# Deep Learning Library: accel-brain-base.¶

accel-brain-base is a basic library of the Deep Learning for rapid development at low cost. This library makes it possible to design and implement deep learning, which must be configured as a complex system or a System of Systems, by combining a plurality of functionally differentiated modules such as a Restricted Boltzmann Machine(RBM), Deep Boltzmann Machines(DBMs), a Stacked-Auto-Encoder, an Encoder/Decoder based on Long Short-Term Memory(LSTM), and a Convolutional Auto-Encoder(CAE).

From the view points of functionally equivalents and structural expansions, this library also prototypes many variants such as energy-based models and Generative models. Typical examples are Generative Adversarial Networks(GANs) and Adversarial Auto-Encoders(AAEs). In addition, it provides deep reinforcement learning that applies the neural network described above as a function approximator.

## Installation¶

Install using pip:

pip install accel-brain-base

### Source code¶

The source code is currently hosted on GitHub.

### Python package index(PyPI)¶

Installers for the latest released version are available at the Python package index.

### Dependencies¶

#### For ML Ops.¶

In this library, almost all models inherit HybridBlock from mxnet.gluon. Functions for common ML Ops such as saving and loading parameters are provided by HybridBlock.

Note that, however, we are using HybridBlocks in mxnet.gluon at present but we will also implement functional equivalents that use TensorFlow and PyTorch in the future. There is no reason to stick to one library.

## Documentation¶

Full documentation is available on https://code.accel-brain.com/Accel-Brain-Base/README.html. This document contains information on functionally reusability, functional scalability and functional extensibility.

## Problem Setting: Deep Learning after the era of “Democratization of Artificial Intelligence(AI)”.¶

How the Research and Development(R&D) on the subject of machine learning including deep learning, after the era of “Democratization of Artificial Intelligence(AI)”, can become possible? Simply implementing the models and algorithms provided by standard machine learning libraries and applications like AutoML would reinvent the wheel. If you just copy and paste the demo code from the library and use it, your R&D would fall into dogmatically authoritarian development, or so-called the Hype driven development.

If you fall in love with the concept of “Democratization of AI,” you may forget the reality that the R&D is under the influence of not only democracy but also capitalism. The R&D provides economic value when its R&D artifacts are distinguished from the models and algorithms realized by standard machine learning libraries and applications such as AutoML. In general terms, R&D must provide a differentiator to maximize the scarcity of its implementation artifacts.

On the other hand, it must be remembered that any R&D builds on the history of the social structure and the semantics of the concepts envisioned by previous studies. Many models and algorithms are variants derived not only from research but also from the relationship with business domains. It is impossible to assume differentiating factors without taking commonality and identity between society and its history.

Considering many variable parts, structural unions, and functional equivalents in the deep learning paradigm, which are variants derived not only from research but also from the relationship with business domains, from perspective of commonality/variability analysis in order to practice object-oriented design, this library provides abstract classes that define the skeleton of the deep Learning algorithm in an operation, deferring some steps in concrete variant algorithms such as the Deep Boltzmann Machines, Stacked Auto-Encoder, Encoder/Decoder based on LSTM, and Convolutional Auto-Encoder to client subclasses. The abstract classes and the interfaces in this library let subclasses redefine certain steps of the deep Learning algorithm without changing the algorithm’s structure.

These abstract classes can also provide new original models and algorithms such as Generative Adversarial Networks(GANs), Deep Reinforcement Learning, or Neural network language model by implementing the variable parts of the fluid elements of objects.

### Problem Solution: Deep Boltzmann Machines.¶

The function of this library is building and modeling Restricted Boltzmann Machine(RBM) and Deep Boltzmann Machine(DBM). The models are functionally equivalent to stacked auto-encoder. The basic function is the same as dimensions reduction or pre-learning for so-called transfer learning.

