The full-size MLPs and the projection layers in attention introduce tremendous model sizes of large language models (LLMs), consuming extensive computational resources in pre-training. We empirically observe that the activations of pre-trained LLMs exhibit low-rank property. Motivated by such observations, we propose **CoLA** and its memory-efficient implementation, **CoLA-M**, to replace these full-size layers with compute-efficient **auto-encoders** that naturally enforce low-rank activations throughout training. This fundamental architectural change eliminates the activation redundancy and significantly boosts model capacity and training efficiency. Experiments on LLaMA models with 60 million to 7 billion parameters show that CoLA reduces the computing cost by 2\pmb{\times} and improves training throughput by 1.86\pmb{\times} while maintaining full-rank level performance. CoLA-M further squeezes memory cost without sacrificing throughput, offering a pre-training approach with collectively superior parameter, computing, and memory efficiency. The LLMs produced are also 2\pmb{\times} smaller, enabling faster inference with lower memory cost on resource-constrained platforms.

Traditional (unstructured) pruning methods for a Transformer model focus on regularizing the individual weights by penalizing them toward zero. In this work, we explore spectral-normalized identity priors (SNIP), a structured pruning approach which penalizes an entire residual module in a Transformer model toward an identity mapping. Our method identifies and discards unimportant non-linear mappings in the residual connections by applying a thresholding operator on the function norm, and is applicable to any structured module including a single attention head, an entire attention blocks, or a feed-forward subnetwork. Furthermore, we introduce spectral normalization to stabilize the distribution of the post-activation values of the Transformer layers, further improving the pruning effectiveness of the proposed methodology. We conduct experiments with BERT on 5 GLUE benchmark tasks to demonstrate that SNIP achieves effective pruning results while maintaining comparable performance. Specifically, we improve the performance over the state-of-the-art by 0.5 to 1.0% on average at 50% compression ratio.

Deep neural networks for machine comprehension typically utilizes only word or character embeddings without explicitly taking advantage of structured linguistic information such as constituency trees and dependency trees. In this paper, we propose structural embedding of syntactic trees (SEST), an algorithm framework to utilize structured information and encode them into vector representations that can boost the performance of algorithms for the machine comprehension. We evaluate our approach using a state-of-the-art neural attention model on the SQuAD dataset. Experimental results demonstrate that our model can accurately identify the syntactic boundaries of the sentences and extract answers that are syntactically coherent over the baseline methods.

In this paper, we describe our participation in phase B of task 5b of the fifth edition of the annual BioASQ challenge, which includes answering factoid, list, yes-no and summary questions from biomedical data. We describe our techniques with an emphasis on ideal answer generation, where the goal is to produce a relevant, precise, non-redundant, query-oriented summary from multiple relevant documents. We make use of extractive summarization techniques to address this task and experiment with different biomedical ontologies and various algorithms including agglomerative clustering, Maximum Marginal Relevance (MMR) and sentence compression. We propose a novel word embedding based tf-idf similarity metric and a soft positional constraint which improve our system performance. We evaluate our techniques on test batch 4 from the fourth edition of the challenge. Our best system achieves a ROUGE-2 score of 0.6534 and ROUGE-SU4 score of 0.6536.