Standard sequential generation methods assume a pre-specified generation order, such as text generation methods which generate words from left to right. In this work, we propose a framework for training models of text generation that operate in non-monotonic orders; the model directly learns good orders, without any additional annotation. Our framework operates by generating a word at an arbitrary position, and then recursively generating words to its left and then words to its right, yielding a binary tree. Learning is framed as imitation learning, including a coaching method which moves from imitating an oracle to reinforcing the policy's own preferences. Experimental results demonstrate that using the proposed method, it is possible to learn policies which generate text without pre-specifying a generation order, while achieving competitive performance with conventional left-to-right generation.

Despite the notable successes in video games such as Atari 2600, current AI is yet to defeat human champions in the domain of real-time strategy (RTS) games. One of the reasons is that an RTS game is a multi-agent game, in which single-agent reinforcement learning methods cannot simply be applied because the environment is not a stationary Markov Decision Process. In this paper, we present a first step toward finding a game-theoretic solution to RTS games by applying Neural Fictitious Self-Play (NFSP), a game-theoretic approach for finding Nash equilibria, to Mini-RTS, a small but nontrivial RTS game provided on the ELF platform. More specifically, we show that NFSP can be effectively combined with policy gradient reinforcement learning and be applied to Mini-RTS. Experimental results also show that the scalability of NFSP can be substantially improved by pretraining the models with simple self-play using policy gradients, which by itself gives a strong strategy despite its lack of theoretical guarantee of convergence.

In the domain of algorithmic music composition, machine learning-driven systems eliminate the need for carefully hand-crafting rules for composition. In particular, the capability of recurrent neural networks to learn complex temporal patterns lends itself well to the musical domain. Promising results have been observed across a number of recent attempts at music composition using deep RNNs. These approaches generally aim at first training neural networks to reproduce subsequences drawn from existing songs. Subsequently, they are used to compose music either at the audio sample-level or at the note-level. We designed a representation that divides polyphonic music into a small number of monophonic streams. This representation greatly reduces the complexity of the problem and eliminates an exponential number of probably poor compositions. On top of our LSTM neural network that learnt musical sequences in this representation, we built an RL agent that learnt to find combinations of songs whose joint dominance produced pleasant compositions. We present \textbf{Amadeus}, an algorithmic music composition system that composes music that consists of intricate melodies, basic chords, and even occasional contrapuntal sequences.

How can we learn to do probabilistic inference in a way that generalizes between models? Amortized variational inference learns for a single model, sharing statistical strength across observations. This benefits scalability and model learning, but does not help with generalization to new models. We propose meta-amortized variational inference, a framework that amortizes the cost of inference over a family of generative models. We apply this approach to deep generative models by introducing the MetaVAE: a variational autoencoder that learns to generalize to new distributions and rapidly solve new unsupervised learning problems using only a small number of target examples. Empirically, we validate the approach by showing that the MetaVAE can: (1) capture relevant sufficient statistics for inference, (2) learn useful representations of data for downstream tasks such as clustering, and (3) perform meta-density estimation on unseen synthetic distributions and out-of-sample Omniglot alphabets.

Quantization of neural networks has become common practice, driven by the need for efficient implementations of deep neural networks on embedded devices. In this paper, we exploit an oft-overlooked degree of freedom in most networks - for a given layer, individual output channels can be scaled by any factor provided that the corresponding weights of the next layer are inversely scaled. Therefore, a given network has many factorizations which change the weights of the network without changing its function. We present a conceptually simple and easy to implement method that uses this property and show that proper factorizations significantly decrease the degradation caused by quantization. We show improvement on a wide variety of networks and achieve state-of-the-art degradation results for MobileNets. While our focus is on quantization, this type of factorization is applicable to other domains such as network-pruning, neural nets regularization and network interpretability.

Due to the high training costs of deep learning, model developers often rent cloud GPU servers to achieve better efficiency. However, this practice raises privacy concerns. An adversarial party may be interested in 1) personal identifiable information encoded in the training data and the learned models, 2) misusing the sensitive models for its own benefits, or 3) launching model inversion (MIA) and generative adversarial network (GAN) attacks to reconstruct replicas of training data (e.g., sensitive images). Learning from encrypted data seems impractical due to the large training data and expensive learning algorithms, while differential-privacy based approaches have to make significant trade-offs between privacy and model quality. We investigate the use of image disguising techniques to protect both data and model privacy. Our preliminary results show that with block-wise permutation and transformations, surprisingly, disguised images still give reasonably well performing deep neural networks (DNN). The disguised images are also resilient to the deep-learning enhanced visual discrimination attack and provide an extra layer of protection from MIA and GAN attacks.

Leveled Homomorphic Encryption (LHE) offers a potential solution that could allow sectors with sensitive data to utilize the cloud and securely deploy their models for remote inference with Deep Neural Networks (DNN). However, this application faces several obstacles due to the limitations of LHE. One of the main problems is the incompatibility of commonly used nonlinear functions in DNN with the operations supported by LHE, i.e. addition and multiplication. As common in LHE approaches, we train a model with a nonlinear function, and replace it with a low-degree polynomial approximation at inference time on private data. While this typically leads to approximation errors and loss in prediction accuracy, we propose a method that reduces this loss to small values or eliminates it entirely, depending on simple hyper-parameters. This is achieved by the introduction of a novel and elegantly simple Min-Max normalization scheme, which scales inputs to nonlinear functions into ranges with low approximation error. While being intuitive in its concept and trivial to implement, we empirically show that it offers a stable and effective approximation solution to nonlinear functions in DNN. In return, this can enable deeper networks with LHE, and facilitate the development of security- and privacy-aware analytics applications.

We propose a modular extension of the backpropagation algorithm for computation of the block diagonal of the training objective's Hessian to various levels of refinement. The approach compartmentalizes the otherwise tedious construction of the Hessian into local modules. It is applicable to feedforward neural network architectures, and can be integrated into existing machine learning libraries with relatively little overhead, facilitating the development of novel second-order optimization methods. Our formulation subsumes several recently proposed block-diagonal approximation schemes as special cases. Our PyTorch implementation is included with the paper.

Abstract--Mind wandering (MW) is a ubiquitous phenomenon which reflects a shift in attention from task-related to task-unrelated thoughts. There is a need for intelligent interfaces that can reorient attention when MW is detected due to its detrimental effects on performance and productivity. In this paper, we propose a deep learning model for MW detection using Electroencephalogram (EEG) signals. Specifically, we develop a channel-wise deep convolutional neural network (CNN) model to classify the features of focusing state and MW extracted from EEG signals. This is the first study that employs CNN to automatically detect MW using only EEG data. The experimental results on the collected dataset demonstrate promising performance with 91.78% accuracy, 92.84% sensitivity, and 90.73% specificity.

The paper surveys recent extensions of the Long-Short Term Memory networks to handle tree structures from the perspective of learning non-trivial forms of isomorph structured transductions. It provides a discussion of modern TreeLSTM models, showing the effect of the bias induced by the direction of tree processing. An empirical analysis is performed on real-world benchmarks, highlighting how there is no single model adequate to effectively approach all transduction problems.