Querying Word Embeddings for Similarity and Relatedness
Fatemeh Torabi Asr, Robert Zinkov, and Michael N. Jones
In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), 2018
{ Abstract | BibTex | PDF }

Word embeddings obtained from neural network models such as Wor2Vec Skipgram have become popular representations of word meaning and have been evaluated on a variety of word similarity and relatedness norming data. Skipgram generates a set of word and context embeddings, the latter typically discarded after training. We demonstrate the usefulness of context embeddings in predicting asymmetric association between words from a recently published dataset of production norms. Our findings suggest that humans respond with words closer to the cue within the context embedding space (rather than the word embedding space), when asked to generate thematically related words.

@inproceedings{asr2018embeddings,
  author = {Fatemeh Torabi Asr and Robert Zinkov and Michael N. Jones},
  title = {Querying Word Embeddings for Similarity and Relatedness},
  booktitle = {Proceedings of the 2018 Conference of the North
                 American Chapter of the Association for Computational
                 Linguistics: Human Language Technologies (NAACL-HLT)},
  year = {2018},
  publisher = {Association for Computational Linguistics},
  pages = {675--684},
}

Automating Expectation Maximixation
Robert Zinkov
In NIPS Workshop on Advances in Approximate Bayesian Inference, 2017
{ Abstract | BibTex | PDF }

Existing probabilistic programming systems frequently treat machine learning problems as purely inference tasks. But real-world problems often require a novel combination of inference and optimization. For example when implementing Expectation Maximization for a model, one must alternate between performing an inference and an optimization step. We show that by treating optimization as a program transformation it is possible to implement Expectation Maximization in a generic and composable way.

@inproceedings{zinkov2017autoem,
  title = {Automating Expectation Maximixation},
  author = {Robert Zinkov},
  booktitle = {NIPS Workshop on Advances in Approximate Bayesian Inference},
  year = {2017},
}

End-to-end Training of Differentiable Pipelines Across Machine Learning Frameworks
Mitar Milutinovic, Atılım Güneş Baydin, Robert Zinkov, William Harvey, Dawn Song, Frank Wood, and Wade Shen
In NIPS Workshop on Autodiff, 2017
{ Abstract | BibTex }

In this work we present a unified interface and methodology for performing end-to-end gradient-based refinement of pipelines of differentiable machine-learning primitives. This is distinguished from recent interoperability efforts such as the Open Neural Network Exchange (ONNX) format and other language-centric cross-compilation approaches in that the final pipeline does not need to be implemented nor trained in the same language nor cross-compiled into any single language; in other words, primitives may be written and pre-trained in PyTorch, TensorFlow, Caffe, scikit-learn or any of the other popular machine learning frameworks and fine-tuned end-to-end while being executed directly in their host frameworks. Provided primitives expose our proposed interface, it is possible to automatically compose all such primitives and refine them based on an end-to-end loss.

@inproceedings{milutinovic2017pipeline,
  title = {End-to-end Training of Differentiable Pipelines Across Machine Learning Frameworks},
  author = {Mitar Milutinovic and Atılım Güneş Baydin and Robert
            Zinkov and William Harvey and Dawn Song and Frank
            Wood and Wade Shen},
  booktitle = {NIPS Workshop on Autodiff},
  year = {2017},
}

Principled Inference Networks in Deep Generative Models
Stefan Webb, Adam Golinski, Robert Zinkov, and Frank Wood
In NIPS Workshop on Bayesian Deep Learning, 2017
{ Abstract | BibTex | PDF }

In learning deep generative models, the encoder for variational inference is typically formed in an ad hoc manner with a structure and parametrization analogous to the forward model. Our chief insight is that this results in coarse approximations to the posterior, and that the d-separation properties of the BN structure of the forward model should be used, in a principled way, to produce ones that are faithful to the posterior, for which we introduce the novel Compact Minimal I-map (CoMI) algorithm. Applying our method to common models reveals that standard encoder design choices lack many important edges, and through experiments we demonstrate that modelling these edges is important for optimal learning. We show how using a faithful encoder is crucial when modelling with continuous relaxations of categorical distributions.

