The inference process for large language models is slow and memory-intensive, with one of the most critical bottlenecks being excessive Key-Value (KV) cache accesses. This paper introduces "Double Sparsity," a novel post-training sparse attention technique designed to alleviate this bottleneck by reducing KV cache access. Double Sparsity combines token sparsity, which focuses on utilizing only the important tokens for computing self-attention, with channel sparsity, an approach that uses important feature channels for identifying important tokens. Our key insight is that the pattern of channel sparsity is relatively static, allowing us to use offline calibration to make it efficient at runtime, thereby enabling accurate and efficient identification of important tokens. Moreover, this method can be combined with offloading to achieve significant memory usage reduction. Experimental results demonstrate that Double Sparsity can achieve $\frac{1}{16}$ token and channel sparsity with minimal impact on accuracy across various tasks, including wiki-2 perplexity, key-value retrieval, and long context benchmarks with models including Llama-2-7B, Llama-2-70B, and Mixtral-8x7B. It brings up to a 14.1$\times$ acceleration in attention operations and a 1.9$\times$ improvement in end-to-end inference on GPUs. With offloading, it achieves a decoding speed acceleration of 16.3$\times$ compared to state-of-the-art solutions at a sequence length of 256K. Our code is publicly available at https://github.com/andy-yang-1/DoubleSparse.

This study aimed to determine if ChatGPT's large language models could match the scoring accuracy of human and machine scores from the ASAP competition. The investigation focused on various prediction models, including linear regression, random forest, gradient boost, and boost. ChatGPT's performance was evaluated against human raters using quadratic weighted kappa (QWK) metrics. Results indicated that while ChatGPT's gradient boost model achieved QWKs close to human raters for some data sets, its overall performance was inconsistent and often lower than human scores. The study highlighted the need for further refinement, particularly in handling biases and ensuring scoring fairness. Despite these challenges, ChatGPT demonstrated potential for scoring efficiency, especially with domain-specific fine-tuning. The study concludes that ChatGPT can complement human scoring but requires additional development to be reliable for high-stakes assessments. Future research should improve model accuracy, address ethical considerations, and explore hybrid models combining ChatGPT with empirical methods.

We highlight a perhaps important but hitherto unobserved insight: The attention module in a transformer is a smoothed cubic spline. Viewed in this manner, this mysterious but critical component of a transformer becomes a natural development of an old notion deeply entrenched in classical approximation theory. More precisely, we show that with ReLU-activation, attention, masked attention, encoder-decoder attention are all cubic splines. As every component in a transformer is constructed out of compositions of various attention modules (= cubic splines) and feed forward neural networks (= linear splines), all its components -- encoder, decoder, and encoder-decoder blocks; multilayered encoders and decoders; the transformer itself -- are cubic or higher-order splines. If we assume the Pierce-Birkhoff conjecture, then the converse also holds, i.e., every spline is a ReLU-activated encoder. Since a spline is generally just $C^2$, one way to obtain a smoothed $C^\infty$-version is by replacing ReLU with a smooth activation; and if this activation is chosen to be SoftMax, we recover the original transformer as proposed by Vaswani et al. This insight sheds light on the nature of the transformer by casting it entirely in terms of splines, one of the best known and thoroughly understood objects in applied mathematics.

Large language models (LLMs) such as GPT-4 sometimes appear to be creative, solving novel tasks often with a few demonstrations in the prompt. These tasks require the models to generalize on distributions different from those from training data -- which is known as out-of-distribution (OOD) generalization. Despite the tremendous success of LLMs, how they approach OOD generalization remains an open and underexplored question. We examine OOD generalization in settings where instances are generated according to hidden rules, including in-context learning with symbolic reasoning. Models are required to infer the hidden rules behind input prompts without any fine-tuning. We empirically examined the training dynamics of Transformers on a synthetic example and conducted extensive experiments on a variety of pretrained LLMs, focusing on a type of components known as induction heads. We found that OOD generalization and composition are tied together -- models can learn rules by composing two self-attention layers, thereby achieving OOD generalization. Furthermore, a shared latent subspace in the embedding (or feature) space acts as a bridge for composition by aligning early layers and later layers, which we refer to as the common bridge representation hypothesis.

Empirical risk minimization, a cornerstone in machine learning, is often hindered by the Optimizer's Curse stemming from discrepancies between the empirical and true data-generating distributions.To address this challenge, the robust satisficing framework has emerged recently to mitigate ambiguity in the true distribution. Distinguished by its interpretable hyperparameter and enhanced performance guarantees, this approach has attracted increasing attention from academia. However, its applicability in tackling general machine learning problems, notably deep neural networks, remains largely unexplored due to the computational challenges in solving this model efficiently across general loss functions. In this study, we delve into the Kullback Leibler divergence based robust satisficing model under a general loss function, presenting analytical interpretations, diverse performance guarantees, efficient and stable numerical methods, convergence analysis, and an extension tailored for hierarchical data structures. Through extensive numerical experiments across three distinct machine learning tasks, we demonstrate the superior performance of our model compared to state-of-the-art benchmarks.

