BiGAN to advance high-fidelity representation synthesis.

Abstract

The ability of the Generative Adversarial Networks (GANs) framework to learn generative models mapping from simple latent distributions to arbitrarily complex data distributions has been demonstrated empirically, with compelling results showing that the latent space of such generators captures semantic variation in the data distribution. Intuitively, models trained to predict these semantic latent representations given data may serve as useful feature representations for auxiliary problems where semantics are relevant. However, in their existing form, GANs have no means of learning the inverse mapping -- projecting data back into the latent space. We propose Bidirectional Generative Adversarial Networks (BiGANs) as a means of learning this inverse mapping, and demonstrate that the resulting learned feature representation is useful for auxiliary supervised discrimination tasks, competitive with contemporary approaches to unsupervised and self-supervised feature learning.

Topics

generative adversarial networks generative model latent factor latent variable representation learning semantics distribution functions network architecture attention mechanism

Keywords

Adversarial feature learning BiGAN encoder Unsupervised training Inverse mapping Latent representations Jensen-Shannon divergence Permutation-invariant MNIST

1 Introduction
2 Preliminaries
3 Bidirectional Generative Adversarial Networks
3.1 Optimal discriminator, generator, & encoder
3.2 Optimal generator & encoder are inverses
3.3 Relationship to autoencoders
3.4 Learning
3.5 Generalized BiGAN
4 Evaluation
4.1 Baseline methods
4.2 Permutation-invariant MNIST
4.3 ImageNet
4.4 Discussion
References
A Additional proofs
A.1 Proof of Proposition (optimal discriminator)
A.2 Proof of Proposition (encoder and generator objective)
A.3 Measure definitions for deterministic EE and GG
A.4 Proof of Theorem (optimal generator and encoder are inverses)
A.5 Proof of Theorem (relationship to autoencoders)
B Learning details
C Model and training details
C.1 Permutation-invariant MNIST
C.2 ImageNet

Knowls

Knowl 1 — Bidirectional Generative Adversarial Networks (BiGAN) Minimax Formulation
model/method
Bidirectional Generative Adversarial Networks (BiGANs) extend standard Generative Adversarial Networks by introducing an encoder network $E: \Omega_X \to \Omega_Z$ alongside the generator network $G: \Omega_Z \to \Omega_X$ , where $\Omega_X$ is the data space and $\Omega_Z$ is a continuous latent feature space. The encoder induces a conditional distribution $p_E(z|x) = \delta(z - E(x))$ given real data $x \sim p_X(x)$ , and the generator defines $p_G(x|z) = \delta(x - G(z))$ given latent variables $z \sim p_Z(z)$ .

The discriminator $D: \Omega_X \times \Omega_Z \to [0, 1]$ discriminates jointly over data and latent pairs, assigning probability $P_D(Y = 1 | x, z)$ indicating that the pair originates from the encoder distribution $(x, E(x))$ rather than the generator distribution $(G(z), z)$ . The BiGAN minimax training objective is defined as:

$\min_{G, E} \max_D V(D, E, G)$

where

$V(D, E, G) := \mathbb{E}_{x \sim p_X} \left[ \mathbb{E}_{z \sim p_E(\cdot|x)} [\log D(x, z)] \right] + \mathbb{E}_{z \sim p_Z} \left[ \mathbb{E}_{x \sim p_G(\cdot|z)} [\log (1 - D(x, z))] \right]$

For deterministic mappings $E$ and $G$ , this simplifies to:

$V(D, E, G) = \mathbb{E}_{x \sim p_X} [\log D(x, E(x))] + \mathbb{E}_{z \sim p_Z} [\log(1 - D(G(z), z))]$
Knowl 2 — Global Optimality and Jensen-Shannon Divergence of BiGANs
theoretical result
Let $P_{EX}$ and $P_{GZ}$ denote the joint probability measures over $\Omega = \Omega_X \times \Omega_Z$ modeled by the encoder ( $p_{EX}(x, z) := p_E(z|x)p_X(x)$ ) and the generator ( $p_{GZ}(x, z) := p_G(x|z)p_Z(z)$ ), respectively.
1. Optimal Discriminator: For any fixed generator $G$ and encoder $E$ , the optimal discriminator $D^*_{EG} := \arg\max_D V(D, E, G)$ equals the Radon-Nikodym derivative of $P_{EX}$ with respect to the sum measure $P_{EX} + P_{GZ}$ :
$D^*_{EG}(x, z) = \frac{d P_{EX}}{d(P_{EX} + P_{GZ})}(x, z)$
1. Joint Divergence Objective: For the optimal discriminator $D^*_{EG}$ , the minimax objective $C(E, G) := \max_D V(D, E, G) = V(D^*_{EG}, E, G)$ reduces directly to the Jensen-Shannon divergence $D_{JS}$ between the joint probability measures $P_{EX}$ and $P_{GZ}$ :
$C(E, G) = 2 D_{JS}(P_{EX} \parallel P_{GZ}) - \log 4$
1. Global Optimum: The global minimum of $C(E, G)$ is achieved if and only if the joint distributions match identically ( $P_{EX} = P_{GZ}$ ). At this global minimum, $C(E, G) = -\log 4$ and the optimal discriminator cannot differentiate the joint pairs, outputting $D^*_{EG}(x, z) = \frac{1}{2}$ everywhere on the support.
Knowl 3 — Inversion Property of Optimal Deterministic BiGAN Modules
theoretical result
Let $E: \Omega_X \to \Omega_Z$ and $G: \Omega_Z \to \Omega_X$ be deterministic encoder and generator mappings, and let $P_X$ and $P_Z$ be the data and latent probability distributions with supports $\hat{\Omega}_X := \text{supp}(P_X)$ and $\hat{\Omega}_Z := \text{supp}(P_Z)$ .

If $E$ and $G$ achieve the global optimum of the BiGAN objective (such that $P_{EX} = P_{GZ}$ ), then $E$ and $G$ invert each other almost everywhere:

$G(E(x)) = x \quad \text{for } P_X\text{-almost every } x \in \Omega_X$ $E(G(z)) = z \quad \text{for } P_Z\text{-almost every } z \in \Omega_Z$

That is, $E = G^{-1}$ almost everywhere on the support of the respective distributions without needing explicit reconstruction loss terms during training.
Knowl 4 — Equivalence of BiGAN Objective to Autoencoder with L0 Loss
theoretical result
Given deterministic mappings $E: \Omega_X \to \Omega_Z$ and $G: \Omega_Z \to \Omega_X$ , the BiGAN objective evaluated at the optimal discriminator $D^*_{EG} = f_{EG} := \frac{d P_{EX}}{d(P_{EX} + P_{GZ})}$ can be rewritten as:

$C(E, G) = \mathbb{E}_{x \sim p_X}\left[ \mathbf{1}_{[E(x) \in \hat{\Omega}_Z \wedge G(E(x)) = x]} \log f_{EG}(x, E(x)) \right] + \mathbb{E}_{z \sim p_Z}\left[ \mathbf{1}_{[G(z) \in \hat{\Omega}_X \wedge E(G(z)) = z]} \log (1 - f_{EG}(G(z), z)) \right]$

where $\hat{\Omega}_X = \text{supp}(P_X)$ , $\hat{\Omega}_Z = \text{supp}(P_Z)$ , and $\log f_{EG}, \log(1 - f_{EG}) \in (-\infty, 0)$ almost everywhere on the support intersection $R^1 = \text{supp}(P_{EX}) \cap \text{supp}(P_{GZ})$ .

The indicator $\mathbf{1}_{[G(E(x)) = x]}$ acts as an $\ell_0$ autoencoder reconstruction loss on real data $x$ , while the indicator $\mathbf{1}_{[E(G(z)) = z]}$ enforces encoder inversion of generated samples. Unlike classical autoencoders, no structural distance metric is assumed on $\Omega_X$ , as feature metrics are learned entirely through the discriminator.
Knowl 5 — Generalized BiGAN Formulation for Asymmetric Resolutions
model/method
To allow the encoder $E$ to take higher-resolution inputs than the generator $G$ produces, the BiGAN objective generalizes with transformation functions $g_X: \Omega_X \to \Omega'_X$ and $g_Z: \Omega_Z \to \Omega'_Z$ , where $\Omega'_X$ and $\Omega'_Z$ represent intermediate or lower-dimensional representation spaces.