#### The structure of RBM.¶

According to graph theory, the structure of RBM corresponds to a complete bipartite graph which is a special kind of bipartite graph where every node in the visible layer is connected to every node in the hidden layer. Based on statistical mechanics and thermodynamics(Ackley, D. H., Hinton, G. E., & Sejnowski, T. J. 1985), the state of this structure can be reflected by the energy function:

where is a bias in visible layer, is a bias in hidden layer, is an activity or a state in visible layer, is an activity or a state in hidden layer, and is a weight matrix in visible and hidden layer. The activities can be calculated as the below product, since the link of activations of visible layer and hidden layer are conditionally independent.

#### The learning equations of RBM.¶

Because of the rules of conditional independence, the learning equations of RBM can be introduced as simple form. The distribution of visible state which is marginalized over the hidden state is as following:

where is a partition function in statistical mechanics or thermodynamics. Let be set of observed data points, then . Therefore the gradients on the parameter of the log-likelihood function are

where is an expected value for . is a sigmoid function.

The learning equations of RBM are introduced by performing control so that those gradients can become zero.

#### Contrastive Divergence as an approximation method.¶

In relation to RBM, Contrastive Divergence(CD) is a method for approximation of the gradients of the log-likelihood(Hinton, G. E. 2002). The procedure of this method is similar to Markov Chain Monte Carlo method(MCMC). However, unlike MCMC, the visbile variables to be set first in visible layer is not randomly initialized but the observed data points in training dataset are set to the first visbile variables. And, like Gibbs sampler, drawing samples from hidden variables and visible variables is repeated k times. Empirically (and surprisingly), k is considered to be 1.

#### The structure of DBM.¶

Salakhutdinov, R., Hinton, G. E. (2009). Deep boltzmann machines. In International conference on artificial intelligence and statistics (pp. 448-455). p451.

As is well known, DBM is composed of layers of RBMs stacked on top of each other(Salakhutdinov, R., & Hinton, G. E. 2009). This model is a structural expansion of Deep Belief Networks(DBN), which is known as one of the earliest models of Deep Learning(Le Roux, N., & Bengio, Y. 2008). Like RBM, DBN places nodes in layers. However, only the uppermost layer is composed of undirected edges, and the other consists of directed edges. DBN with R hidden layers is below probabilistic model:

where r = 0 points to visible layer. Considerling simultaneous distribution in top two layer,

and conditional distributions in other layers are as follows:

The pre-learning of DBN engages in a procedure of recursive learning in layer-by-layer. However, as you can see from the difference of graph structure, DBM is slightly different from DBN in the form of pre-learning. For instance, if r = 1, the conditional distribution of visible layer is

.

On the other hand, the conditional distribution in the intermediate layer is

where 2 has been introduced considering that the intermediate layer r receives input data from Shallower layer r-1 and deeper layer r+1. DBM sets these parameters as initial states.

#### DBM as a Stacked Auto-Encoder.¶

DBM is functionally equivalent to a Stacked Auto-Encoder, which is-a neural network that tries to reconstruct its input. To encode the observed data points, the function of DBM is as linear transformation of feature map below

.

On the other hand, to decode this feature points, the function of DBM is as linear transformation of feature map below

.

The reconstruction error should be calculated in relation to problem setting such as the Representation Learning. In default, this library computes Mean Squared Error(MSE) or L2 norm. For instance, my jupyter notebook: demo/Deep-Boltzmann-Machines-for-Representation-Learning.ipynb demonstrates that DBM which is a Stacked Auto-Encoder can minimize the reconstruction errors based on the Representation Learning.

### Functionally equivalent: Encoder/Decoder based on LSTM.¶

The methodology of equivalent-functionalism enables us to introduce more functional equivalents and compare problem solutions structured with different algorithms and models in common problem setting. For example, in dimension reduction problem, the function of Encoder/Decoder schema is equivalent to DBM as a Stacked Auto-Encoder.

Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078., p2.

According to the neural networks theory, and in relation to manifold hypothesis, it is well known that multilayer neural networks can learn features of observed data points and have the feature points in hidden layer. High-dimensional data can be converted to low-dimensional codes by training the model such as Stacked Auto-Encoder and Encoder/Decoder with a small central layer to reconstruct high-dimensional input vectors. This function of dimensionality reduction facilitates feature expressions to calculate similarity of each data point.