This paper proposes design techniques for partially-calibrated sparse linear subarrays and algorithms to perform direction-of-arrival (DOA) estimation. First, we introduce array architectures that incorporate two distinct array categories, namely type-I and type-II arrays. The former breaks down a known sparse linear geometry into as many pieces as we need, and the latter employs each subarray such as it fits a preplanned sparse linear geometry. Moreover, we devise two Direction of Arrival (DOA) estimation algorithms that are suitable for partially-calibrated array scenarios within the coarray domain. The algorithms are capable of estimating a greater number of sources than the number of available physical sensors, while maintaining the hardware and computational complexity within practical limits for real-time implementation. To this end, we exploit the intersection of projections onto affine spaces by devising the Generalized Coarray Multiple Signal Classification (GCA-MUSIC) in conjunction with the estimation of a refined projection matrix related to the noise subspace, as proposed in the GCA root-MUSIC algorithm. An analysis is performed for the devised subarray configurations in terms of degrees of freedom, as well as the computation of the Cram\`er-Rao Lower Bound for the utilized data model, in order to demonstrate the good performance of the proposed methods. Simulations assess the performance of the proposed design methods and algorithms against existing approaches.

Large Language Models (LLMs) are highly sensitive to prompts, including additional context provided therein. As LLMs grow in capability, understanding their prompt-sensitivity becomes increasingly crucial for ensuring reliable and robust performance, particularly since evaluating these models becomes more challenging. In this work, we investigate how current models (Llama, Mixtral, Falcon) respond when presented with additional input from another model, mimicking a scenario where a more capable model -- or a system with access to more external information -- provides supplementary information to the target model. Across a diverse spectrum of question-answering tasks, we study how an LLM's response to multiple-choice questions changes when the prompt includes a prediction and explanation from another model. Specifically, we explore the influence of the presence of an explanation, the stated authoritativeness of the source, and the stated confidence of the supplementary input. Our findings reveal that models are strongly influenced, and when explanations are provided they are swayed irrespective of the quality of the explanation. The models are more likely to be swayed if the input is presented as being authoritative or confident, but the effect is small in size. This study underscores the significant prompt-sensitivity of LLMs and highlights the potential risks of incorporating outputs from external sources without thorough scrutiny and further validation. As LLMs continue to advance, understanding and mitigating such sensitivities will be crucial for their reliable and trustworthy deployment.

This study introduces the Federated Medical Knowledge Injection (FEDMEKI) platform, a new benchmark designed to address the unique challenges of integrating medical knowledge into foundation models under privacy constraints. By leveraging a cross-silo federated learning approach, FEDMEKI circumvents the issues associated with centralized data collection, which is often prohibited under health regulations like the Health Insurance Portability and Accountability Act (HIPAA) in the USA. The platform is meticulously designed to handle multi-site, multi-modal, and multi-task medical data, which includes 7 medical modalities, including images, signals, texts, laboratory test results, vital signs, input variables, and output variables. The curated dataset to validate FEDMEKI covers 8 medical tasks, including 6 classification tasks (lung opacity detection, COVID-19 detection, electrocardiogram (ECG) abnormal detection, mortality prediction, sepsis prediction, and enlarged cardiomediastinum detection) and 2 generation tasks (medical visual question answering (MedVQA) and ECG noise clarification). This comprehensive dataset is partitioned across several clients to facilitate the decentralized training process under 16 benchmark approaches. FEDMEKI not only preserves data privacy but also enhances the capability of medical foundation models by allowing them to learn from a broader spectrum of medical knowledge without direct data exposure, thereby setting a new benchmark in the application of foundation models within the healthcare sector.

The versatility of self-attention mechanism earned transformers great success in almost all data modalities, with limitations on the quadratic complexity and difficulty of training. To apply transformers across different data modalities, practitioners have to make specific clever data-modality-dependent constructions. In this paper, we propose Sampling Foundational Transformer (SFT) that can work on multiple data modalities (e.g., point cloud, graph, and sequence) and constraints (e.g., rotational-invariant). The existence of such model is important as contemporary foundational modeling requires operability on multiple data sources. For efficiency on large number of tokens, our model relies on our context aware sampling-without-replacement mechanism for both linear asymptotic computational complexity and real inference time gain. For efficiency, we rely on our newly discovered pseudoconvex formulation of transformer layer to increase model's convergence rate. As a model working on multiple data modalities, SFT has achieved competitive results on many benchmarks, while being faster in inference, compared to other very specialized models.

In this paper, we introduce a probabilistic extension to Kolmogorov Arnold Networks (KANs) by incorporating Gaussian Process (GP) as non-linear neurons, which we refer to as GP-KAN. A fully analytical approach to handling the output distribution of one GP as an input to another GP is achieved by considering the function inner product of a GP function sample with the input distribution. These GP neurons exhibit robust non-linear modelling capabilities while using few parameters and can be easily and fully integrated in a feed-forward network structure. They provide inherent uncertainty estimates to the model prediction and can be trained directly on the log-likelihood objective function, without needing variational lower bounds or approximations. In the context of MNIST classification, a model based on GP-KAN of 80 thousand parameters achieved 98.5% prediction accuracy, compared to current state-of-the-art models with 1.5 million parameters.