We present a framework for autonomously learning a portable representation that describes a collection of low-level continuous environments. We show that these abstract representations can be learned in a task-independent egocentric space specific to the agent that, when grounded with problem-specific information, are provably sufficient for planning. We demonstrate transfer in two different domains, where an agent learns a portable, task-independent symbolic vocabulary, as well as rules expressed in that vocabulary, and then learns to instantiate those rules on a per-task basis. This reduces the number of samples required to learn a representation of a new task.

A recent strategy to circumvent the exploding and vanishing gradient problem in RNNs, and to allow the stable propagation of signals over long time scales, is to constrain recurrent connectivity matrices to be orthogonal or unitary. This ensures eigenvalues with unit norm and thus stable dynamics and training. However this comes at the cost of reduced expressivity due to the limited variety of orthogonal transformations. We propose a novel connectivity structure based on the Schur decomposition and a splitting of the Schur form into normal and non-normal parts. This allows to parametrize matrices with unit-norm eigenspectra without orthogonality constraints on eigenbases. The resulting architecture ensures access to a larger space of spectrally constrained matrices, of which orthogonal matrices are a subset. This crucial difference retains the stability advantages and training speed of orthogonal RNNs while enhancing expressivity, especially on tasks that require computations over ongoing input sequences.

Performance estimation aims at estimating the loss that a predictive model will incur on unseen data. These procedures are part of the pipeline in every machine learning project and are used for assessing the overall generalisation ability of predictive models. In this paper we address the application of these methods to time series forecasting tasks. For independent and identically distributed data the most common approach is cross-validation. However, the dependency among observations in time series raises some caveats about the most appropriate way to estimate performance in this type of data and currently there is no settled way to do so. We compare different variants of cross-validation and of out-of-sample approaches using two case studies: One with 62 real-world time series and another with three synthetic time series. Results show noticeable differences in the performance estimation methods in the two scenarios. In particular, empirical experiments suggest that cross-validation approaches can be applied to stationary time series. However, in real-world scenarios, when different sources of non-stationary variation are at play, the most accurate estimates are produced by out-of-sample methods that preserve the temporal order of observations.

We perform an experimental study of the dynamics of Stochastic Gradient Descent (SGD) in learning deep neural networks for several real and synthetic classification tasks. We show that in the initial epochs, almost all of the performance improvement of the classifier obtained by SGD can be explained by a linear classifier. More generally, we give evidence for the hypothesis that, as iterations progress, SGD learns functions of increasing complexity. This hypothesis can be helpful in explaining why SGD-learned classifiers tend to generalize well even in the over-parameterized regime. We also show that the linear classifier learned in the initial stages is "retained" throughout the execution even if training is continued to the point of zero training error, and complement this with a theoretical result in a simplified model. Key to our work is a new measure of how well one classifier explains the performance of another, based on conditional mutual information.

Massively multi-label prediction/classification problems arise in environments like health-care or biology where very precise predictions are useful. One challenge with massively multi-label problems is that there is often a long-tailed frequency distribution for the labels, which results in few positive examples for the rare labels. We propose a solution to this problem by modifying the output layer of a neural network to create a Bayesian network of sigmoids which takes advantage of ontology relationships between the labels to help share information between the rare and the more common labels. We apply this method to the two massively multi-label tasks of disease prediction (ICD-9 codes) and protein function prediction (Gene Ontology terms) and obtain significant improvements in per-label AUROC and average precision for less common labels.

Neuromorphic hardware architectures represent a growing family of potential post-Moore's Law Era platforms. Largely due to event-driving processing inspired by the human brain, these computer platforms can offer significant energy benefits compared to traditional von Neumann processors. Unfortunately there still remains considerable difficulty in successfully programming, configuring and deploying neuromorphic systems. We present the Fugu framework as an answer to this need. Rather than necessitating a developer attain intricate knowledge of how to program and exploit spiking neural dynamics to utilize the potential benefits of neuromorphic computing, Fugu is designed to provide a higher level abstraction as a hardware-independent mechanism for linking a variety of scalable spiking neural algorithms from a variety of sources. Individual kernels linked together provide sophisticated processing through compositionality. Fugu is intended to be suitable for a wide-range of neuromorphic applications, including machine learning, scientific computing, and more brain-inspired neural algorithms. Ultimately, we hope the community adopts this and other open standardization attempts allowing for free exchange and easy implementations of the ever-growing list of spiking neural algorithms.