The modified modules are $E: \Omega_X \to \Omega'_Z$ , $G: \Omega_Z \to \Omega'_X$ , and discriminator $D: \Omega'_X \times \Omega'_Z \to [0, 1]$ , with value function:

$V(D, E, G) = \mathbb{E}_{x \sim p_X} [\log D(g_X(x), E(x))] + \mathbb{E}_{z \sim p_Z} [\log(1 - D(G(z), g_Z(z)))]$

For visual representation learning, $g_X(x)$ downsamples a high-resolution input $x \in \Omega_X$ to a lower resolution matching the generator output $G(z) \in \Omega'_X$ , while $g_Z(z) = z$ is the identity map. At global optimality, $P_{EX'} = P_{GZ'}$ holds over $\Omega'_X \times \Omega'_Z$ , and for deterministic models, $\exists z \in \text{supp}(P_Z)$ such that $E(x) = g_Z(z) \wedge G(z) = g_X(x)$ for $P_X$ -almost every $x$ .
Knowl 6 — BiGAN Training Procedure with Simultaneous Updates and Inverse Generator-Encoder Objective
algorithm
BiGAN optimization utilizes mini-batch stochastic gradient steps over parameters $\theta_D$ , $\theta_G$ , and $\theta_E$ . To avoid gradient saturation during generator and encoder training, $G$ and $E$ are trained using an inverted label objective $\Lambda(D, G, E)$ with identical fixed-point characteristics:

$\Lambda(D, G, E) = \mathbb{E}_{x \sim p_X} [\log(1 - D(x, E(x)))] + \mathbb{E}_{z \sim p_Z} [\log D(G(z), z)]$
Input: Data distribution $p_X$ , latent prior distribution $p_Z$ , batch size $n$ , Adam hyperparameters $\alpha=2\times 10^{-4}, \beta_1=0.5, \beta_2=0.999$ , weight decay $\lambda=2.5\times 10^{-5}$
Output: Trained encoder $E_{\theta_E}$ , generator $G_{\theta_G}$ , and discriminator $D_{\theta_D}$
while training not converged do
    Sample mini-batch of $n$ samples $\{x^{(1)}, \dots, x^{(n)}\} \sim p_X$
    Sample mini-batch of $n$ noise vectors $\{z^{(1)}, \dots, z^{(n)}\} \sim p_Z$

    Compute discriminator gradient:
     $g_D \leftarrow \nabla_{\theta_D} \frac{1}{n} \sum_{i=1}^n \left[ \log D_{\theta_D}(x^{(i)}, E_{\theta_E}(x^{(i)})) + \log(1 - D_{\theta_D}(G_{\theta_G}(z^{(i)}), z^{(i)})) \right] - \lambda \theta_D$

    Compute generator-encoder inverse gradient:
     $g_{E,G} \leftarrow \nabla_{\theta_E, \theta_G} \frac{1}{n} \sum_{i=1}^n \left[ \log(1 - D_{\theta_D}(x^{(i)}, E_{\theta_E}(x^{(i)}))) + \log D_{\theta_D}(G_{\theta_G}(z^{(i)}), z^{(i)}) \right] - \lambda (\theta_E + \theta_G)$

    Update $\theta_D$ with Adam step in direction $+g_D$
    Update $\theta_E, \theta_G$ with Adam step in direction $+g_{E,G}$
end while
All parameters $\theta_D, \theta_G, \theta_E$ are updated simultaneously on each mini-batch iteration without alternating discriminator and generator/encoder update loops. The learning rate $\alpha$ is decayed exponentially from $2 \times 10^{-4}$ to $2 \times 10^{-6}$ over training.
Knowl 7 — BiGAN Discriminator Latent Conditioning Architecture
model/method
In the BiGAN discriminator network $D(x, z)$ , the data sample $x$ is supplied as the initial input to the first layer. At each subsequent linear layer, the latent representation $z$ is injected into the intermediate computation.

Specifically, $z$ is transformed via a learned linear projection to match the channel or hidden dimension of that layer and is added directly to the spatial convolution or linear output before the non-linear activation is applied.

To balance gradient flow between $x$ and $z$ , weights in $D$ that multiply $z$ to be added to spatial convolution outputs are initialized with standard deviation scaled by the convolution kernel size. For a $5 \times 5$ spatial kernel, these weights are initialized from a zero-mean normal distribution with standard deviation $\sigma = 0.5$ ( $25 \times$ the standard baseline initialization of $0.02$ ).

Knowl 8 — Permutation-Invariant MNIST Unsupervised Feature Classification

data/table

On permutation-invariant MNIST, each $28 \times 28$ image is flattened to an unstructured 784-dimensional vector and processed with multi-layer perceptron (MLP) architectures for $D$ , $G$ , and $E$ , each having two hidden layers of 1024 units with batch normalization. The latent space is a 50D continuous uniform distribution $p_Z = [\mathcal{U}(-1, 1)]^{50}$ .

