Figure 1: A summary of several outlier statistics recorded from ImageNet validation set on ViT. We use zero-based indexing for dimensions. BERTRecall from Figure 1 that all the outliers are only present in hidden dimensions #123, #180,4 #225, #308, #381, #526, #720 (with the majority of them in #180, #720). In Figures 9 and 10 we show more6 examples of the discovered self-attention patterns for attention heads #3 and #12 ( hidden dim #1807 and #720, respectively). We also show self-attention patterns in attention heads and layers which are8 not associated with the outliers in Figures 11 and 12, respectively.9

Transformer models have been widely adopted in various domains over the last years, and especially large language models have advanced the field of AI significantly. Due to their size, the capability of these networks has increased tremendously, but this has come at the cost of a significant increase in necessary compute. Quantization is one of the most effective ways to reduce the computational time and memory consumption of neural networks. Many studies have shown, however, that modern transformer models tend to learn strong outliers in their activations, making them difficult to quantize. To retain acceptable performance, the existence of these outliers requires activations to be in higher bitwidth or the use of different numeric formats, extra fine-tuning, or other workarounds.

We also present some additional ablation studies.

Due to their size, the capability of these networks has increased tremendously, but this has come at the cost of a significant increase in necessary compute.

With the rise of Transformer models in NLP and CV domain, Multi-Head Attention has been proven to be a game-changer. However, its expensive computation poses challenges to the model throughput and efficiency, especially for the long sequence tasks. Exploiting the sparsity in attention has been proven to be an effective way to reduce computation. Nevertheless, prior works do not consider the various distributions among different heads and lack a systematic method to determine the threshold. To address these challenges, we propose Low-Precision Approximate Attention with Head-wise Trainable Threshold for Efficient Transformer (LATTE). LATTE employs a headwise threshold-based filter with the low-precision dot product and computation reuse mechanism to reduce the computation of MHA. Moreover, the trainable threshold is introduced to provide a systematic method for adjusting the thresholds and enable end-to-end optimization. Experimental results indicate LATTE can smoothly adapt to both NLP and CV tasks, offering significant computation savings with only a minor compromise in performance. Also, the trainable threshold is shown to be essential for the leverage between the performance and the computation. As a result, LATTE filters up to 85.16% keys with only a 0.87% accuracy drop in the CV task and 89.91% keys with a 0.86 perplexity increase in the NLP task.

Because protein-protein interactions (PPIs) are crucial to understand living systems, harvesting these data is essential to probe disease development and discern gene/protein functions and biological processes. Some curated datasets contain PPI data derived from the literature and other sources (e.g., IntAct, BioGrid, DIP, and HPRD). However, they are far from exhaustive, and their maintenance is a labor-intensive process. On the other hand, machine learning methods to automate PPI knowledge extraction from the scientific literature have been limited by a shortage of appropriate annotated data. This work presents a unified, multi-source PPI corpora with vetted interaction definitions augmented by binary interaction type labels and a Transformer-based deep learning method that exploits entities' relational context information for relation representation to improve relation classification performance. The model's performance is evaluated on four widely studied biomedical relation extraction datasets, as well as this work's target PPI datasets, to observe the effectiveness of the representation to relation extraction tasks in various data. Results show the model outperforms prior state-of-the-art models. The code and data are available at: https://github.com/BNLNLP/PPI-Relation-Extraction

Many computational factors limit broader deployment of large language models. In this paper, we focus on a memory bottleneck imposed by the key-value (KV) cache, a computational shortcut that requires storing previous KV pairs during decoding. While existing KV cache methods approach this problem by pruning or evicting large swaths of relatively less important KV pairs to dramatically reduce the memory footprint of the cache, they can have limited success in tasks that require recollecting a majority of previous tokens. To alleviate this issue, we propose LESS, a simple integration of a (nearly free) constant sized cache with eviction-based cache methods, such that all tokens can be queried at later decoding steps. Its ability to retain information throughout time shows merit on a variety of tasks where we demonstrate LESS can help reduce the performance gap from caching everything, sometimes even matching it, all while being efficient.

Transformer models have been widely adopted in various domains over the last years, and especially large language models have advanced the field of AI significantly. Due to their size, the capability of these networks has increased tremendously, but this has come at the cost of a significant increase in necessary compute. Quantization is one of the most effective ways to reduce the computational time and memory consumption of neural networks. Many studies have shown, however, that modern transformer models tend to learn strong outliers in their activations, making them difficult to quantize. To retain acceptable performance, the existence of these outliers requires activations to be in higher bitwidth or the use of different numeric formats, extra fine-tuning, or other workarounds. We show that strong outliers are related to very specific behavior of attention heads that try to learn a "no-op" or just a partial update of the residual. To achieve the exact zeros needed in the attention matrix for a no-update, the input to the softmax is pushed to be larger and larger during training, causing outliers in other parts of the network. Based on these observations, we propose two simple (independent) modifications to the attention mechanism - clipped softmax and gated attention. We empirically show that models pre-trained using our methods learn significantly smaller outliers while maintaining and sometimes even improving the floating-point task performance. This enables us to quantize transformers to full INT8 quantization of the activations without any additional effort. We demonstrate the effectiveness of our methods on both language models (BERT, OPT) and vision transformers.

Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior "in the wild" in a language model. We evaluate the reliability of our explanation using three quantitative criteria--faithfulness, completeness and minimality. Though these criteria support our explanation, they also point to remaining gaps in our understanding. Our work provides evidence that a mechanistic understanding of large ML models is feasible, opening opportunities to scale our understanding to both larger models and more complex tasks.

This paper describes a novel study on using `Attention Mask' input in transformers and using this approach for detecting offensive content in both English and Persian languages. The paper's principal focus is to suggest a methodology to enhance the performance of the BERT-based models on the `Offensive Language Detection' task. Therefore, we customize attention probabilities by changing the `Attention Mask' input to create more efficacious word embeddings. To do this, we firstly tokenize the training set of the exploited datasets (by BERT tokenizer). Then, we apply Multinomial Naive Bayes to map these tokens to two probabilities. These probabilities indicate the likelihood of making a text non-offensive or offensive, provided that it contains that token. Afterwards, we use these probabilities to define a new term, namely Offensive Score. Next, we create two separate (because of the differences in the types of the employed datasets) equations based on Offensive Scores for each language to re-distribute the `Attention Mask' input for paying more attention to more offensive phrases. Eventually, we put the F1-macro score as our evaluation metric and fine-tune several combinations of BERT with ANNs, CNNs and RNNs to examine the effect of using this methodology on various combinations. The results indicate that all models will enhance with this methodology. The most improvement was 2% and 10% for English and Persian languages, respectively.