Bingli Liao

Bingli Liao, Danilo Vasconcellos Vargas

The efficiency of large language models (LLMs) remains a critical challenge, particularly in contexts where computational resources are limited. Traditional attention mechanisms in these models, while powerful, require significant computational and memory resources due to the necessity of recalculating and storing attention weights across different layers. This paper introduces a novel Shared Attention (SA) mechanism, designed to enhance the efficiency of LLMs by directly sharing computed attention weights across multiple layers. Unlike previous methods that focus on sharing intermediate Key-Value (KV) caches, our approach utilizes the isotropic tendencies of attention distributions observed in advanced LLMs post-pretraining to reduce both the computational flops and the size of the KV cache required during inference. We empirically demonstrate that implementing SA across various LLMs results in minimal accuracy loss on standard benchmarks. Our findings suggest that SA not only conserves computational resources but also maintains robust model performance, thereby facilitating the deployment of more efficient LLMs in resource-constrained environments.

Bingli Liao, Danilo Vasconcellos Vargas

Large Language Models (LLMs) are pivotal in advancing natural language processing but often struggle with complex reasoning tasks due to inefficient attention distributions. In this paper, we explore the effect of increased computed tokens on LLM performance and introduce a novel method for extending computed tokens in the Chain-of-Thought (CoT) process, utilizing attention mechanism optimization. By fine-tuning an LLM on a domain-specific, highly structured dataset, we analyze attention patterns across layers, identifying inefficiencies caused by non-semantic tokens with outlier high attention scores. To address this, we propose an algorithm that emulates early layer attention patterns across downstream layers to re-balance skewed attention distributions and enhance knowledge abstraction. Our findings demonstrate that our approach not only facilitates a deeper understanding of the internal dynamics of LLMs but also significantly improves their reasoning capabilities, particularly in non-STEM domains. Our study lays the groundwork for further innovations in LLM design, aiming to create more powerful, versatile, and responsible models capable of tackling a broad range of real-world applications.

Danilo Vasconcellos Vargas, Bingli Liao, Takahiro Kanzaki

Adversarial examples have shown that albeit highly accurate, models learned by machines, differently from humans, have many weaknesses. However, humans' perception is also fundamentally different from machines, because we do not see the signals which arrive at the retina but a rather complex recreation of them. In this paper, we explore how machines could recreate the input as well as investigate the benefits of such an augmented perception. In this regard, we propose Perceptual Deep Neural Networks ($\varphi$DNN) which also recreate their own input before further processing. The concept is formalized mathematically and two variations of it are developed (one based on inpainting the whole image and the other based on a noisy resized super resolution recreation). Experiments reveal that $\varphi$DNNs and their adversarial training variations can increase the robustness substantially, surpassing both state-of-the-art defenses and pre-processing types of defenses in 100% of the tests. $\varphi$DNNs are shown to scale well to bigger image sizes, keeping a similar high accuracy throughout; while the state-of-the-art worsen up to 35%. Moreover, the recreation process intentionally corrupts the input image. Interestingly, we show by ablation tests that corrupting the input is, although counter-intuitive, beneficial. Thus, $\varphi$DNNs reveal that input recreation has strong benefits for artificial neural networks similar to biological ones, shedding light into the importance of purposely corrupting the input as well as pioneering an area of perception models based on GANs and autoencoders for robust recognition in artificial intelligence.