This library provides Encoder/Decoder based on LSTM, which is a reconstruction model and makes it possible to extract series features embedded in deeper layers. The LSTM encoder learns a fixed length vector of time-series observed data points and the LSTM decoder uses this representation to reconstruct the time-series using the current hidden state and the value inferenced at the previous time-step.

One interesting application example is the Encoder/Decoder for Anomaly Detection (EncDec-AD) paradigm (Malhotra, P., et al. 2016). This reconstruction model learns to reconstruct normal time-series behavior, and thereafter uses reconstruction error to detect anomalies. Malhotra, P., et al. (2016) showed that EncDec-AD paradigm is robust and can detect anomalies from predictable, unpredictable, periodic, aperiodic, and quasi-periodic time-series. Further, they showed that the paradigm is able to detect anomalies from short time-series (length as small as 30) as well as long time-series (length as large as 500).

As the prototype is exemplified in demo/Encoder-Decoder-based-on-LSTM-for-Anomaly-Detection.ipynb, this library provides Encoder/Decoder based on LSTM as a EncDec-AD scheme.

### Functionally equivalent: Convolutional Auto-Encoder.¶

Convolutional Auto-Encoder(Masci, J., et al., 2011) is a functionally equivalent of Stacked Auto-Encoder in relation to problem settings such as image segmentation, object detection, inpainting and graphics. A stack of Convolutional Auto-Encoder forms a convolutional neural network(CNN), which are among the most successful models for supervised image classification. Each Convolutional Auto-Encoder is trained using conventional on-line gradient descent without additional regularization terms.

Reconstructed image by Convolutional Auto-Encoder.

This library can draw a distinction between Stacked Auto-Encoder and Convolutional Auto-Encoder, and is able to design and implement respective models. Stacked Auto-Encoder ignores the 2 dimentional image structures. In many cases, the rank of observed tensors extracted from image dataset is more than 3. This is not only a problem when dealing with realistically sized inputs, but also introduces redundancy in the parameters, forcing each feature to be global. Like Shape-BM, Convolutional Auto-Encoder differs from Stacked Auto-Encoder as their weights are shared among all locations in the input, preserving spatial locality. Hence, the reconstructed image data is due to a linear combination of basic image patches based on the latent code.

In this library, Convolutional Auto-Encoder is also based on Encoder/Decoder scheme. The encoder is to the decoder what the Convolution is to the Deconvolution. The Deconvolution also called transposed convolutions “work by swapping the forward and backward passes of a convolution.” (Dumoulin, V., & Visin, F. 2016, p20.)

In relation to the Representation Learning, like DBM, this model also can minimize the reconstruction errors. An example can be found in demo/Convolutional-Auto-Encoder-for-Representation-Learning.ipynb.

### Functionally equivalent: Convolutional Contractive Auto-Encoder.¶

This library also provides some functionally equivalents of the Convolutional Auto-Encoder. For instance, Convolutional Contractive Auto-Encoder(Contractive CAE) is a Convolutional Auto-Encoder based on the First-Order Contractive Auto-Encoder(Rifai, S., et al., 2011), which executes the representation learning by adding a penalty term to the classical reconstruction cost function. This penalty term corresponds to the Frobenius norm of the Jacobian matrix of the encoder activations with respect to the input and results in a localized space contraction which in turn yields robust features on the activation layer.

Analogically, the Contractive Convolutional Auto-Encoder calculates the penalty term. But it differs in that the operation of the deconvolution intervenes insted of inner product. The prototype is exemplified in demo/Contractive-Convolutional-Auto-Encoder-for-Representation-Learning.ipynb.

## Issue: Structural extension from Auto-Encoders and Encoder/Decoders to energy-based models and Generative models.¶

Auto-Encoders, such as the Encoder/Decoder, the Convolutional Auto-Encoder, and the DBM have in common that these models are Stacked Auto-Encoders or the reconstruction models. On the other hand, the Auto-Encoders and the Encoder/Decoders are not statistical mechanical energy-based models unlike with RBM or DBM.

