Knowledge editing (KE) aims to efficiently and precisely modify the behavior of large language models (LLMs) to update specific knowledge without negatively influencing other knowledge. Current research primarily focuses on white-box LLMs editing, overlooking an important scenario: black-box LLMs editing, where LLMs are accessed through interfaces and only textual output is available. To address the limitations of existing evaluations that are not inapplicable to black-box LLM editing and lack comprehensiveness, we propose a multi-perspective evaluation framework, incorporating the assessment of style retention for the first time. To tackle privacy leaks of editing data and style over-editing in current methods, we introduce a novel postEdit framework, resolving privacy concerns through downstream post-processing and maintaining textual style consistency via fine-grained editing to original responses. Experiments and analysis on two benchmarks demonstrate that postEdit outperforms all baselines and achieves strong generalization, especially with huge improvements on style retention (average $+20.82\%\uparrow$).
In this paper, we propose a class of non-parametric classifiers, that learn arbitrary boundaries and generalize well. Our approach is based on a novel way to regularize 1NN classifiers using a greedy approach. We refer to this class of classifiers as Watershed Classifiers. 1NN classifiers are known to trivially over-fit but have very large VC dimension, hence do not generalize well. We show that watershed classifiers can find arbitrary boundaries on any dense enough dataset, and, at the same time, have very small VC dimension; hence a watershed classifier leads to good generalization. Traditional approaches to regularize 1NN classifiers are to consider $K$ nearest neighbours. Neighbourhood component analysis (NCA) proposes a way to learn representations consistent with ($n-1$) nearest neighbour classifier, where $n$ denotes the size of the dataset. In this article, we propose a loss function which can learn representations consistent with watershed classifiers, and show that it outperforms the NCA baseline.
Vision Transformers (ViTs) have emerged as a state-of-the-art solution for object classification tasks. However, their computational demands and high parameter count make them unsuitable for real-time inference, prompting the need for efficient hardware implementations. Existing hardware accelerators for ViTs suffer from frequent off-chip memory access, restricting the achievable throughput by memory bandwidth. In devices with a high compute-to-communication ratio (e.g., edge FPGAs with limited bandwidth), off-chip memory access imposes a severe bottleneck on overall throughput. This work proposes ME-ViT, a novel \underline{M}emory \underline{E}fficient FPGA accelerator for \underline{ViT} inference that minimizes memory traffic. We propose a \textit{single-load policy} in designing ME-ViT: model parameters are only loaded once, intermediate results are stored on-chip, and all operations are implemented in a single processing element. To achieve this goal, we design a memory-efficient processing element (ME-PE), which processes multiple key operations of ViT inference on the same architecture through the reuse of \textit{multi-purpose buffers}. We also integrate the Softmax and LayerNorm functions into the ME-PE, minimizing stalls between matrix multiplications. We evaluate ME-ViT on systolic array sizes of 32 and 16, achieving up to a 9.22$\times$ and 17.89$\times$ overall improvement in memory bandwidth, and a 2.16$\times$ improvement in throughput per DSP for both designs over state-of-the-art ViT accelerators on FPGA. ME-ViT achieves a power efficiency improvement of up to 4.00$\times$ (1.03$\times$) over a GPU (FPGA) baseline. ME-ViT enables up to 5 ME-PE instantiations on a Xilinx Alveo U200, achieving a 5.10$\times$ improvement in throughput over the state-of-the art FPGA baseline, and a 5.85$\times$ (1.51$\times$) improvement in power efficiency over the GPU (FPGA) baseline.
The generative paradigm has become increasingly important in machine learning and deep learning models. Among popular generative models are normalizing flows, which enable exact likelihood estimation by transforming a base distribution through diffeomorphic transformations. Extending the normalizing flow framework to handle time-indexed flows gave dynamic normalizing flows, a powerful tool to model time series, stochastic processes, and neural stochastic differential equations (SDEs). In this work, we propose a novel variant of dynamic normalizing flows, a Time Changed Normalizing Flow (TCNF), based on time deformation of a Brownian motion which constitutes a versatile and extensive family of Gaussian processes. This approach enables us to effectively model some SDEs, that cannot be modeled otherwise, including standard ones such as the well-known Ornstein-Uhlenbeck process, and generalizes prior methodologies, leading to improved results and better inference and prediction capability.