While there are convergence guarantees for temporal difference (TD) learning when using linear function approximators, the situation for nonlinear models is far less understood, and divergent examples are known. Here we take a first step towards extending theoretical convergence guarantees to TD learning with nonlinear function approximation. More precisely, we consider the expected dynamics of the TD(0) algorithm. We prove that this ODE is attracted to a compact set for smooth homogeneous functions including some ReLU networks. For over-parametrized and well-conditioned functions in sufficiently reversible environments we prove convergence to the global optimum. This result improves when using $k$-step or $ \lambda$ returns. Finally, we generalize a divergent counterexample to a family of divergent problems to motivate the assumptions needed to prove convergence.

The goal of feature selection is to identify important features that are relevant to explain an outcome variable. Most of the work in this domain has focused on identifying globally relevant features, which are features that are related to the outcome using evidence across the entire dataset. We study a more fine-grained statistical problem: conditional feature selection, where a feature may be relevant depending on the values of the other features. For example in genetic association studies, variant $A$ could be associated with the phenotype in the entire dataset, but conditioned on variant $B$ being present it might be independent of the phenotype. In this sense, variant $A$ is globally relevant, but conditioned on $B$ it is no longer locally relevant in that region of the feature space. We present a generalization of the knockoff procedure that performs conditional feature selection while controlling a generalization of the false discovery rate (FDR) to the conditional setting. By exploiting the feature/response model-free framework of the knockoffs, the quality of the statistical FDR guarantee is not degraded even when we perform conditional feature selections. We implement this method and present an algorithm that automatically partitions the feature space such that it enhances the differences between selected sets in different regions, and validate the statistical theoretical results with experiments.

Encoder-decoder recurrent neural network models (Seq2Seq) have achieved great success in ubiquitous areas of computation and applications. It was shown to be successful in modeling data with both temporal and spatial dependencies for translation or prediction tasks. In this study, we propose a dimension reduction approach to visualize and interpret the representation of the data by these models. We propose to view the hidden states of the encoder and the decoder as spatio-temporal snapshots of network dynamics and to apply proper orthogonal decomposition to their concatenation to compute a low-dimensional embedding for hidden state dynamics. Projection of the decoder states onto such interpretable embedding space shows that Seq2Seq training to predict sequences using gradient-descent back propagation effectively performs dimension reduction consisting of only a small percentage of dimensions of the network's hidden units. Furthermore, sequences are being clustered into well separable clusters in the low dimensional space each of which corresponds to a different type of dynamics. The projection methodology also clarifies the roles of the encoder and the decoder components of the network. We show that the projection of encoder hidden states onto the low dimensional space provides an initializing trajectory directing the sequence to the cluster which corresponds to that particular type of distinct dynamics and the projection of the decoder hidden states constitutes the embedded cluster attractor. Inspection of the low dimensional space and the projections onto it during training shows that the estimation of clusters separability in the embedding can be utilized to estimate the optimality of model training. We test and demonstrate our proposed interpretability methodology on synthetic examples (dynamics on a circle and an ellipse) and on 3D human body movement data.

We introduce a new class of context dependent, incomplete information games to serve as structured prediction models for settings with significant strategic interactions. Our games map the input context to outcomes by first condensing the input into private player types that specify the utilities, weighted interactions, as well as the initial strategies for the players. The game is played over multiple rounds where players respond to weighted aggregates of their neighbors' strategies. The predicted output from the model is a mixed strategy profile (a near-Nash equilibrium) and each observation is thought of as a sample from this strategy profile. We introduce two new aggregator paradigms with provably convergent game dynamics, and characterize the conditions under which our games are identifiable from data. Our games can be parameterized in a transferable manner so that the sets of players can change from one game to another. We demonstrate empirically that our games as models can recover meaningful strategic interactions from real voting data.