Large language models (LLMs) have achieved remarkable success in the field of natural language processing, enabling better human-computer interaction using natural language. However, the seamless integration of speech signals into LLMs has not been explored well. The "decoder-only" architecture has also not been well studied for speech processing tasks. In this research, we introduce Speech-LLaMA, a novel approach that effectively incorporates acoustic information into text-based large language models. Our method leverages Connectionist Temporal Classification and a simple audio encoder to map the compressed acoustic features to the continuous semantic space of the LLM. In addition, we further probe the decoder-only architecture for speech-to-text tasks by training a smaller scale randomly initialized speech-LLaMA model from speech-text paired data alone. We conduct experiments on multilingual speech-to-text translation tasks and demonstrate a significant improvement over strong baselines, highlighting the potential advantages of decoder-only models for speech-to-text conversion.
相關內容
Large language models (LLMs) have demonstrated remarkable prediction performance for a growing array of tasks. However, their rapid proliferation and increasing opaqueness have created a growing need for interpretability. Here, we ask whether we can automatically obtain natural language explanations for black box text modules. A "text module" is any function that maps text to a scalar continuous value, such as a submodule within an LLM or a fitted model of a brain region. "Black box" indicates that we only have access to the module's inputs/outputs. We introduce Summarize and Score (SASC), a method that takes in a text module and returns a natural language explanation of the module's selectivity along with a score for how reliable the explanation is. We study SASC in 3 contexts. First, we evaluate SASC on synthetic modules and find that it often recovers ground truth explanations. Second, we use SASC to explain modules found within a pre-trained BERT model, enabling inspection of the model's internals. Finally, we show that SASC can generate explanations for the response of individual fMRI voxels to language stimuli, with potential applications to fine-grained brain mapping. All code for using SASC and reproducing results is made available on Github.
Generalized linear models (GLMs) are popular for data-analysis in almost all quantitative sciences, but the choice of likelihood family and link function is often difficult. This motivates the search for likelihoods and links that minimize the impact of potential misspecification. We perform a large-scale simulation study on double-bounded and lower-bounded response data where we systematically vary both true and assumed likelihoods and links. In contrast to previous studies, we also study posterior calibration and uncertainty metrics in addition to point-estimate accuracy. Our results indicate that certain likelihoods and links can be remarkably robust to misspecification, performing almost on par with their respective true counterparts. Additionally, normal likelihood models with identity link (i.e., linear regression) often achieve calibration comparable to the more structurally faithful alternatives, at least in the studied scenarios. On the basis of our findings, we provide practical suggestions for robust likelihood and link choices in GLMs.
It has been widely observed that exact or approximate MAP (mode-seeking) decoding from natural language generation (NLG) models consistently leads to degenerate outputs (Stahlberg and Byrne, 2019, Holtzman et al., 2019). This has generally been attributed to either a fundamental inadequacy of modes in models or weaknesses in language modeling. Contrastingly in this work, we emphasize that degenerate modes can even occur in the absence of any model error, due to contamination of the training data. Specifically, we show that mixing even a tiny amount of low-entropy noise with a population text distribution can cause the data distribution's mode to become degenerate, implying that any models trained on it will be as well. As the unconditional mode of NLG models will often be degenerate, we therefore propose to apply MAP decoding to the model's distribution conditional on avoiding specific degeneracies. Using exact-search, we empirically verify that the length-conditional modes of machine translation models and language models are indeed more fluent and topical than their unconditional modes. For the first time, we also share many examples of exact modal sequences from these models, and from several variants of the LLaMA-7B model. Notably, the modes of the LLaMA models are still degenerate, showing that improvements in modeling have not fixed this issue. Because of the cost of exact mode finding algorithms, we develop an approximate mode finding approach, ACBS, which finds sequences that are both high-likelihood and high-quality. We apply this approach to LLaMA-7B, a model which was not trained for instruction following, and find that we are able to elicit reasonable outputs without any finetuning.
Gaussian processes (GPs) are a popular class of Bayesian nonparametric models, but its training can be computationally burdensome for massive training datasets. While there has been notable work on scaling up these models for big data, existing methods typically rely on a stationary GP assumption for approximation, and can thus perform poorly when the underlying response surface is non-stationary, i.e., it has some regions of rapid change and other regions with little change. Such non-stationarity is, however, ubiquitous in real-world problems, including our motivating application for surrogate modeling of computer experiments. We thus propose a new Product of Sparse GP (ProSpar-GP) method for scalable GP modeling with massive non-stationary data. The ProSpar-GP makes use of a carefully-constructed product-of-experts formulation of sparse GP experts, where different experts are placed within local regions of non-stationarity. These GP experts are fit via a novel variational inference approach, which capitalizes on mini-batching and GPU acceleration for efficient optimization of inducing points and length-scale parameters for each expert. We further show that the ProSpar-GP is Kolmogorov-consistent, in that its generative distribution defines a valid stochastic process over the prediction space; such a property provides essential stability for variational inference, particularly in the presence of non-stationarity. We then demonstrate the improved performance of the ProSpar-GP over the state-of-the-art, in a suite of numerical experiments and an application for surrogate modeling of a satellite drag simulator.
