We investigate what grade of sensor data is required for training an imitation-learning-based AV planner on human expert demonstration. Machine-learned planners [1] are very hungry for training data, which is usually collected using vehicles equipped with the same sensors used for autonomous operation [1]. This is costly and non-scalable. If cheaper sensors could be used for collection instead, data availability would go up, which is crucial in a field where data volume requirements are large and availability is small. We present experiments using up to 1000 hours worth of expert demonstration and find that training with 10x lower-quality data outperforms 1x AV-grade data in terms of planner performance (see Fig. 1). The important implication of this is that cheaper sensors can indeed be used. This serves to improve data access and democratize the field of imitation-based motion planning. Alongside this, we perform a sensitivity analysis of planner performance as a function of perception range, field-of-view, accuracy, and data volume, and reason about why lower-quality data still provide good planning results.

The field of statistical relational learning aims at unifying logic and probability to reason and learn from data. Perhaps the most successful paradigm in the field is probabilistic logic programming (PLP): the enabling of stochastic primitives in logic programming. While many systems offer inference capabilities, the more significant challenge is that of learning meaningful and interpretable symbolic representations from data. In that regard, inductive logic programming and related techniques have paved much of the way for the last few decades, but a major limitation of this exciting landscape is that only discrete features and distributions are handled. Many disciplines express phenomena in terms of continuous models. In this paper, we propose a new computational framework for inducing probabilistic logic programs over continuous and mixed discrete-continuous data. Most significantly, we show how to learn these programs while making no assumption about the true underlying density. Our experiments show the promise of the proposed framework.

There is a growing trend in building deep learning patient representations from health records to obtain a comprehensive view of a patient’s data for machine learning tasks. This paper proposes a reproducible approach to generate patient pathways from health records and to transform them into a machine-processable image-like structure useful for deep learning tasks. Based on this approach, we generated over a million pathways from FAIR synthetic health records and used them to train a convolutional neural network. Our initial experiments show the accuracy of the CNN on a prediction task is comparable or better than other autoencoders trained on the same data, while requiring significantly less computational resources for training. We also assess the impact of the size of the training dataset on autoencoders performances. The source code for generating pathways from health records is provided as open source.

We study the problem of the unsupervised learning of graphical models in mixed discrete-continuous domains. The problem of unsupervised learning of such models in discrete domains alone is notoriously challenging, compounded by the fact that inference is computationally demanding. The situation is generally believed to be significantly worse in discrete-continuous domains: estimating the unknown probability distribution of given samples is often limited in practice to a handful of parametric forms, and in addition to that, computing conditional queries needs to carefully handle low-probability regions in safety-critical applications. In recent years, the regime of tractable learning has emerged, which attempts to learn a graphical model that permits efficient inference. Most of the results in this regime are based on arithmetic circuits, for which inference is linear in the size of the obtained circuit. In this work, we show how, with minimal modifications, such regimes can be generalized by leveraging efficient density estimation schemes based on piecewise polynomial approximations. Our framework is realized on a recent computational abstraction that permits efficient inference for a range of queries in the underlying language. Our empirical results show that our approach is effective, and allows a study of the trade-off between the granularity of the learned model and its predictive power.