However, Auto-Encoders have traditionally been used to represent energy-based models. According to the statistical mechanical theory for energy-based models, Auto-Encoders constructed by neural networks can be associated with an energy landscape, akin to negative log-probability in a probabilistic model, which measures how well the Auto-Encoder can represent regions in the input space. The energy landscape has been commonly inferred heuristically, by using a training criterion that relates the Auto-Encoder to a probabilistic model such as a RBM. The energy function is identical to the free energy of the corresponding RBM, showing that Auto-Encoders and RBMs may be viewed as two different ways to derive training criteria for forming the same type of analytically defined energy landscape.

The view of the Auto-Encoder as a dynamical system allows us to understand how an energy function may be derived for the Auto-Encoder. This makes it possible to assign energies to Auto-Encoders with many different types of activation functions and outputs, and consider minimanization of reconstruction errors as energy minimanization(Kamyshanska, H., & Memisevic, R., 2014).

When trained with some regularization terms, the Auto-Encoders have the ability to learn an energy manifold without supervision or negative examples(Zhao, J., et al., 2016). This means that even when an energy-based Auto-Encoding model is trained to reconstruct a real sample, the model contributes to discovering the data manifold by itself.

This library provides energy-based Auto-Encoders such as Contractive Convolutional Auto-Encoder(Rifai, S., et al., 2011), Repelling Convolutional Auto-Encoder(Zhao, J., et al., 2016), Denoising Auto-Encoders(Bengio, Y., et al., 2013), and Ladder Networks(Valpola, H., 2015). But it is more usefull to redescribe the Auto-Encoders in the framework of Generative Adversarial Networks(GANs)(Goodfellow, I., et al., 2014) to make those models function as not only energy-based models but also Generative models. For instance, theory of an Adversarial Auto-Encoders(AAEs)(Makhzani, A., et al., 2015) and energy-based GANs(EBGANs)(Zhao, J., et al., 2016) enables us to turn Auto-Encoders into a Generative models which referes energy functions.

### Problem Solution: Generative Adversarial Networks(GANs).¶

The Generative Adversarial Networks(GANs) (Goodfellow et al., 2014) framework establishes a min-max adversarial game between two neural networks – a generative model, G, and a discriminative model, D. The discriminator model, D(x), is a neural network that computes the probability that a observed data point x in data space is a sample from the data distribution (positive samples) that we are trying to model, rather than a sample from our generative model (negative samples). Concurrently, the generator uses a function G(z) that maps samples z from the prior p(z) to the data space. G(z) is trained to maximally confuse the discriminator into believing that samples it generates come from the data distribution. The generator is trained by leveraging the gradient of D(x) w.r.t. x, and using that to modify its parameters.

### Problem Solution: Conditional GANs (or cGANs).¶

The Conditional GANs (or cGANs) is a simple extension of the basic GAN model which allows the model to condition on external information. This makes it possible to engage the learned generative model in different “modes” by providing it with different contextual information (Gauthier, J. 2014).

Mirza, M., & Osindero, S. (2014). Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784., p3.

This model can be constructed by simply feeding the data, y, to condition on to both the generator and discriminator. In an unconditioned generative model, because the maps samples z from the prior p(z) are drawn from uniform or normal distribution, there is no control on modes of the data being generated. On the other hand, it is possible to direct the data generation process by conditioning the model on additional information (Mirza, M., & Osindero, S. 2014).

This library also provides the Adversarial Auto-Encoders(AAEs), which is a probabilistic Auto-Encoder that uses GANs to perform variational inference by matching the aggregated posterior of the feature points in hidden layer of the Auto-Encoder with an arbitrary prior distribution(Makhzani, A., et al., 2015). Matching the aggregated posterior to the prior ensures that generating from any part of prior space results in meaningful samples. As a result, the decoder of the Adversarial Auto-Encoder learns a deep generative model that maps the imposed prior to the data distribution.

Makhzani, A., Shlens, J., Jaitly, N., Goodfellow, I., & Frey, B. (2015). Adversarial autoencoders. arXiv preprint arXiv:1511.05644., p2.