@inproceedings{webb2017deepbayes,
  title = {Principled Inference Networks in Deep Generative Models},
  author = {Stefan Webb and Adam Golinski and Robert Zinkov and Frank Wood},
  booktitle = {NIPS Workshop on Bayesian Deep Learning},
  year = {2017},
}

Composing inference algorithms as program transformations
Robert Zinkov and Chung-chieh Shan
In Proceedings of Uncertainty in Artificial Intelligence, 2017
{ Abstract | BibTex | PDF }

Probabilistic inference procedures are usually coded painstakingly from scratch, for each target model and each inference algorithm. We reduce this effort by generating inference procedures from models automatically. We make this code generation modular by decomposing inference algorithms into reusable program-to-program transformations. These transformations perform exact inference as well as generate probabilistic programs that compute expectations, densities, and MCMC samples. The resulting inference procedures are about as accurate and fast as other probabilistic programming systems on real-world problems.

@inproceedings{zinkov-composing,
  title = {Composing inference algorithms as program transformations},
  author = {Robert Zinkov and Shan, Chung-chieh},
  booktitle = {Proceedings of Uncertainty in Artificial Intelligence},
  year = {2017},
}

Using Synthetic Data to Train Neural Networks is Model-Based Reasoning
Tuan Anh Le, Atılım Güneş Baydin, Robert Zinkov, and Frank Wood
In International Joint Conference on Neural Networks (IJCNN), 2017
{ Abstract | BibTex | PDF }

We draw a formal connection between using synthetic training data to optimize neural network parameters and approximate, Bayesian, model-based reasoning. In particular, training a neural network using synthetic data can be viewed as learning a proposal distribution generator for approximate inference in the synthetic-data generative model. We demonstrate this connection in a recognition task where we develop a novel Captcha-breaking architecture and train it using synthetic data, demonstrating both state-of-the-art performance and a way of computing task-specific posterior uncertainty. Using a neural network trained this way, we also demonstrate successful breaking of real-world Captchas currently used by Facebook and Wikipedia. Reasoning from these empirical results and drawing connections with Bayesian modeling, we discuss the robustness of synthetic data results and suggest important considerations for ensuring good neural network generalization when training with synthetic data.

@article{le2017using,
  title = {Using Synthetic Data to Train Neural Networks is Model-Based Reasoning},
  author = {Le, Tuan Anh and Baydin, At{\i}l{\i}m G{\"u}ne\c{s} and Zinkov, Robert and Wood, Frank},
  booktitle = {International Joint Conference on Neural Networks ({IJCNN})},
  pages = {3514--3521},
  year = {2017},
}

Probabilistic Inference by Program Transformation in Hakaru (System Description)
Praveen Narayanan, Jacques Carette, Wren Romano, Chung-chieh Shan, and Robert Zinkov
In Functional and Logic Programming: 13th International Symposium, FLOPS 2016, 2016
{ Abstract | BibTex | PDF }

We present Hakaru, a new probabilistic programming system that allows modular and composable reuse of distributions, queries, and inference algorithms, all expressed in a single language of measures. The system implements two automatic and semantics-preserving transformations — disintegration, for calculating conditional distributions, and simplification, for subsuming exact inference via computer algebra. We show how these features work together by describing the ideal workflow of a Hakaru user on two small problems. We highlight details of the system’s design, and discuss our experience building it.

@inproceedings{hakaru-system-description,
  author = {Praveen Narayanan and Jacques Carette and Wren Romano and
            Chung-chieh Shan and Robert Zinkov},
  title = {Probabilistic Inference by Program Transformation in {Hakaru} (System Description)},
  booktitle = {Functional and Logic Programming: 13th International Symposium, {FLOPS} 2016},
  year = {2016},
  address = {Berlin},
  publisher = {Springer},
  series = {{L}ecture {N}otes in {C}omputer {S}cience},
  number = {9613},
  pages = {62--79},
}

Building blocks for exact and approximate inference
Jacques Carette, Praveen Narayanan, Wren Romano, Chung-chieh Shan, and Robert Zinkov
In NIPS Workshop on Black Box Learning and Inference, 2015
{ Abstract | BibTex | PDF }

A promising approach to implementing black-box inference is to search a space of inference strategies. However, a search space has not been defined to date that includes combinations of exact and approximate inference methods, such as Rao-Blackwellized Metropolis-Hastings and Gibbs sampling of continuous distributions. We present a handful of inference building blocks that not only constitute black-box inference methods themselves but also generate a search space of inference strategies that includes such combinations. The building blocks include simplifying and sampling from distributions, based in turn on calculating their expectation, disintegration, and density. We have implemented these building blocks as probabilistic-program transformations and specified them in terms of a distribution semantics.