Modern computationally-intensive applications often operate under time constraints, necessitating acceleration methods and distribution of computational workloads across multiple entities. However, the outcome is either achieved within the desired timeline or not, and in the latter case, valuable resources are wasted. In this paper, we introduce solutions for layered-resolution computation. These solutions allow lower-resolution results to be obtained at an earlier stage than the final result. This innovation notably enhances the deadline-based systems, as if a computational job is terminated due to time constraints, an approximate version of the final result can still be generated. Moreover, in certain operational regimes, a high-resolution result might be unnecessary, because the low-resolution result may already deviate significantly from the decision threshold, for example in AI-based decision-making systems. Therefore, operators can decide whether higher resolution is needed or not based on intermediate results, enabling computations with adaptive resolution. We present our framework for two critical and computationally demanding jobs: distributed matrix multiplication (linear) and model inference in machine learning (nonlinear). Our theoretical and empirical results demonstrate that the execution delay for the first resolution is significantly shorter than that for the final resolution, while maintaining overall complexity comparable to the conventional one-shot approach. Our experiments further illustrate how the layering feature increases the likelihood of meeting deadlines and enables adaptability and transparency in massive, large-scale computations.
Denoising diffusion models have become ubiquitous for generative modeling. The core idea is to transport the data distribution to a Gaussian by using a diffusion. Approximate samples from the data distribution are then obtained by estimating the time-reversal of this diffusion using score matching ideas. We follow here a similar strategy to sample from unnormalized probability densities and compute their normalizing constants. However, the time-reversed diffusion is here simulated by using an original iterative particle scheme relying on a novel score matching loss. Contrary to standard denoising diffusion models, the resulting Particle Denoising Diffusion Sampler (PDDS) provides asymptotically consistent estimates under mild assumptions. We demonstrate PDDS on multimodal and high dimensional sampling tasks.
In this paper we build a case for providing job completion time predictions to cloud users, similar to the delivery date of a package or arrival time of a booked ride. Our analysis reveals that providing predictability can come at the expense of performance and fairness. Existing cloud scheduling systems optimize for extreme points in the trade-off space, making them either extremely unpredictable or impractical. To address this challenge, we present PCS, a new scheduling framework that aims to provide predictability while balancing other traditional objectives. The key idea behind PCS is to use Weighted-Fair-Queueing (WFQ) and find a suitable configuration of different WFQ parameters (e.g., class weights) that meets specific goals for predictability. It uses a simulation-aided search strategy, to efficiently discover WFQ configurations that lie on the Pareto front of the trade-off space between these objectives. We implement and evaluate PCS in the context of DNN job scheduling on GPUs. Our evaluation, on a small scale GPU testbed and larger-scale simulations, shows that PCS can provide accurate completion time estimates while marginally compromising on performance and fairness.
Zero-shot text-to-speech (TTS) has gained significant attention due to its powerful voice cloning capabilities, requiring only a few seconds of unseen speaker voice prompts. However, all previous work has been developed for cloud-based systems. Taking autoregressive models as an example, although these approaches achieve high-fidelity voice cloning, they fall short in terms of inference speed, model size, and robustness. Therefore, we propose MobileSpeech, which is a fast, lightweight, and robust zero-shot text-to-speech system based on mobile devices for the first time. Specifically: 1) leveraging discrete codec, we design a parallel speech mask decoder module called SMD, which incorporates hierarchical information from the speech codec and weight mechanisms across different codec layers during the generation process. Moreover, to bridge the gap between text and speech, we introduce a high-level probabilistic mask that simulates the progression of information flow from less to more during speech generation. 2) For speaker prompts, we extract fine-grained prompt duration from the prompt speech and incorporate text, prompt speech by cross attention in SMD. We demonstrate the effectiveness of MobileSpeech on multilingual datasets at different levels, achieving state-of-the-art results in terms of generating speed and speech quality. MobileSpeech achieves RTF of 0.09 on a single A100 GPU and we have successfully deployed MobileSpeech on mobile devices. Audio samples are available at \url{https://mobilespeech.github.io/} .
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.
We investigate decentralized online convex optimization (D-OCO), in which a set of local learners are required to minimize a sequence of global loss functions using only local computations and communications. Previous studies have established $O(n^{5/4}\rho^{-1/2}\sqrt{T})$ and ${O}(n^{3/2}\rho^{-1}\log T)$ regret bounds for convex and strongly convex functions respectively, where $n$ is the number of local learners, $\rho<1$ is the spectral gap of the communication matrix, and $T$ is the time horizon. However, there exist large gaps from the existing lower bounds, i.e., $\Omega(n\sqrt{T})$ for convex functions and $\Omega(n)$ for strongly convex functions. To fill these gaps, in this paper, we first develop novel D-OCO algorithms that can respectively reduce the regret bounds for convex and strongly convex functions to $\tilde{O}(n\rho^{-1/4}\sqrt{T})$ and $\tilde{O}(n\rho^{-1/2}\log T)$. The primary technique is to design an online accelerated gossip strategy that enjoys a faster average consensus among local learners. Furthermore, by carefully exploiting the spectral properties of a specific network topology, we enhance the lower bounds for convex and strongly convex functions to $\Omega(n\rho^{-1/4}\sqrt{T})$ and $\Omega(n\rho^{-1/2})$, respectively. These lower bounds suggest that our algorithms are nearly optimal in terms of $T$, $n$, and $\rho$.