This paper introduces MazeBase: an environment for simple 2D games, designed as a sandbox for machine learning approaches to reasoning and planning. Within it, we create 10 simple games embodying a range of algorithmic tasks (e.g. if-then statements or set negation). A variety of neural models (fully connected, convolutional network, memory network) are deployed via reinforcement learning on these games, with and without a procedurally generated curriculum. Despite the tasks' simplicity, the performance of the models is far from optimal, suggesting directions for future development. We also demonstrate the versatility of MazeBase by using it to emulate small combat scenarios from StarCraft. Models trained on the MazeBase version can be directly applied to StarCraft, where they consistently beat the in-game AI.

Streaming variational Bayes (SVB) is successful in learning LDA models in an online manner. However previous attempts toward developing online Monte-Carlo methods for LDA have little success, often by having much worse perplexity than their batch counterparts. We present a streaming Gibbs sampling (SGS) method, an online extension of the collapsed Gibbs sampling (CGS). Our empirical study shows that SGS can reach similar perplexity as CGS, much better than SVB. Our distributed version of SGS, DSGS, is much more scalable than SVB mainly because the updates' communication complexity is small.

We consider the problem of learning user preferences over robot trajectories for environments rich in objects and humans. This is challenging because the criterion defining a good trajectory varies with users, tasks and interactions in the environment. We represent trajectory preferences using a cost function that the robot learns and uses it to generate good trajectories in new environments. We design a crowdsourcing system - PlanIt, where non-expert users label segments of the robot's trajectory. PlanIt allows us to collect a large amount of user feedback, and using the weak and noisy labels from PlanIt we learn the parameters of our model. We test our approach on 122 different environments for robotic navigation and manipulation tasks. Our extensive experiments show that the learned cost function generates preferred trajectories in human environments. Our crowdsourcing system is publicly available for the visualization of the learned costs and for providing preference feedback: \url{http://planit.cs.cornell.edu}

In previous work, we proposed a logic-based framework in which computation is the execution of actions in an attempt to make reactive rules of the form if antecedent then consequent true in a canonical model of a logic program determined by an initial state, sequence of events, and the resulting sequence of subsequent states. In this model-theoretic semantics, reactive rules are the driving force, and logic programs play only a supporting role. In the canonical model, states, actions and other events are represented with timestamps. But in the operational semantics, for the sake of efficiency, timestamps are omitted and only the current state is maintained. State transitions are performed reactively by executing actions to make the consequents of rules true whenever the antecedents become true. This operational semantics is sound, but incomplete. It cannot make reactive rules true by preventing their antecedents from becoming true, or by proactively making their consequents true before their antecedents become true. In this paper, we characterize the notion of reactive model, and prove that the operational semantics can generate all and only such models. In order to focus on the main issues, we omit the logic programming component of the framework.

We consider the problem of learning decision rules for prediction with feature budget constraint. In particular, we are interested in pruning an ensemble of decision trees to reduce expected feature cost while maintaining high prediction accuracy for any test example. We propose a novel 0-1 integer program formulation for ensemble pruning. Our pruning formulation is general - it takes any ensemble of decision trees as input. By explicitly accounting for feature-sharing across trees together with accuracy/cost trade-off, our method is able to significantly reduce feature cost by pruning subtrees that introduce more loss in terms of feature cost than benefit in terms of prediction accuracy gain. Theoretically, we prove that a linear programming relaxation produces the exact solution of the original integer program. This allows us to use efficient convex optimization tools to obtain an optimally pruned ensemble for any given budget. Empirically, we see that our pruning algorithm significantly improves the performance of the state of the art ensemble method BudgetRF.

Gaussian Processes (GPs) are widely used tools in statistics, machine learning, robotics, computer vision, and scientific computation. However, despite their popularity, they can be difficult to apply; all but the simplest classification or regression applications require specification and inference over complex covariance functions that do not admit simple analytical posteriors. This paper shows how to embed Gaussian processes in any higher-order probabilistic programming language, using an idiom based on memoization, and demonstrates its utility by implementing and extending classic and state-of-the-art GP applications. The interface to Gaussian processes, called gpmem, takes an arbitrary real-valued computational process as input and returns a statistical emulator that automatically improve as the original process is invoked and its input-output behavior is recorded. The flexibility of gpmem is illustrated via three applications: (i) robust GP regression with hierarchical hyper-parameter learning, (ii) discovering symbolic expressions from time-series data by fully Bayesian structure learning over kernels generated by a stochastic grammar, and (iii) a bandit formulation of Bayesian optimization with automatic inference and action selection. All applications share a single 50-line Python library and require fewer than 20 lines of probabilistic code each.