@article{caretteblocks,
  title = {Building blocks for exact and approximate inference},
  author = {Carette, Jacques and Narayanan, Praveen and Romano, Wren
            and Shan, Chung-chieh and Zinkov, Robert},
  booktitle = {NIPS Workshop on Black Box Learning and Inference},
  year = {2015},
}

Designing a MCMC library for Probabilistic Programming
Robert Zinkov, Praveen Narayanan, and Chung-chieh Shan
In PLDI Workshop on Probabilistic and Approximate Computing, 2014
{ Abstract | BibTex }

Markov Chain Monte Carlo methods have proven effective for performing Bayesian inference. Unfortunately, writing an MCMC sampler is tedious and error-prone. We present a design for a combinator library which shows how samplers may be described as the composition of transition probability kernels. This representation makes it easy to extend an existing sampler and highlights the relationships between existing MCMC methods.

@article{zinkov2014approx,
  title = {Designing a MCMC library for Probabilistic Programming},
  author = {Robert Zinkov and Praveen Narayanan and Chung-chieh Shan},
  booktitle = {PLDI Workshop on Probabilistic and Approximate Computing},
  year = {2014},
}

Sensitivity analysis for distributed optimization with resource constraints
Emma Bowring, Zhengyu Yin, Rob Zinkov, and Milind Tambe
In 8th International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2009
{ Abstract | BibTex | PDF }

Previous work in multiagent coordination has addressed the challenge of planning in domains where agents must optimize a global goal, while satisfying local resource constraints. However, the imposition of resource constraints naturally raises the question of whether the agents could significantly improve their team performance if a few more resources were made available. Sensitivity analysis aims to answer that question. This paper focuses on sensitivity analysis in the context of the distributed coordination framework, Multiply-Constrained DCOP (MC-DCOP). There are three main challenges in performing sensitivity analysis: (i) to perform it in a distributed fashion, (ii) to avoid re-solving an NP-hard MC-DCOP optimization from scratch, and (iii) to avoid considering unproductive uses for extra resources. To meet these challenges, this paper presents three types of locally optimal algorithms: link analysis, local reoptimization and local constraint propagation. These algorithms are distributed and avoid redundant computation by ascertaining just the effects of local perturbations on the original problem. Deploying our algorithms on a large number of MC-DCOP problems revealed several results. While our cheapest algorithm successfully identified quality improvements for a few problems, our more complex techniques were necessary to identify the best uses for additional resources. Furthermore, we identified two heuristics that can help identify a priori which agents might benefit most from additional resources: density rank, which works well when nodes received identical resources and remaining resource rank, which works well when nodes received resources based on the number of neighbors they had.

@inproceedings{BowringYZT09,
  author = {Emma Bowring and
               Zhengyu Yin and
               Rob Zinkov and
               Milind Tambe},
  title = {Sensitivity analysis for distributed optimization with resource constraints},
  booktitle = {8th International Joint Conference on Autonomous Agents and Multiagent
               Systems ({AAMAS})},
  pages = {633--640},
  year = {2009},
}

Potential-based Shaping in Model-based Reinforcement Learning
John Asmuth, Michael L. Littman, and Robert Zinkov
In Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence (AAAI), 2008
{ Abstract | BibTex | PDF }

Potential-based shaping was designed as a way of introducing background knowledge into model-free reinforcement-learning algorithms. By identifying states that are likely to have high value, this approach can decrease experience complexity—the number of trials needed to find near-optimal behavior. An orthogonal way of decreasing experience complexity is to use a model-based learning approach, building and exploiting an explicit transition model. In this paper, we show how potential-based shaping can be redefined to work in the model-based setting to produce an algorithm that shares the benefits of both ideas.

@inproceedings{asmuth2008potential,
  title = {Potential-based Shaping in Model-based Reinforcement Learning},
  author = {Asmuth, John and Littman, Michael L. and Zinkov, Robert},
  booktitle = {Proceedings of the Twenty-Third {AAAI} Conference on Artificial Intelligence
               ({AAAI})},
  pages = {604--609},
  year = {2008},
}