Jiaji Huang – Research Summary

Pretrained Language Models (LM), e.g., BERT, are very successful. However, it is still mysterious why they can generalize so well. The joint force of interpretability and language modeling has created a new research community called Bertology.

Probing tasks are widely used in Bertology. That is, applying an LM to a probing task. If the LM works well, then we would agree that task-relevant information (e.g., syntatic, semantic) is captured. In other words, “probing” is an extrinsic approach. It attempts to understand a blackbox (the LM) via another box (the probing task) that is “less black”.

Different from probing, we take a more intrinsic approach, through comparing and contrasting the LMs. We discover interesting similarity patterns among a zoo of LMs. The left figure is one visualization of the similarities among 34 language models (downloadable from huggingface model hub). The pattern hints on many possibilities of practical interests, e.g., picking the right LM for your tasks (see our Neurips2021 paper). In addition, intriguing similarity pattern also exists inside the building blocks of an LM. For example, among attention heads. One immediate follow-up question is, “can we prune the LMs more effectively by taking advantage of this similarity pattern?”

This onging effort connects to many fields in machine learning and NLP. Other than the above mentioned applications, it also seeks to answer fundamental questions in transferability and learning dynamics.

Nearest Neighbor (NN) search may be the most general and widely applied retrieval method in multi-dimensional feature spaces. However, NN is often degraded by a phenomenon called hubness. Hubness is a tendency in high dimensional space that some data points, called “hubs”, are suspiciously close to many others. As a result, NN search may retrieve these “hubby” points more often than they should be.

Earlier works, Inverted SoFtmax (ISF, 2017) and Cross-domain Similarity Local Scaling (CSLS, 2018) are heurestics that mitigate hubness and improve upon NN for bilingual lexicon induction tasks. Nevertheless, still missing are a rigorious formulation and a thorough understanding of the hubness-reduction problem.

The proposed Hubless Nearest Neighbor (HNN) search hinges upon an Equal Preference Assumption. We observe that the assumption approximately holds for bilingual lexicon induction between languages that share their origins. Essentially, it is a constraint that all items in the gallery are equally likely to be retrieved. The retrieval task is then re-formulated into a linear assignment problem. By relaxing its objective function, we can further derive an efficient solver that is parallelizable over GPUs.

Deep neural networks have proved very successful in domains where large training sets are available, but when the number of training samples is small, their performance suffers from overfitting. To reduce overfitting, people have proposed various data independent regularizations, e.g. weight decay, dropout.