We propose multi-way, multilingual neural machine translation. The proposed approach enables a single neural translation model to translate between multiple languages, with a number of parameters that grows only linearly with the number of languages. This is made possible by having a single attention mechanism that is shared across all language pairs. We train the proposed multi-way, multilingual model on ten language pairs from WMT'15 simultaneously and observe clear performance improvements over models trained on only one language pair. In particular, we observe that the proposed model significantly improves the translation quality of low-resource language pairs.

The community detection problem refers to finding the underlying communities within a network using only the knowledge of the network topology [16, 31]. This paper considers the following probabilistic model for generating a network with underlying community structures: Suppose that out of a total of n vertices, rK of them are partitioned into r clusters of size K, and the remaining n rK vertices do not belong to any clusters (called outlier vertices); a random graph G is generated based on the cluster structure, where each pair of vertices is connected independently with probability p if they are in the same cluster or q otherwise. In particular, an outlier vertex is connected to any other vertex with probability q. This random graph ensemble is known as the planted cluster model [10] with parameters n, r, K N and p, q [0, 1] such that n rK. In particular, we call p and q the in-cluster and cross-cluster edge density, respectively. The planted cluster model encompasses several classical planted random graph models including planted clique [5], planted coloring [4], planted dense subgraph [6], planted partition [11], and the stochastic block model [23], which have been widely used for studying the community detection and graph partitioning problem (see, e.g., [26, 12, 30, 9] and the references therein). This paper was accepted to IEEE Transactions on Information Theory on January 3, 2016. The material in this paper was presented in part at the IEEE International Symposium on Information Theory, Hong Kong, June, 2015 [20].

This paper considers probabilistic estimation of a low-rank matrix from non-linear element-wise measurements of its elements. We derive the corresponding approximate message passing (AMP) algorithm and its state evolution. Relying on non-rigorous but standard assumptions motivated by statistical physics, we characterize the minimum mean squared error (MMSE) achievable information theoretically and with the AMP algorithm. Unlike in related problems of linear estimation, in the present setting the MMSE depends on the output channel only trough a single parameter - its Fisher information. We illustrate this striking finding by analysis of submatrix localization, and of detection of communities hidden in a dense stochastic block model. For this example we locate the computational and statistical boundaries that are not equal for rank larger than four.

The apparent stochasticity of in-vivo neural circuits has long been hypothesized to represent a signature of ongoing stochastic inference in the brain. More recently, a theoretical framework for neural sampling has been proposed, which explains how sample-based inference can be performed by networks of spiking neurons. One particular requirement of this approach is that the neural response function closely follows a logistic curve. Analytical approaches to calculating neural response functions have been the subject of many theoretical studies. In order to make the problem tractable, particular assumptions regarding the neural or synaptic parameters are usually made. However, biologically significant activity regimes exist which are not covered by these approaches: Under strong synaptic bombardment, as is often the case in cortex, the neuron is shifted into a high-conductance state (HCS) characterized by a small membrane time constant. In this regime, synaptic time constants and refractory periods dominate membrane dynamics. The core idea of our approach is to separately consider two different "modes" of spiking dynamics: burst spiking and transient quiescence, in which the neuron does not spike for longer periods. We treat the former by propagating the PDF of the effective membrane potential from spike to spike within a burst, while using a diffusion approximation for the latter. We find that our prediction of the neural response function closely matches simulation data. Moreover, in the HCS scenario, we show that the neural response function becomes symmetric and can be well approximated by a logistic function, thereby providing the correct dynamics in order to perform neural sampling. We hereby provide not only a normative framework for Bayesian inference in cortex, but also powerful applications of low-power, accelerated neuromorphic systems to relevant machine learning tasks.