### Problem Solution: Energy-based Generative Adversarial Network(EBGAN).¶

Reusing the Auto-Encoders, this library introduces the Energy-based Generative Adversarial Network (EBGAN) model(Zhao, J., et al., 2016) which views the discriminator as an energy function that attributes low energies to the regions near the data manifold and higher energies to other regions. THe Auto-Encoders have traditionally been used to represent energy-based models. When trained with some regularization terms, the Auto-Encoders have the ability to learn an energy manifold without supervision or negative examples. This means that even when an energy-based Auto-Encoding model is trained to reconstruct a real sample, the model contributes to discovering the data manifold by itself.

Zhao, J., Mathieu, M., & LeCun, Y. (2016). Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126., p4.

### Functionally equivalent: Energy-based Adversarial Auto-Encoders(EBAAEs).¶

This library models the Energy-based Adversarial-Auto-Encoder(EBAAE) by structural coupling between AAEs and EBGAN. As the prototype is exemplified in demo/Energy-based-Adversarial-Auto-Encoder-for-Representation-Learning.ipynb, the learning algorithm equivalents an adversarial training of AAEs as a generator and EBGAN as a discriminator.

## Issue: How unsupervised learning like Auto-Encoder, Energy-based Model, and Generative Model can function in classification problem?¶

In most classification problems, finding and producing labels for the samples is hard. In many cases plenty of unlabeled data existand it seems obvious that using them should improve the results. For instance, there are plenty of unlabeled images available and in most image classification tasks there are vastly more bits of information in the statistical structure of input images than in their labels.

It is argued here that the reason why unsupervised learning has not been able to improve results is that most current versions are incompatible with supervised learning. The problem is that many un-supervised learning methods try to represent as much information about the original data as possible whereas supervised learning tries to filter out all the information which is irrelevant for the task at hand.

Ladder network is an Auto-Encoder which can discard information Unsupervised learning needs to toleratediscarding information in order to work well with supervised learning. Many unsupervised learning methods are not good at this but one class of models stands out as an exception: hierarchical latent variable models. Unfortunately their derivation can be quite complicated and often involves approximations which compromise their per-formance.

A simpler alternative is offered by Auto-Encoders which also have the benefit of being compatible with standard supervised feedforward networks. They would be a promising candidate for combining supervised and unsupervised learning but unfortunately Auto-Encoders normally correspond to latent variable models with a single layer of stochastic variables, that is, they do not tolerate discarding information.

Valpola, H. (2015). From neural PCA to deep unsupervised learning. In Advances in independent component analysis and learning machines (pp. 143-171). Academic Press., p6.

Ladder network makes it possible to solve that problem by settting recursive derivation of the learning rule with a distributed cost function, building denoisng Auto-Encoder recursively. Normally denoising Auto-Encoders have a fixed input but the cost functions on the higher layers can influence their input mappings and this creates a bias towards PCA-type solutions.

In relation to problem settings such as the Representation Learning, the Ladder Networks is also functionally equivalent of standard CAE as the prototype is exemplified in demo/Convolutional-Ladder-Networks-for-Representation-Learning.ipynb.

### Problem Solution: Deep Reconstruction-Classification Networks(DRCN or DRCNetworks).¶

Deep Reconstruction-Classification Network(DRCN or DRCNetworks) is a convolutional network that jointly learns two tasks:

1. supervised source label prediction.
2. unsupervised target data reconstruction.

Ideally, a discriminative representation should model both the label and the structure of the data. Based on that intuition, Ghifary, M., et al.(2016) hypothesize that a domain-adaptive representation should satisfy two criteria:

1. classify well the source domain labeled data.
2. reconstruct well the target domain unlabeled data, which can be viewed as an approximate of the ideal discriminative representation.

Ghifary, M., Kleijn, W. B., Zhang, M., Balduzzi, D., & Li, W. (2016, October). Deep reconstruction-classification networks for unsupervised domain adaptation. In European Conference on Computer Vision (pp. 597-613). Springer, Cham., p5.

The encoding parameters of the DRCN are shared across both tasks, while the decoding parameters are sepa-rated. The aim is that the learned label prediction function can perform well onclassifying images in the target domain thus the data reconstruction can beviewed as an auxiliary task to support the adaptation of the label prediction.