Evaluated using 1-Nearest-Neighbor (1NN) classification error on the learned feature representations:

Here, $D$ denotes intermediate discriminator features, LR is a Latent Regressor minimizing $\mathcal{L}(z, E(G(z)))$ after GAN training, JLR trains LR jointly with GAN training, and AE denotes an autoencoder trained directly with $\ell_1$ or $\ell_2$ pixel reconstruction losses. All methods achieve comparable performance around 97.1–97.6%, showing that BiGAN successfully learns representations on unstructured vectors where the generator accurately covers the data distribution.

Knowl 9 — Transfer Learning on PASCAL VOC Classification, Detection, and Segmentation

data/table

Unsupervised feature representations pretrained on ImageNet LSVRC are transferred to PASCAL VOC tasks using AlexNet architectures. Transfer is evaluated on VOC 2007 classification (mean average precision, mAP) under frozen fc layers (fc8, fc6-8, and full fine-tuning all), VOC 2007 Fast R-CNN object detection (all fine-tuned mAP), and VOC 2012 Fully Convolutional Network (FCN) semantic segmentation (mean intersection over union, % mIU):

		Classification (% mAP)			FRCN (% mAP)	FCN (% mIU)
	Pretraining Method	fc8	fc6-8	all	all	all
Supervised	ImageNet (Krizhevsky et al., 2012)	77.0	78.8	78.3	56.8	48.0
Self-Sup.	Agrawal et al. (2015)	31.2	31.0	54.2	43.9	-
	Pathak et al. (2016)	30.5	34.6	56.5	44.5	30.0
	Wang Gupta (2015)	28.4	55.6	63.1	47.4	-
	Doersch et al. (2015)	44.7	55.1	65.3	51.1	-
Unsup.	$k$ -means (Krähenbühl et al., 2016)	32.0	39.2	56.6	45.6	32.6
	Discriminator ( $D$ )	30.7	40.5	56.4	-	-
	Latent Regressor (LR)	36.9	47.9	57.1	-	-
	Joint LR	37.1	47.9	56.5	-	-
	Autoencoder ( $\ell_2$ )	24.8	16.0	53.8	41.9	-
	BiGAN (ours, $64 \times 64$ )	37.5	48.7	58.9	46.2	34.9
	BiGAN ( $112 \times 112$ $E$ , ours)	41.7	52.5	60.3	46.9	35.2

BiGAN outperforms all purely unsupervised baselines (including discriminator-only features and latent regression) and autoencoders across all transfer regimes, and achieves competitive results with self-supervised methods that rely on handcrafted pretext tasks or video priors.

Knowl 10 — Layer-Wise Frozen Evaluation on ImageNet Classification

data/table

Unsupervised pretraining methods are evaluated on the ImageNet LSVRC validation set by freezing the first $N$ convolutional layers of an AlexNet encoder (from conv1 up to convN), randomly reinitializing the remaining layers, and training them fully supervised for 1000-way classification:

Method	conv1	conv2	conv3	conv4	conv5
Random	48.5	41.0	34.8	27.1	12.0
Wang Gupta (2015)	51.8	46.9	42.8	38.8	29.8
Doersch et al. (2015)	53.1	47.6	48.7	45.6	30.4
Noroozi Favaro (2016)*	57.1	56.0	52.4	48.3	38.1
BiGAN (ours)	56.2	54.4	49.4	43.9	33.3
BiGAN ( $112 \times 112$ $E$ , ours)	55.3	53.2	49.3	44.4	34.8

*(Note: Noroozi & Favaro use an altered convnet with larger intermediate feature maps). BiGAN preserves high accuracy across conv1–conv5 layers without requiring hand-designed domain-specific prediction tasks, outperforming baseline random initialization and video-based tracking representations.

Coverage note — None was omitted; all key theoretical definitions, optimality proofs/characterizations, algorithmic update rules, architectures, and empirical benchmark tables across MNIST, ImageNet, and PASCAL VOC were extracted.

