Autoregressive decoding with generative Large Language Models (LLMs) on accelerators (GPUs/TPUs) is often memory-bound where most of the time is spent on transferring model parameters from high bandwidth memory (HBM) to cache. On the other hand, recent works show that LLMs can maintain quality with significant sparsity/redundancy in the feedforward (FFN) layers by appropriately training the model to operate on a top-$k$ fraction of rows/columns (where $k \approx 0.05$), there by suggesting a way to reduce the transfer of model parameters, and hence latency. However, exploiting this sparsity for improving latency is hindered by the fact that identifying top rows/columns is data-dependent and is usually performed using full matrix operations, severely limiting potential gains. To address these issues, we introduce HiRE (High Recall Approximate Top-k Estimation). HiRE comprises of two novel components: (i) a compression scheme to cheaply predict top-$k$ rows/columns with high recall, followed by full computation restricted to the predicted subset, and (ii) DA-TOP-$k$: an efficient multi-device approximate top-$k$ operator. We demonstrate that on a one billion parameter model, HiRE applied to both the softmax as well as feedforward layers, achieves almost matching pretraining and downstream accuracy, and speeds up inference latency by $1.47\times$ on a single TPUv5e device.

Generating multiple translation candidates would enable users to choose the one that satisfies their needs. Although there has been work on diversified generation, there exists room for improving the diversity mainly because the previous methods do not address the overcorrection problem -- the model underestimates a prediction that is largely different from the training data, even if that prediction is likely. This paper proposes methods that generate more diverse translations by introducing perturbed k-nearest neighbor machine translation (kNN-MT). Our methods expand the search space of kNN-MT and help incorporate diverse words into candidates by addressing the overcorrection problem. Our experiments show that the proposed methods drastically improve candidate diversity and control the degree of diversity by tuning the perturbation's magnitude.

Projecting intermediate representations onto the vocabulary is an increasingly popular interpretation tool for transformer-based LLMs, also known as the logit lens. We propose a quantitative extension to this approach and define spectral filters on intermediate representations based on partitioning the singular vectors of the vocabulary embedding and unembedding matrices into bands. We find that the signals exchanged in the tail end of the spectrum are responsible for attention sinking (Xiao et al. 2023), of which we provide an explanation. We find that the loss of pretrained models can be kept low despite suppressing sizable parts of the embedding spectrum in a layer-dependent way, as long as attention sinking is preserved. Finally, we discover that the representation of tokens that draw attention from many tokens have large projections on the tail end of the spectrum.

Recent studies have shown that recommender systems (RSs) are highly vulnerable to data poisoning attacks. Understanding attack tactics helps improve the robustness of RSs. We intend to develop efficient attack methods that use limited resources to generate high-quality fake user profiles to achieve 1) transferability among black-box RSs 2) and imperceptibility among detectors. In order to achieve these goals, we introduce textual reviews of products to enhance the generation quality of the profiles. Specifically, we propose a novel attack framework named R-Trojan, which formulates the attack objectives as an optimization problem and adopts a tailored transformer-based generative adversarial network (GAN) to solve it so that high-quality attack profiles can be produced. Comprehensive experiments on real-world datasets demonstrate that R-Trojan greatly outperforms state-of-the-art attack methods on various victim RSs under black-box settings and show its good imperceptibility.

Standardising structure volume names in radiotherapy (RT) data is necessary to enable data mining and analyses, especially across multi-institutional centres. This process is time and resource intensive, which highlights the need for new automated and efficient approaches to handle the task. Several machine learning-based methods have been proposed and evaluated to standardise nomenclature. However, no studies have considered that RT patient records are distributed across multiple data centres. This paper introduces a method that emulates real-world environments to establish standardised nomenclature. This is achieved by integrating decentralised real-time data and federated learning (FL). A multimodal deep artificial neural network was proposed to standardise RT data in federated settings. Three types of possible attributes were extracted from the structures to train the deep learning models: tabular, visual, and volumetric. Simulated experiments were carried out to train the models across several scenarios including multiple data centres, input modalities, and aggregation strategies. The models were compared against models developed with single modalities in federated settings, in addition to models trained in centralised settings. Categorical classification accuracy was calculated on hold-out samples to inform the models performance. Our results highlight the need for fusing multiple modalities when training such models, with better performance reported with tabular-volumetric models. In addition, we report comparable accuracy compared to models built in centralised settings. This demonstrates the suitability of FL for handling the standardization task. Additional ablation analyses showed that the total number of samples in the data centres and the number of data centres highly affects the training process and should be carefully considered when building standardisation models.