Using this library, for instance, we can extend the Convolutional Auto-Encoder in DRCNetworks to the Convolutional Ladder Networks as mentioned in demo/DRCNetworks-for-Dataset-Bias-Problem.ipynb.

## Issue: Structural extension for Deep Reinforcement Learning.¶

The Reinforcement learning theory presents several issues from a perspective of deep learning theory(Mnih, V., et al. 2013). Firstly, deep learning applications have required large amounts of hand-labelled training data. Reinforcement learning algorithms, on the other hand, must be able to learn from a scalar reward signal that is frequently sparse, noisy and delayed.

The difference between the two theories is not only the type of data but also the timing to be observed. The delay between taking actions and receiving rewards, which can be thousands of timesteps long, seems particularly daunting when compared to the direct association between inputs and targets found in supervised learning.

Another issue is that deep learning algorithms assume the data samples to be independent, while in reinforcement learning one typically encounters sequences of highly correlated states. Furthermore, in Reinforcement learning, the data distribution changes as the algorithm learns new behaviours, presenting aspects of recursive learning, which can be problematic for deep learning methods that assume a fixed underlying distribution.

## Problem Re-setting: Generalisation, or a function approximation.¶

This library considers problem setteing in which an agent interacts with an environment , in a sequence of actions, observations and rewards. At each time-step the agent selects an action at from the set of possible actions, . The state/action-value function is .

The goal of the agent is to interact with the by selecting actions in a way that maximises future rewards. We can make the standard assumption that future rewards are discounted by a factor of $\gamma$ per time-step, and define the future discounted return at time as

,

where is the time-step at which the agent will reach the goal. This library defines the optimal state/action-value function as the maximum expected return achievable by following any strategy, after seeing some state and then taking some action ,

,

where is a policy mapping sequences to actions (or distributions over actions).

The optimal state/action-value function obeys an important identity known as the Bellman equation. This is based on the following intuition: if the optimal value of the sequence at the next time-step was known for all possible actions , then the optimal strategy is to select the action maximising the expected value of

,
.

The basic idea behind many reinforcement learning algorithms is to estimate the state/action-value function, by using the Bellman equation as an iterative update,

.

Such value iteration algorithms converge to the optimal state/action-value function, as .

But increasing the complexity of states/actions is equivalent to increasing the number of combinations of states/actions. If the value function is continuous and granularities of states/actions are extremely fine, the combinatorial explosion will be encountered. In other words, this basic approach is totally impractical, because the state/action-value function is estimated separately for each sequence, without any generalisation. Instead, it is common to use a function approximator to estimate the state/action-value function,

So the Reduction of complexities is required.

### Problem Solution: Deep Q-Network¶

In this problem setting, the function of nerual network or deep learning is a function approximation with weights as a Q-Network. A Q-Network can be trained by minimising a loss functions that changes at each iteration ,

where

is the target for iteration and is a so-called behaviour distribution. This is probability distribution over states and actions. The parameters from the previous iteration are held fixed when optimising the loss function . Differentiating the loss function with respect to the weights we arrive at the following gradient,

### Functional equivalent: MobileNet.¶

If you pay attention to the calculation speed, it is better to extend the CNN part that is the function approximation to MobileNet. As mentioned in demo/MobileNet-v2-for-Image-Classification.ipynb, this library provides the MobileNet V2(Sandler, M., et al., 2018).

### Functional equivalent: LSTM.¶

It is not inevitable to functionally reuse CNN as a function approximator. In the above problem setting of generalisation and Combination explosion, for instance, Long Short-Term Memory(LSTM) networks, which is-a special Reccurent Neural Network(RNN) structure, and CNN as a function approximator are functionally equivalent. In the same problem setting, functional equivalents can be functionally replaced. Considering that the feature space of the rewards has the time-series nature, LSTM will be more useful.

