Abstract--Procedural content generation via machine learning (PCGML) is typically framed as the task of fitting a generative model to full-scale examples of a desired content distribution. This approach presents a fundamental tension: the more design effort expended to produce detailed training examples for shaping a generator, the lower the return on investment from applying PCGML in the first place. In response, we propose the use of discriminative models (which capture the validity of a design rather the distribution of the content) trained on positive and negative examples. Through a modest modification of WaveFunctionCollapse, a commercially-adopted PCG approach that we characterize as using elementary machine learning, we demonstrate a new mode of control for learning-based generators. We demonstrate how an artist might craft a focused set of additional positive and negative examples by critique of the generator's previous outputs. This interaction mode bridges PCGML with mixed-initiative design assistance tools by working with a machine to define a space of valid designs rather than just one new design. Procedural Content Generation via Machine Learning (PCGML) is the recent term for the strategy of controlling content generators using examples [1]. Existing PCGML approaches train their statistical models based on preexisting artist-provided samples of the desired content. However, there is a fundamental tension here: machine learning often works better with more training data, but the effort to produce quality training data is frequently costly enough that the artists might be better off just making the content themselves. Rather than attempting to train a generative statistical model (capturing the distribution of desired content), we focus on applying discriminative learning. In discriminative learning, the model learns to judge whether a candidate content artifact would be valid or desirable, but it does not learn how to generate candidates. Pairing a discriminative model with a preexisting content generator, we realize example-driven generation that can be influenced by both positive and negative examples of valid design patterns.

We study the sample complexity of semi-supervised learning (SSL) and introduce new assumptions based on the mismatch between a mixture model learned from unlabeled data and the true mixture model induced by the (unknown) class conditional distributions. Under these assumptions, we establish an $\Omega(K\log K)$ labeled sample complexity bound without imposing parametric assumptions, where $K$ is the number of classes. Our results suggest that even in nonparametric settings it is possible to learn a near-optimal classifier using only a few labeled samples. Unlike previous theoretical work which focuses on binary classification, we consider general multiclass classification ($K>2$), which requires solving a difficult permutation learning problem. This permutation defines a classifier whose classification error is controlled by the Wasserstein distance between mixing measures, and we provide finite-sample results characterizing the behaviour of the excess risk of this classifier. Finally, we describe three algorithms for computing these estimators based on a connection to bipartite graph matching, and perform experiments to illustrate the superiority of the MLE over the majority vote estimator.

Many modern machine learning classifiers are shown to be vulnerable to adversarial perturbations of the instances that can "evade" the classifier and get misclassified. Despite a massive amount of work focusing on making classifiers robust, the task seems quite challenging. In this work, through a theoretical study, we investigate the adversarial risk and robustness of classifiers and draw a connection to the well-known phenomenon of "concentration of measure" in metric measure spaces. We show that if the metric probability space of the test instance is concentrated, any classifier with some initial constant error is inherently vulnerable to adversarial perturbations. One class of concentrated metric probability spaces are the so-called Levy families that include many natural distributions. In this special case, our attacks only need to perturb the test instance by at most $O(\sqrt n)$ to make it misclassified, where $n$ is the data dimension. Using our general result about Levy instance spaces, we first recover as special case some of the previously proved results about the existence of adversarial examples. However, many more Levy families are known for which we immediately obtain new attacks finding adversarial examples (e.g., product distribution under the Hamming distance). Finally, we show that concentration of measure for product spaces implies the existence of so called "poisoning" attacks in which the adversary tampers with the training data with the goal of increasing the error of the classifier. We show that for any deterministic learning algorithm that uses $m$ training examples, there is an adversary who substitutes $O(\sqrt m)$ of the examples with other (still correctly labeled) ones and can almost fully degrade the confidence parameter of any PAC learning algorithm or alternatively increase the risk to almost 1 if the adversary also knows the final test example.

We investigate how generative adversarial nets (GANs) can help semi-supervised learning on graphs. We first provide insights on working principles of adversarial learning over graphs and then present GraphSGAN, a novel approach to semi-supervised learning on graphs. In GraphSGAN, generator and classifier networks play a novel competitive game. At equilibrium, generator generates fake samples in low-density areas between subgraphs. In order to discriminate fake samples from the real, classifier implicitly takes the density property of subgraph into consideration. An efficient adversarial learning algorithm has been developed to improve traditional normalized graph Laplacian regularization with a theoretical guarantee. Experimental results on several different genres of datasets show that the proposed GraphSGAN significantly outperforms several state-of-the-art methods. GraphSGAN can be also trained using mini-batch, thus enjoys the scalability advantage.

