We describe an approach for aligning an LLM-based dialogue agent based on global (i.e., dialogue-level) rewards, while also taking into account naturally-occurring multimodal signals. At a high level, our approach (dubbed GELI) learns a local, turn-level reward model by decomposing the human-provided Global Explicit (GE) session-level reward, using Local Implicit (LI) multimodal reward signals to crossmodally shape the reward decomposition step. This decomposed reward model is then used as part of the standard RHLF pipeline improve an LLM-based dialog agent. We run quantitative and qualitative human studies to evaluate the performance of our GELI approach, and find that it shows consistent improvements across various conversational metrics compared to baseline methods.
相關內容
Despite the notable success of language models (LMs) in various natural language processing (NLP) tasks, the reliability of LMs is susceptible to backdoor attacks. Prior research attempts to mitigate backdoor learning while training the LMs on the poisoned dataset, yet struggles against complex backdoor attacks in real-world scenarios. In this paper, we investigate the learning mechanisms of backdoor LMs in the frequency space by Fourier analysis. Our findings indicate that the backdoor mapping presented on the poisoned datasets exhibits a more discernible inclination towards lower frequency compared to clean mapping, resulting in the faster convergence of backdoor mapping. To alleviate this dilemma, we propose Multi-Scale Low-Rank Adaptation (MuScleLoRA), which deploys multiple radial scalings in the frequency space with low-rank adaptation to the target model and further aligns the gradients when updating parameters. Through downscaling in the frequency space, MuScleLoRA encourages the model to prioritize the learning of relatively high-frequency clean mapping, consequently mitigating backdoor learning. Experimental results demonstrate that MuScleLoRA outperforms baselines significantly. Notably, MuScleLoRA reduces the average success rate of diverse backdoor attacks to below 15\% across multiple datasets and generalizes to various backbone LMs, including BERT, RoBERTa, GPT2-XL, and Llama2. The codes are publicly available at //github.com/ZrW00/MuScleLoRA.
This paper studies a beam tracking problem in which an access point (AP), in collaboration with a reconfigurable intelligent surface (RIS), dynamically adjusts its downlink beamformers and the reflection pattern at the RIS in order to maintain reliable communications with multiple mobile user equipments (UEs). Specifically, the mobile UEs send uplink pilots to the AP periodically during the channel sensing intervals, the AP then adaptively configures the beamformers and the RIS reflection coefficients for subsequent data transmission based on the received pilots. This is an active sensing problem, because channel sensing involves configuring the RIS coefficients during the pilot stage and the optimal sensing strategy should exploit the trajectory of channel state information (CSI) from previously received pilots. Analytical solution to such an active sensing problem is very challenging. In this paper, we propose a deep learning framework utilizing a recurrent neural network (RNN) to automatically summarize the time-varying CSI obtained from the periodically received pilots into state vectors. These state vectors are then mapped to the AP beamformers and RIS reflection coefficients for subsequent downlink data transmissions, as well as the RIS reflection coefficients for the next round of uplink channel sensing. The mappings from the state vectors to the downlink beamformers and the RIS reflection coefficients for both channel sensing and downlink data transmission are performed using graph neural networks (GNNs) to account for the interference among the UEs. Simulations demonstrate significant and interpretable performance improvement of the proposed approach over the existing data-driven methods with nonadaptive channel sensing schemes.
Functional data analysis (FDA) finds widespread application across various fields, due to data being recorded continuously over a time interval or at several discrete points. Since the data is not observed at every point but rather across a dense grid, smoothing techniques are often employed to convert the observed data into functions. In this work, we propose a novel Bayesian approach for selecting basis functions for smoothing one or multiple curves simultaneously. Our method differentiates from other Bayesian approaches in two key ways: (i) by accounting for correlated errors and (ii) by developing a variational EM algorithm instead of a Gibbs sampler. Simulation studies demonstrate that our method effectively identifies the true underlying structure of the data across various scenarios and it is applicable to different types of functional data. Our variational EM algorithm not only recovers the basis coefficients and the correct set of basis functions but also estimates the existing within-curve correlation. When applied to the motorcycle dataset, our method demonstrates comparable, and in some cases superior, performance in terms of adjusted $R^2$ compared to other techniques such as regression splines, Bayesian LASSO and LASSO. Additionally, when assuming independence among observations within a curve, our method, utilizing only a variational Bayes algorithm, is in the order of thousands faster than a Gibbs sampler on average. Our proposed method is implemented in R and codes are available at //github.com/acarolcruz/VB-Bases-Selection.
Despite several works that succeed in generating synthetic data with differential privacy (DP) guarantees, they are inadequate for generating high-quality synthetic data when the input data has missing values. In this work, we formalize the problems of DP synthetic data with missing values and propose three effective adaptive strategies that significantly improve the utility of the synthetic data on four real-world datasets with different types and levels of missing data and privacy requirements. We also identify the relationship between privacy impact for the complete ground truth data and incomplete data for these DP synthetic data generation algorithms. We model the missing mechanisms as a sampling process to obtain tighter upper bounds for the privacy guarantees to the ground truth data. Overall, this study contributes to a better understanding of the challenges and opportunities for using private synthetic data generation algorithms in the presence of missing data.
