Alexander Scott Warstadt

Clément Guerner, Tianyu Liu, Anej Svete et al. · allen-ai, eth-zurich

The linear subspace hypothesis (Bolukbasi et al., 2016) states that, in a language model's representation space, all information about a concept such as verbal number is encoded in a linear subspace. Prior work has relied on auxiliary classification tasks to identify and evaluate candidate subspaces that might give support for this hypothesis. We instead give a set of intrinsic criteria which characterize an ideal linear concept subspace and enable us to identify the subspace using only the language model distribution. Our information-theoretic framework accounts for spuriously correlated features in the representation space (Kumar et al., 2022) by reconciling the statistical notion of concept information and the geometric notion of how concepts are encoded in the representation space. As a byproduct of this analysis, we hypothesize a causal process for how a language model might leverage concepts during generation. Empirically, we find that linear concept erasure is successful in erasing most concept information under our framework for verbal number as well as some complex aspect-level sentiment concepts from a restaurant review dataset. Our causal intervention for controlled generation shows that, for at least one concept across two languages models, the concept subspace can be used to manipulate the concept value of the generated word with precision.

Kevin Du, Lucas Torroba Hennigen, Niklas Stoehr et al. · eth-zurich

Many popular feature-attribution methods for interpreting deep neural networks rely on computing the gradients of a model's output with respect to its inputs. While these methods can indicate which input features may be important for the model's prediction, they reveal little about the inner workings of the model itself. In this paper, we observe that the gradient computation of a model is a special case of a more general formulation using semirings. This observation allows us to generalize the backpropagation algorithm to efficiently compute other interpretable statistics about the gradient graph of a neural network, such as the highest-weighted path and entropy. We implement this generalized algorithm, evaluate it on synthetic datasets to better understand the statistics it computes, and apply it to study BERT's behavior on the subject-verb number agreement task (SVA). With this method, we (a) validate that the amount of gradient flow through a component of a model reflects its importance to a prediction and (b) for SVA, identify which pathways of the self-attention mechanism are most important.

Theodor Amariucai, Alex Warstadt

In contrast to children, language models (LMs) exhibit considerably inferior data efficiency when acquiring language. In this submission to the BabyLM Challenge (Warstadt et al., 2023), we test the hypothesis that this data efficiency gap is partly caused by a lack of multimodal input and grounding in the learning environment of typical language models. Although previous work looking into this question found that multimodal training can even harm language-only performance, we speculate that these findings can be attributed to catastrophic forgetting of complex language due to fine-tuning on captions data. To test our hypothesis, we perform an ablation study on FLAVA (Singh et al., 2022), a multimodal vision-and-language model, independently varying the volume of text and vision input to quantify how much text data (if any) can be offset by vision at different data scales. We aim to limit catastrophic forgetting through a multitask pretraining regime that includes unimodal text-only tasks and data sampled from WiT, the relatively diverse Wikipedia-based dataset (Srinivasan et al., 2021). Our results are largely negative: Multimodal pretraining does not harm our models' language performance but does not consistently help either. That said, our conclusions are limited by our having been able to conduct only a small number of runs. While we must leave open the possibility that multimodal input explains some of the gap in data efficiency between LMs and humans, positive evidence for this hypothesis will require better architectures and techniques for multimodal training.