Michael Wilson

Michael Wilson, Jackson Petty, Robert Frank

Language models are typically evaluated on their success at predicting the distribution of specific words in specific contexts. Yet linguistic knowledge also encodes relationships between contexts, allowing inferences between word distributions. We investigate the degree to which pre-trained Transformer-based large language models (LLMs) represent such relationships, focusing on the domain of argument structure. We find that LLMs perform well in generalizing the distribution of a novel noun argument between related contexts that were seen during pre-training (e.g., the active object and passive subject of the verb spray), succeeding by making use of the semantically-organized structure of the embedding space for word embeddings. However, LLMs fail at generalizations between related contexts that have not been observed during pre-training, but which instantiate more abstract, but well-attested structural generalizations (e.g., between the active object and passive subject of an arbitrary verb). Instead, in this case, LLMs show a bias to generalize based on linear order. This finding points to a limitation with current models and points to a reason for which their training is data-intensive.s reported here are available at https://github.com/clay-lab/structural-alternations.

Jackson Petty, Michael Wilson, Robert Frank

How is knowledge of position-role mappings in natural language learned? We explore this question in a computational setting, testing whether a variety of well-performing pertained language models (BERT, RoBERTa, and DistilBERT) exhibit knowledge of these mappings, and whether this knowledge persists across alternations in syntactic, structural, and lexical alternations. In Experiment 1, we show that these neural models do indeed recognize distinctions between theme and recipient roles in ditransitive constructions, and that these distinct patterns are shared across construction type. We strengthen this finding in Experiment 2 by showing that fine-tuning these language models on novel theme- and recipient-like tokens in one paradigm allows the models to make correct predictions about their placement in other paradigms, suggesting that the knowledge of these mappings is shared rather than independently learned. We do, however, observe some limitations of this generalization when tasks involve constructions with novel ditransitive verbs, hinting at a degree of lexical specificity which underlies model performance.

Luigi Del Debbio, Joe Marsh Rossney, Michael Wilson

A trivializing map is a field transformation whose Jacobian determinant exactly cancels the interaction terms in the action, providing a representation of the theory in terms of a deterministic transformation of a distribution from which sampling is trivial. Recently, a proof-of-principle study by Albergo, Kanwar and Shanahan [arXiv:1904.12072] demonstrated that approximations of trivializing maps can be `machine-learned' by a class of invertible, differentiable neural models called \textit{normalizing flows}. By ensuring that the Jacobian determinant can be computed efficiently, asymptotically exact sampling from the theory of interest can be performed by drawing samples from a simple distribution and passing them through the network. From a theoretical perspective, this approach has the potential to become more efficient than traditional Markov Chain Monte Carlo sampling techniques, where autocorrelations severely diminish the sampling efficiency as one approaches the continuum limit. A major caveat is that it is not yet understood how the size of models and the cost of training them is expected to scale. As a first step, we have conducted an exploratory scaling study using two-dimensional $φ^4$ with up to $20^2$ lattice sites. Although the scope of our study is limited to a particular model architecture and training algorithm, initial results paint an interesting picture in which training costs grow very quickly indeed. We describe a candidate explanation for the poor scaling, and outline our intentions to clarify the situation in future work.

Tegan Emerson, Sarah Tymochko, George Stantchev et al.

Detecting small targets at range is difficult because there is not enough spatial information present in an image sub-region containing the target to use correlation-based methods to differentiate it from dynamic confusers present in the scene. Moreover, this lack of spatial information also disqualifies the use of most state-of-the-art deep learning image-based classifiers. Here, we use characteristics of target tracks extracted from video sequences as data from which to derive distinguishing topological features that help robustly differentiate targets of interest from confusers. In particular, we calculate persistent homology from time-delayed embeddings of dynamic statistics calculated from motion tracks extracted from a wide field-of-view video stream. In short, we use topological methods to extract features related to target motion dynamics that are useful for classification and disambiguation and show that small targets can be detected at range with high probability.