## References¶

### Deep Boltzmann machines.¶

• Ackley, D. H., Hinton, G. E., & Sejnowski, T. J. (1985). A learning algorithm for Boltzmann machines. Cognitive science, 9(1), 147-169.
• Boulanger-Lewandowski, N., Bengio, Y., & Vincent, P. (2012). Modeling temporal dependencies in high-dimensional sequences: Application to polyphonic music generation and transcription. arXiv preprint arXiv:1206.6392.
• Eslami, S. A., Heess, N., Williams, C. K., & Winn, J. (2014). The shape boltzmann machine: a strong model of object shape. International Journal of Computer Vision, 107(2), 155-176.
• Hinton, G. E. (2002). Training products of experts by minimizing contrastive divergence. Neural computation, 14(8), 1771-1800.
• Le Roux, N., & Bengio, Y. (2008). Representational power of restricted Boltzmann machines and deep belief networks. Neural computation, 20(6), 1631-1649.
• Lyu, Q., Wu, Z., Zhu, J., & Meng, H. (2015, June). Modelling High-Dimensional Sequences with LSTM-RTRBM: Application to Polyphonic Music Generation. In IJCAI (pp. 4138-4139).
• Lyu, Q., Wu, Z., & Zhu, J. (2015, October). Polyphonic music modelling with LSTM-RTRBM. In Proceedings of the 23rd ACM international conference on Multimedia (pp. 991-994). ACM.
• Salakhutdinov, R., & Hinton, G. E. (2009). Deep boltzmann machines. InInternational conference on artificial intelligence and statistics (pp. 448-455).
• Sutskever, I., Hinton, G. E., & Taylor, G. W. (2009). The recurrent temporal restricted boltzmann machine. In Advances in Neural Information Processing Systems (pp. 1601-1608).

### Auto-Encoders.¶

• Baccouche, M., Mamalet, F., Wolf, C., Garcia, C., & Baskurt, A. (2012, September). Spatio-Temporal Convolutional Sparse Auto-Encoder for Sequence Classification. In BMVC (pp. 1-12).
• Bengio, Y., Yao, L., Alain, G., & Vincent, P. (2013). Generalized denoising auto-encoders as generative models. In Advances in neural information processing systems (pp. 899-907).
• Chong, Y. S., & Tay, Y. H. (2017, June). Abnormal event detection in videos using spatiotemporal autoencoder. In International Symposium on Neural Networks (pp. 189-196). Springer, Cham.
• Masci, J., Meier, U., Cireşan, D., & Schmidhuber, J. (2011, June). Stacked convolutional auto-encoders for hierarchical feature extraction. In International Conference on Artificial Neural Networks (pp. 52-59). Springer, Berlin, Heidelberg.
• Patraucean, V., Handa, A., & Cipolla, R. (2015). Spatio-temporal video autoencoder with differentiable memory. arXiv preprint arXiv:1511.06309.
• Rifai, S., Vincent, P., Muller, X., Glorot, X., & Bengio, Y. (2011, June). Contractive auto-encoders: Explicit invariance during feature extraction. In Proceedings of the 28th International Conference on International Conference on Machine Learning (pp. 833-840). Omnipress.
• Rifai, S., Mesnil, G., Vincent, P., Muller, X., Bengio, Y., Dauphin, Y., & Glorot, X. (2011, September). Higher order contractive auto-encoder. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases (pp. 645-660). Springer, Berlin, Heidelberg.
• Seung, H. S. (1998). Learning continuous attractors in recurrent networks. In Advances in neural information processing systems (pp. 654-660).
• Zhao, J., Mathieu, M., & LeCun, Y. (2016). Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126.

### Encoder/Decoder schemes with an Attention mechanism.¶

• Bahdanau, D., Cho, K., & Bengio, Y. (2014). Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473.
• Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078.
• Malhotra, P., Ramakrishnan, A., Anand, G., Vig, L., Agarwal, P., & Shroff, G. (2016). LSTM-based encoder-decoder for multi-sensor anomaly detection. arXiv preprint arXiv:1607.00148.
• Xingjian, S. H. I., Chen, Z., Wang, H., Yeung, D. Y., Wong, W. K., & Woo, W. C. (2015). Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in neural information processing systems (pp. 802-810).
• Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems (pp. 5998-6008).

• Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., … & Bengio, Y. (2014). Generative adversarial nets. In Advances in neural information processing systems (pp. 2672-2680).
• Makhzani, A., Shlens, J., Jaitly, N., Goodfellow, I., & Frey, B. (2015). Adversarial autoencoders. arXiv preprint arXiv:1511.05644.
• Mirza, M., & Osindero, S. (2014). Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784.
• Zhao, J., Mathieu, M., & LeCun, Y. (2016). Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126.

### Unsupervised / Supervised pre-training¶

• Bengio, Y., Lamblin, P., Popovici, D., & Larochelle, H. (2007). Greedy layer-wise training of deep networks. In Advances in neural information processing systems (pp. 153-160).
• Erhan, D., Bengio, Y., Courville, A., Manzagol, P. A., Vincent, P., & Bengio, S. (2010). Why does unsupervised pre-training help deep learning?. Journal of Machine Learning Research, 11(Feb), 625-660.

### Semi-supervised learning.¶

• Ghifary, M., Kleijn, W. B., Zhang, M., Balduzzi, D., & Li, W. (2016, October). Deep reconstruction-classification networks for unsupervised domain adaptation. In European Conference on Computer Vision (pp. 597-613). Springer, Cham.
• Rasmus, A., Berglund, M., Honkala, M., Valpola, H., & Raiko, T. (2015). Semi-supervised learning with ladder networks. In Advances in neural information processing systems (pp. 3546-3554).
• Valpola, H. (2015). From neural PCA to deep unsupervised learning. In Advances in Independent Component Analysis and Learning Machines (pp. 143-171). Academic Press.

### Deep Reinforcement Learning.¶

• Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078.
• Egorov, M. (2016). Multi-agent deep reinforcement learning.
• Gupta, J. K., Egorov, M., & Kochenderfer, M. (2017, May). Cooperative multi-agent control using deep reinforcement learning. In International Conference on Autonomous Agents and Multiagent Systems (pp. 66-83). Springer, Cham.
• Malhotra, P., Ramakrishnan, A., Anand, G., Vig, L., Agarwal, P., & Shroff, G. (2016). LSTM-based encoder-decoder for multi-sensor anomaly detection. arXiv preprint arXiv:1607.00148.
• Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., & Riedmiller, M. (2013). Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602.
• Sainath, T. N., Vinyals, O., Senior, A., & Sak, H. (2015, April). Convolutional, long short-term memory, fully connected deep neural networks. In Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on (pp. 4580-4584). IEEE.
• Xingjian, S. H. I., Chen, Z., Wang, H., Yeung, D. Y., Wong, W. K., & Woo, W. C. (2015). Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in neural information processing systems (pp. 802-810).
• Zaremba, W., Sutskever, I., & Vinyals, O. (2014). Recurrent neural network regularization. arXiv preprint arXiv:1409.2329.

### Optimizations.¶

• Bengio, Y., Boulanger-Lewandowski, N., & Pascanu, R. (2013, May). Advances in optimizing recurrent networks. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (pp. 8624-8628). IEEE.
• Duchi, J., Hazan, E., & Singer, Y. (2011). Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul), 2121-2159.
• Dozat, T. (2016). Incorporating nesterov momentum into adam., Workshop track - ICLR 2016.
• Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.

### Algorithms, Arithmetic, Regularizations, and Representations learning.¶

• Dumoulin, V., & Visin, F. (2016). A guide to convolution arithmetic for deep learning. arXiv preprint arXiv:1603.07285.
• Erhan, D., Courville, A., & Bengio, Y. (2010). Understanding representations learned in deep architectures. Department dInformatique et Recherche Operationnelle, University of Montreal, QC, Canada, Tech. Rep, 1355, 1.
• Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning (adaptive computation and machine learning series). Adaptive Computation and Machine Learning series, 800.
• He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
• Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167.
• Kamyshanska, H., & Memisevic, R. (2014). The potential energy of an autoencoder. IEEE transactions on pattern analysis and machine intelligence, 37(6), 1261-1273.
• Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L. C. (2018). Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4510-4520).
• Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1), 1929-1958.
• Zaremba, W., Sutskever, I., & Vinyals, O. (2014). Recurrent neural network regularization. arXiv preprint arXiv:1409.2329.

• accel-brain

## Author URI¶

• https://accel-brain.co.jp/
• https://accel-brain.com/