References

1.Pulkit Agrawal, Joao Carreira, and Jitendra Malik. Learning to see by moving. In ICCV, 2015.
2.Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2015.
3.Emily L. Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep generative image models using a Laplacian pyramid of adversarial networks. In NIPS, 2015.
4.Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015.
5.Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. DeCAF: A deep convolutional activation feature for generic visual recognition. In ICML, 2014.
6.Vincent Dumoulin, Ishmael Belghazi, Ben Poole, Alex Lamb, Martin Arjovsky, Olivier Mastropietro, and Aaron Courville. Adversarially learned inference. arXiv:1606.00704, 2016.
7.Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The PASCAL Visual Object Classes challenge: A retrospective. IJCV, 2014.
8.Ross Girshick. Fast R-CNN. In ICCV, 2015.
9.Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014.
10.Ian Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. Maxout networks. In ICML, 2013.
11.Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014.
12.Alex Graves, Abdel-rahman Mohamed, and Geoffrey E. Hinton. Speech recognition with deep recurrent neural networks. In ICASSP, 2013.
13.Geoffrey E. Hinton and Ruslan R. Salakhutdinov. Reducing the dimensionality of data with neural networks. Science, 2006.
14.Geoffrey E. Hinton, Simon Osindero, and Yee-Whye Teh. A fast learning algorithm for deep belief nets. Neural Computation, 2006.
15.Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015.
16.Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv:1408.5093, 2014.
17.Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.
18.Philipp Krähenbühl, Carl Doersch, Jeff Donahue, and Trevor Darrell. Data-dependent initializations of convolutional neural networks. In ICLR, 2016.
19.Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classification with deep convolutional neural networks. In NIPS, 2012.
20.Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proc. IEEE, 1998.
21.Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015.
22.Andrew L. Maas, Awni Y. Hannun, and Andrew Y. Ng. Rectifier nonlinearities improve neural network acoustic models. In ICML, 2013.
23.Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016.
24.Deepak Pathak, Philipp Krähenbühl, Jeff Donahue, Trevor Darrell, and Alexei A. Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016.
25.Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In ICLR, 2016.
26.Ali Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. CNN features off-the-shelf: an astounding baseline for recognition. In CVPR Workshops, 2014.
27.Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Fei-Fei Li. ImageNet large scale visual recognition challenge. IJCV, 2015.
28.Ruslan Salakhutdinov and Geoffrey E. Hinton. Deep Boltzmann machines. In AISTATS, 2009.
29.Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to sequence learning with neural networks. In NIPS, 2014.
30.Theano Development Team. Theano: A Python framework for fast computation of mathematical expressions. arXiv:1605.02688, 2016.
31.Oriol Vinyals, Łukasz Kaiser, Terry Koo, Slav Petrov, Ilya Sutskever, and Geoffrey E. Hinton. Grammar as a foreign language. In NIPS, 2015.
32.Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, 2015.
33.Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014.

Citation

MLA

Donahue, J., et al. “Adversarial Feature Learning”. arXiv, 2016, http://arxiv.org/abs/1605.09782v7.

APA

Donahue, J., Krähenbühl, P., & Darrell, T. (2016). Adversarial Feature Learning. arXiv. http://arxiv.org/abs/1605.09782v7

Chicago

Donahue, J., P. Krähenbühl, and T. Darrell. 2016. “Adversarial Feature Learning”. arXiv. http://arxiv.org/abs/1605.09782v7.

Harvard

Donahue, J., Krähenbühl, P. and Darrell, T. (2016) “Adversarial Feature Learning”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1605.09782v7.

Vancouver

1. Donahue J, Krähenbühl P, Darrell T (2016) Adversarial Feature Learning. arXiv

BibTeX

@article{donahue2016adversarial,
  title = {Adversarial Feature Learning},
  author = {Donahue, Jeff and Krähenbühl, Philipp and Darrell, Trevor},
  year = {2016},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1605.09782v7},
  eprint = {1605.09782}
}

Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors

Method	BiGAN	$D$	LR	JLR	AE ( $\ell_2$ )	AE ( $\ell_1$ )
1NN Accuracy (%)	97.39	97.30	97.44	97.13	97.58	97.63

Adversarial Feature Learning

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Bidirectional Generative Adversarial Networks (BiGAN) Minimax Formulation

Knowl 2 — Global Optimality and Jensen-Shannon Divergence of BiGANs

Knowl 3 — Inversion Property of Optimal Deterministic BiGAN Modules

Knowl 4 — Equivalence of BiGAN Objective to Autoencoder with L0 Loss

Knowl 5 — Generalized BiGAN Formulation for Asymmetric Resolutions

Knowl 6 — BiGAN Training Procedure with Simultaneous Updates and Inverse Generator-Encoder Objective

Knowl 7 — BiGAN Discriminator Latent Conditioning Architecture

Knowl 8 — Permutation-Invariant MNIST Unsupervised Feature Classification

Knowl 9 — Transfer Learning on PASCAL VOC Classification, Detection, and Segmentation

Knowl 10 — Layer-Wise Frozen Evaluation on ImageNet Classification

References

Citation

BibTeX

Source Code

Access the Paper