These representations are able to capture the variability of the demonstrations from a teacher as a probability distribution over trajectories, providing a sensible region of exploration and the ability to adapt to changes in the robot environment. However, to be able to capture variability and correlations between different joints, a probabilistic movement primitive requires the estimation of a larger number of parameters compared to their deterministic counterparts, that focus on modeling only the mean behavior. In this paper, we make use of prior distributions over the parameters of a probabilistic movement primitive to make robust estimates of the parameters with few training instances. In addition, we introduce general purpose operators to adapt movement primitives in joint and task space. The proposed training method and adaptation operators are tested in a coffee preparation and in robot table tennis task. In the coffee preparation task we evaluate the generalization performance to changes in the location of the coffee grinder and brewing chamber in a target area, achieving the desired behavior after only two demonstrations. In the table tennis task we evaluate the hit and return rates, outperforming previous approaches while using fewer task specific heuristics.

In this paper, we study a generic learning framework for building robust text classification model that achieves accuracy comparable to standard full models under test-time budget constraints. Our approach learns a selector to identify words that are relevant to the prediction tasks and only passes these words to the classifier for processing. The selector is trained jointly with the classifier and directly learns to incorporate with the classifier. We further propose a data aggregation scheme to improve the robustness of the classifier. Our learning framework is general and can be incorporated with any type of text classification model. On real-world data, we show that the proposed approach improves the performance of a given classifier and speeds up the model with a mere loss in accuracy performance.

One way to interpret neural model predictions is to highlight the most important input features---for example, a heatmap visualization over the words in an input sentence. In existing interpretation methods for NLP, a word's importance is determined by either input perturbation---measuring the decrease in model confidence when that word is removed---or by the gradient with respect to that word. To understand the limitations of these methods, we use input reduction, which iteratively removes the least important word from the input. This exposes pathological behaviors of neural models: the remaining words appear nonsensical to humans and are not the ones determined as important by interpretation methods. As we confirm with human experiments, the reduced examples lack information to support the prediction of any label, but models still make the same predictions with high confidence. To explain these counterintuitive results, we draw connections to adversarial examples and confidence calibration: pathological behaviors reveal difficulties in interpreting neural models trained with maximum likelihood. To mitigate their deficiencies, we fine-tune the models by encouraging high entropy outputs on reduced examples. Fine-tuned models become more interpretable under input reduction without accuracy loss on regular examples.

Recent machine learning strategies for segmentation tasks have shown great ability when trained on large pixel-wise annotated image datasets. It remains a major challenge however to aggregate such datasets, as the time and monetary cost associated with collecting extensive annotations is extremely high. This is particularly the case for generating precise pixel-wise annotations in video and volumetric image data. To this end, this work presents a novel framework to produce pixel-wise segmentations using minimal supervision. Our method relies on 2D point supervision, whereby a single 2D location within an object of interest is provided on each image of the data. Our method then estimates the object appearance in a semi-supervised fashion by learning object-image-specific features and by using these in a semi-supervised learning framework. Our object model is then used in a graph-based optimization problem that takes into account all provided locations and the image data in order to infer the complete pixel-wise segmentation. In practice, we solve this optimally as a tracking problem using a K-shortest path approach. Both the object model and segmentation are then refined iteratively to further improve the final segmentation. We show that by collecting 2D locations using a gaze tracker, our approach can provide state-of-the-art segmentations on a range of objects and image modalities (video and 3D volumes), and that these can then be used to train supervised machine learning classifiers.

In recent years, non-monotonic Inductive Logic Programming has received growing interest. Specifically, several new learning frameworks and algorithms have been introduced for learning under the answer set semantics, allowing the learning of common-sense knowledge involving defaults and exceptions, which are essential aspects of human reasoning. In this paper, we present a noise-tolerant generalisation of the learning from answer sets framework. We evaluate our ILASP3 system, both on synthetic and on real datasets, represented in the new framework. In particular, we show that on many of the datasets ILASP3 achieves a higher accuracy than other ILP systems that have previously been applied to the datasets, including a recently proposed differentiable learning framework.

Over the past decades, researchers and ML practitioners have come up with better and better ways to build, understand and improve the quality of ML models, but mostly under the key assumption that the training data is distributed identically to the testing data. In many real-world applications, however, some potential training examples are unknown to the modeler, due to sample selection bias or, more generally, covariate shift, i.e., a distribution shift between the training and deployment stage. The resulting discrepancy between training and testing distributions leads to poor generalization performance of the ML model and hence biased predictions. We provide novel algorithms that estimate the number and properties of these unknown training examples---unknown unknowns. This information can then be used to correct the training set, prior to seeing any test data. The key idea is to combine species-estimation techniques with data-driven methods for estimating the feature values for the unknown unknowns. Experiments on a variety of ML models and datasets indicate that taking the unknown examples into account can yield a more robust ML model that generalizes better.