Recently, Large Language Model based Autonomous system(LLMAS) has gained great popularity for its potential to simulate complicated behaviors of human societies. One of its main challenges is to present and analyze the dynamic events evolution of LLMAS. In this work, we present a visualization approach to explore detailed statuses and agents' behavior within LLMAS. We propose a general pipeline that establishes a behavior structure from raw LLMAS execution events, leverages a behavior summarization algorithm to construct a hierarchical summary of the entire structure in terms of time sequence, and a cause trace method to mine the causal relationship between agent behaviors. We then develop AgentLens, a visual analysis system that leverages a hierarchical temporal visualization for illustrating the evolution of LLMAS, and supports users to interactively investigate details and causes of agents' behaviors. Two usage scenarios and a user study demonstrate the effectiveness and usability of our AgentLens.

Transformer, as one of the most advanced neural network models in Natural Language Processing (NLP), exhibits diverse applications in the field of anomaly detection. To inspire research on Transformer-based anomaly detection, this review offers a fresh perspective on the concept of anomaly detection. We explore the current challenges of anomaly detection and provide detailed insights into the operating principles of Transformer and its variants in anomaly detection tasks. Additionally, we delineate various application scenarios for Transformer-based anomaly detection models and discuss the datasets and evaluation metrics employed. Furthermore, this review highlights the key challenges in Transformer-based anomaly detection research and conducts a comprehensive analysis of future research trends in this domain. The review includes an extensive compilation of over 100 core references related to Transformer-based anomaly detection. To the best of our knowledge, this is the first comprehensive review that focuses on the research related to Transformer in the context of anomaly detection. We hope that this paper can provide detailed technical information to researchers interested in Transformer-based anomaly detection tasks.

Existing knowledge hypergraph embedding methods mainly focused on improving model performance, but their model structures are becoming more complex and redundant. Furthermore, due to the inherent complex semantic knowledge, the computation of knowledge hypergraph embedding models is often very expensive, leading to low efficiency. In this paper, we propose a feature interaction and extraction-enhanced 3D circular convolutional embedding model, HyCubE, which designs a novel 3D circular convolutional neural network and introduces the alternate mask stack strategy to achieve efficient n-ary knowledge hypergraph embedding. By adaptively adjusting the 3D circular convolution kernel size and uniformly embedding the entity position information, HyCubE improves the model performance with fewer parameters and reaches a better trade-off between model performance and efficiency. In addition, we use 1-N multilinear scoring based on the entity mask mechanism to further accelerate the model training efficiency. Finally, extensive experimental results on all datasets demonstrate that HyCubE consistently outperforms state-of-the-art baselines, with an average improvement of 4.08%-10.77% and a maximum improvement of 21.16% across all metrics. Commendably, HyCubE speeds up by an average of 7.55x and reduces memory usage by an average of 77.02% compared to the latest state-of-the-art baselines.

Large language models (LLMs) have performed well on several reasoning benchmarks, including ones that test analogical reasoning abilities. However, it has been debated whether they are actually performing humanlike abstract reasoning or instead employing less general processes that rely on similarity to what has been seen in their training data. Here we investigate the generality of analogy-making abilities previously claimed for LLMs (Webb, Holyoak, & Lu, 2023). We take one set of analogy problems used to evaluate LLMs and create a set of "counterfactual" variants-versions that test the same abstract reasoning abilities but that are likely dissimilar from any pre-training data. We test humans and three GPT models on both the original and counterfactual problems, and show that, while the performance of humans remains high for all the problems, the GPT models' performance declines sharply on the counterfactual set. This work provides evidence that, despite previously reported successes of LLMs on analogical reasoning, these models lack the robustness and generality of human analogy-making.

Neural networks sometimes exhibit grokking, a phenomenon where perfect or near-perfect performance is achieved on a validation set well after the same performance has been obtained on the corresponding training set. In this workshop paper, we introduce a robust technique for measuring grokking, based on fitting an appropriate functional form. We then use this to investigate the sharpness of transitions in training and validation accuracy under two settings. The first setting is the theoretical framework developed by Levi et al. (2023) where closed form expressions are readily accessible. The second setting is a two-layer MLP trained to predict the parity of bits, with grokking induced by the concealment strategy of Miller et al. (2023). We find that trends between relative grokking gap and grokking sharpness are similar in both settings when using absolute and relative measures of sharpness. Reflecting on this, we make progress toward explaining some trends and identify the need for further study to untangle the various mechanisms which influence the sharpness of grokking.