Addressing Discrepancies in Semantic and Visual Alignment in Neural Networks

For the task of image classification, neural networks primarily rely on visual patterns. In robust networks, we would expect for visually similar classes to be represented similarly. We consider the problem of when semantically similar classes are visually dissimilar, and when visual similarity is present among non-similar classes. We propose a data augmentation technique with the goal of better aligning semantically similar classes with arbitrary (non-visual) semantic relationships. We leverage recent work in diffusion-based semantic mixing to generate semantic hybrids of two classes, and these hybrids are added to the training set as augmented data. We evaluate whether the method increases semantic alignment by evaluating model performance on adversarially perturbed data, with the idea that it should be easier for an adversary to switch one class to a similarly represented class. Results demonstrate that there is an increase in alignment of semantically similar classes when using our proposed data augmentation method.

Evaluating the Capabilities of Multi-modal Reasoning Models with Synthetic Task Data

The impressive advances and applications of large language and joint language-and-visual understanding models has led to an increased need for methods of probing their potential reasoning capabilities. However, the difficulty of gather naturally-occurring data for complex multi-modal reasoning tasks bottlenecks the evaluation of AI methods on tasks which are not already covered by an academic dataset. In this work, we leverage recent advances in high resolution text-to-image generation to develop a framework for generating evaluation data for multi-modal reasoning tasks. We apply this framework to generate context-dependent anomaly data, creating a synthetic dataset on a challenging task which is not well covered by existing datasets. We benchmark the performance of a state-of-the-art visual question answering (VQA) model against data generated with this method, and demonstrate that while the task is tractable, the model performs significantly worse on the context-dependent anomaly detection task than on standard VQA tasks.

Addressing Mistake Severity in Neural Networks with Semantic Knowledge

Robustness in deep neural networks and machine learning algorithms in general is an open research challenge. In particular, it is difficult to ensure algorithmic performance is maintained on out-of-distribution inputs or anomalous instances that cannot be anticipated at training time. Embodied agents will be deployed in these conditions, and are likely to make incorrect predictions. An agent will be viewed as untrustworthy unless it can maintain its performance in dynamic environments. Most robust training techniques aim to improve model accuracy on perturbed inputs; as an alternate form of robustness, we aim to reduce the severity of mistakes made by neural networks in challenging conditions. We leverage current adversarial training methods to generate targeted adversarial attacks during the training process in order to increase the semantic similarity between a model's predictions and true labels of misclassified instances. Results demonstrate that our approach performs better with respect to mistake severity compared to standard and adversarially trained models. We also find an intriguing role that non-robust features play with regards to semantic similarity.

Commander's Intent: A Dataset and Modeling Approach for Human-AI Task Specification in Strategic Play

Effective Human-AI teaming requires the ability to communicate the goals of the team and constraints under which you need the agent to operate. Providing the ability to specify the shared intent or operation criteria of the team can enable an AI agent to perform its primary function while still being able to cater to the specific desires of the current team. While significant work has been conducted to instruct an agent to perform a task, via language or demonstrations, prior work lacks a focus on building agents which can operate within the parameters specified by a team. Worse yet, there is a dearth of research pertaining to enabling humans to provide their specifications through unstructured, naturalist language. In this paper, we propose the use of goals and constraints as a scaffold to modulate and evaluate autonomous agents. We contribute to this field by presenting a novel dataset, and an associated data collection protocol, which maps language descriptions to goals and constraints corresponding to specific strategies developed by human participants for the board game Risk. Leveraging state-of-the-art language models and augmentation procedures, we develop a machine learning framework which can be used to identify goals and constraints from unstructured strategy descriptions. To empirically validate our approach we conduct a human-subjects study to establish a human-baseline for our dataset. Our results show that our machine learning architecture is better able to interpret unstructured language descriptions into strategy specifications than human raters tasked with performing the same machine translation task (F(1,272.53) = 17.025, p < 0.001).

Context-Dependent Anomaly Detection with Knowledge Graph Embedding Models

Increasing the semantic understanding and contextual awareness of machine learning models is important for improving robustness and reducing susceptibility to data shifts. In this work, we leverage contextual awareness for the anomaly detection problem. Although graphed-based anomaly detection has been widely studied, context-dependent anomaly detection is an open problem and without much current research. We develop a general framework for converting a context-dependent anomaly detection problem to a link prediction problem, allowing well-established techniques from this domain to be applied. We implement a system based on our framework that utilizes knowledge graph embedding models and demonstrates the ability to detect outliers using context provided by a semantic knowledge base. We show that our method can detect context-dependent anomalies with a high degree of accuracy and show that current object detectors can detect enough classes to provide the needed context for good performance within our example domain.