A significant limitation of the LTE-V2X and NR-V2X sidelink scheduling mechanisms is their difficulty coping with variations in inter packet arrival times, also known as aperiodic packets. This conflicts with the fundamental characteristics of most V2X services which are triggered based on an event. e.g. ETSI Cooperative Awareness Messages (CAMs) - vehicle kinematics, Cooperative Perception Messages (CPMs) - object sensing and Decentralised Event Notification Messages (DENMs) - event occurrences. Furthermore, network management techniques such as congestion control mechanisms can result in varied inter packet arrival times. To combat this, NR-V2X introduced a dynamic grant mechanism, which we show is ineffective unless there is background periodic traffic to stabilise the sensing history upon which the scheduler makes it decisions. The characteristics of V2X services make it implausible that such periodic application traffic will exist. To overcome this significant drawback, we demonstrate that the standardised scheduling algorithms can be made effective if the event triggered arrival rate of packets can be accurately predicted. These predictions can be used to tune the Resource Reservation Interval (RRI) parameter of the MAC scheduler to negate the negative impact of aperiodicity. Such an approach allows the scheduler to achieve comparable performance to a scenario where packets arrive periodically. To demonstrate the effectiveness of our approach, an ML model has been devised for the prediction of cooperative awareness messages, but the same principle can be abstracted to other V2X service types.
Feedforward neural networks (FNNs) are typically viewed as pure prediction algorithms, and their strong predictive performance has led to their use in many machine-learning applications. However, their flexibility comes with an interpretability trade-off; thus, FNNs have been historically less popular among statisticians. Nevertheless, classical statistical theory, such as significance testing and uncertainty quantification, is still relevant. Supplementing FNNs with methods of statistical inference, and covariate-effect visualisations, can shift the focus away from black-box prediction and make FNNs more akin to traditional statistical models. This can allow for more inferential analysis, and, hence, make FNNs more accessible within the statistical-modelling context.
Pre-trained language models have demonstrated powerful capabilities in the field of natural language processing (NLP). Recently, code pre-trained model (PTM), which draw from the experiences of the NLP field, have also achieved state-of-the-art results in many software engineering (SE) downstream tasks. These code PTMs take into account the differences between programming languages and natural languages during pre-training and make adjustments to pre-training tasks and input data. However, researchers in the SE community still inherit habits from the NLP field when using these code PTMs to generate embeddings for SE downstream classification tasks, such as generating semantic embeddings for code snippets through special tokens and inputting code and text information in the same way as pre-training the PTMs. In this paper, we empirically study five different PTMs (i.e. CodeBERT, CodeT5, PLBART, CodeGPT and CodeGen) with three different architectures (i.e. encoder-only, decoder-only and encoder-decoder) on four SE downstream classification tasks (i.e. code vulnerability detection, code clone detection, just-in-time defect prediction and function docstring mismatch detection) with respect to the two aforementioned aspects. Our experimental results indicate that (1) regardless of the architecture of the code PTMs used, embeddings obtained through special tokens do not sufficiently aggregate the semantic information of the entire code snippet; (2) the quality of code embeddings obtained by combing code data and text data in the same way as pre-training the PTMs is poor and cannot guarantee richer semantic information; (3) using the method that aggregates the vector representations of all code tokens, the decoder-only PTMs can obtain code embeddings with semantics as rich as or even better quality than those obtained from the encoder-only and encoder-decoder PTMs.
Traditional perturbative statistical disclosure control (SDC) approaches such as microaggregation, noise addition, rank swapping, etc, perturb the data in an ``ad-hoc" way in the sense that while they manage to preserve some particular aspects of the data, they end up modifying others. Synthetic data approaches based on the fully conditional specification data synthesis paradigm, on the other hand, aim to generate new datasets that follow the same joint probability distribution as the original data. These synthetic data approaches, however, rely either on parametric statistical models, or non-parametric machine learning models, which need to fit well the original data in order to generate credible and useful synthetic data. Another important drawback is that they tend to perform better when the variables are synthesized in the correct causal order (i.e., in the same order as the true data generating process), which is often unknown in practice. To circumvent these issues, we propose a fully non-parametric and model free perturbative SDC approach that approximates the joint distribution of the original data via sequential applications of restricted permutations to the numerical microdata (where the restricted permutations are guided by the joint distribution of a discretized version of the data). Empirical comparisons against popular SDC approaches, using both real and simulated datasets, suggest that the proposed approach is competitive in terms of the trade-off between confidentiality and data utility.
The joint modeling of multiple longitudinal biomarkers together with a time-to-event outcome is a challenging modeling task of continued scientific interest. In particular, the computational complexity of high dimensional (generalized) mixed effects models often restricts the flexibility of shared parameter joint models, even when the subject-specific marker trajectories follow highly nonlinear courses. We propose a parsimonious multivariate functional principal components representation of the shared random effects. This allows better scalability, as the dimension of the random effects does not directly increase with the number of markers, only with the chosen number of principal component basis functions used in the approximation of the random effects. The functional principal component representation additionally allows to estimate highly flexible subject-specific random trajectories without parametric assumptions. The modeled trajectories can thus be distinctly different for each biomarker. We build on the framework of flexible Bayesian additive joint models implemented in the R-package 'bamlss', which also supports estimation of nonlinear covariate effects via Bayesian P-splines. The flexible yet parsimonious functional principal components basis used in the estimation of the joint model is first estimated in a preliminary step. We validate our approach in a simulation study and illustrate its advantages by analyzing a study on primary biliary cholangitis.
This study compares the performance of (1) fine-tuned models and (2) extremely large language models on the task of check-worthy claim detection. For the purpose of the comparison we composed a multilingual and multi-topical dataset comprising texts of various sources and styles. Building on this, we performed a benchmark analysis to determine the most general multilingual and multi-topical claim detector. We chose three state-of-the-art models in the check-worthy claim detection task and fine-tuned them. Furthermore, we selected three state-of-the-art extremely large language models without any fine-tuning. We made modifications to the models to adapt them for multilingual settings and through extensive experimentation and evaluation. We assessed the performance of all the models in terms of accuracy, recall, and F1-score in in-domain and cross-domain scenarios. Our results demonstrate that despite the technological progress in the area of natural language processing, the models fine-tuned for the task of check-worthy claim detection still outperform the zero-shot approaches in a cross-domain settings.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191