Label noise, commonly found in real-world datasets, has a detrimental impact on a model's generalization. To effectively detect incorrectly labeled instances, previous works have mostly relied on distinguishable training signals, such as training loss, as indicators to differentiate between clean and noisy labels. However, they have limitations in that the training signals incompletely reveal the model's behavior and are not effectively generalized to various noise types, resulting in limited detection accuracy. In this paper, we propose DynaCor framework that distinguishes incorrectly labeled instances from correctly labeled ones based on the dynamics of the training signals. To cope with the absence of supervision for clean and noisy labels, DynaCor first introduces a label corruption strategy that augments the original dataset with intentionally corrupted labels, enabling indirect simulation of the model's behavior on noisy labels. Then, DynaCor learns to identify clean and noisy instances by inducing two clearly distinguishable clusters from the latent representations of training dynamics. Our comprehensive experiments show that DynaCor outperforms the state-of-the-art competitors and shows strong robustness to various noise types and noise rates.
Multivariate time series may be subject to partial structural changes over certain frequency band, for instance, in neuroscience. We study the change point detection problem with high dimensional time series, within the framework of frequency domain. The overarching goal is to locate all change points and delineate which series are activated by the change, over which frequencies. In practice, the number of activated series per change and frequency could span from a few to full participation. We solve the problem by first computing a CUSUM tensor based on spectra estimated from blocks of the time series. A frequency-specific projection approach is applied for dimension reduction. The projection direction is estimated by a proposed tensor decomposition algorithm that adjusts to the sparsity level of changes. Finally, the projected CUSUM vectors across frequencies are aggregated for change point detection. We provide theoretical guarantees on the number of estimated change points and the convergence rate of their locations. We derive error bounds for the estimated projection direction for identifying the frequency-specific series activated in a change. We provide data-driven rules for the choice of parameters. The efficacy of the proposed method is illustrated by simulation and a stock returns application.
This work investigates the performance of intelligent reflective surfaces (IRSs) assisted uplink non-orthogonal multiple access (NOMA) in energy-constrained networks. Specifically, we formulate and solve two optimization problems; the first aims at minimizing the sum of users' transmit power, while the second targets maximizing the system level energy efficiency (EE). The two problems are solved by jointly optimizing the users' transmit powers and the beamforming coefficients at IRS subject to the users' individual uplink rate and transmit power constraints. A novel and low complexity algorithm is developed to optimize the IRS beamforming coefficients by optimizing the objective function over a \textit{complex circle manifold} (CCM). To efficiently optimize the IRS phase shifts over the manifold, the optimization problem is reformulated into a feasibility expansion problem which is reduced to a max-min signal-plus-interference-ratio (SINR). Then, with the aid of a smoothing technique, the exact penalty method is applied to transform the problem from constrained to unconstrained. The proposed solution is compared against three semi-definite programming (SDP)-based benchmarks which are semi-definite relaxation (SDR), SDP-difference of convex (SDP-DC) and sequential rank-one constraint relaxation (SROCR). The results show that the manifold algorithm provides better performance than the SDP-based benchmarks, and at a much lower computational complexity for both the transmit power minimization and EE maximization problems. The results also reveal that IRS-NOMA is only superior to orthogonal multiple access (OMA) when the users' target achievable rate requirements are relatively high.
One principal approach for illuminating a black-box neural network is feature attribution, i.e. identifying the importance of input features for the network's prediction. The predictive information of features is recently proposed as a proxy for the measure of their importance. So far, the predictive information is only identified for latent features by placing an information bottleneck within the network. We propose a method to identify features with predictive information in the input domain. The method results in fine-grained identification of input features' information and is agnostic to network architecture. The core idea of our method is leveraging a bottleneck on the input that only lets input features associated with predictive latent features pass through. We compare our method with several feature attribution methods using mainstream feature attribution evaluation experiments. The code is publicly available.
Translational distance-based knowledge graph embedding has shown progressive improvements on the link prediction task, from TransE to the latest state-of-the-art RotatE. However, N-1, 1-N and N-N predictions still remain challenging. In this work, we propose a novel translational distance-based approach for knowledge graph link prediction. The proposed method includes two-folds, first we extend the RotatE from 2D complex domain to high dimension space with orthogonal transforms to model relations for better modeling capacity. Second, the graph context is explicitly modeled via two directed context representations. These context representations are used as part of the distance scoring function to measure the plausibility of the triples during training and inference. The proposed approach effectively improves prediction accuracy on the difficult N-1, 1-N and N-N cases for knowledge graph link prediction task. The experimental results show that it achieves better performance on two benchmark data sets compared to the baseline RotatE, especially on data set (FB15k-237) with many high in-degree connection nodes.
Detecting carried objects is one of the requirements for developing systems to reason about activities involving people and objects. We present an approach to detect carried objects from a single video frame with a novel method that incorporates features from multiple scales. Initially, a foreground mask in a video frame is segmented into multi-scale superpixels. Then the human-like regions in the segmented area are identified by matching a set of extracted features from superpixels against learned features in a codebook. A carried object probability map is generated using the complement of the matching probabilities of superpixels to human-like regions and background information. A group of superpixels with high carried object probability and strong edge support is then merged to obtain the shape of the carried object. We applied our method to two challenging datasets, and results show that our method is competitive with or better than the state-of-the-art.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191