BAdam: A Memory Efficient Full Parameter Training Method for Large Language Models

This work presents BAdam, an optimizer that leverages the block coordinate optimization framework with Adam as the inner solver. BAdam offers a memory efficient approach to the full parameter finetuning of large language models and reduces running time of the backward process thanks to the chain rule property. Experimentally, we apply BAdam to instruction-tune the Llama 2-7B model on the Alpaca-GPT4 dataset using a single RTX3090-24GB GPU. The results indicate that BAdam exhibits superior convergence behavior in comparison to LoRA and LOMO. Furthermore, our downstream performance evaluation of the instruction-tuned models using the MT-bench shows that BAdam modestly surpasses LoRA and more substantially outperforms LOMO. Finally, we compare BAdam with Adam on a medium-sized task, i.e., finetuning RoBERTa-large on the SuperGLUE benchmark. The results demonstrate that BAdam is capable of narrowing the performance gap with Adam. Our code is available at https://github.com/Ledzy/BAdam.

Autonomous Driving With Perception Uncertainties: Deep-Ensemble Based Adaptive Cruise Control

Autonomous driving depends on perception systems to understand the environment and to inform downstream decision-making. While advanced perception systems utilizing black-box Deep Neural Networks (DNNs) demonstrate human-like comprehension, their unpredictable behavior and lack of interpretability may hinder their deployment in safety critical scenarios. In this paper, we develop an Ensemble of DNN regressors (Deep Ensemble) that generates predictions with quantification of prediction uncertainties. In the scenario of Adaptive Cruise Control (ACC), we employ the Deep Ensemble to estimate distance headway to the lead vehicle from RGB images and enable the downstream controller to account for the estimation uncertainty. We develop an adaptive cruise controller that utilizes Stochastic Model Predictive Control (MPC) with chance constraints to provide a probabilistic safety guarantee. We evaluate our ACC algorithm using a high-fidelity traffic simulator and a real-world traffic dataset and demonstrate the ability of the proposed approach to effect speed tracking and car following while maintaining a safe distance headway. The out-of-distribution scenarios are also examined.

RCoCo: Contrastive Collective Link Prediction across Multiplex Network in Riemannian Space

Link prediction typically studies the probability of future interconnection among nodes with the observation in a single social network. More often than not, real scenario is presented as a multiplex network with common (anchor) users active in multiple social networks. In the literature, most existing works study either the intra-link prediction in a single network or inter-link prediction among networks (a.k.a. network alignment), and consider two learning tasks are independent from each other, which is still away from the fact. On the representation space, the vast majority of existing methods are built upon the traditional Euclidean space, unaware of the inherent geometry of social networks. The third issue is on the scarce anchor users. Annotating anchor users is laborious and expensive, and thus it is impractical to work with quantities of anchor users. Herein, in light of the issues above, we propose to study a challenging yet practical problem of Geometry-aware Collective Link Prediction across Multiplex Network. To address this problem, we present a novel contrastive model, RCoCo, which collaborates intra- and inter-network behaviors in Riemannian spaces. In RCoCo, we design a curvature-aware graph attention network ($\kappa-$GAT), conducting attention mechanism in Riemannian manifold whose curvature is estimated by the Ricci curvatures over the network. Thereafter, we formulate intra- and inter-contrastive loss in the manifolds, in which we augment graphs by exploring the high-order structure of community and information transfer on anchor users. Finally, we conduct extensive experiments with 14 strong baselines on 8 real-world datasets, and show the effectiveness of RCoCo.

FormulaQA: A Question Answering Dataset for Formula-Based Numerical Reasoning

The application of formulas is a fundamental ability of humans when addressing numerical reasoning problems. However, existing numerical reasoning datasets seldom explicitly indicate the formulas employed during the reasoning steps. To bridge this gap, we propose a question answering dataset for formula-based numerical reasoning called FormulaQA, from junior high school physics examinations. We further conduct evaluations on LLMs with size ranging from 7B to over 100B parameters utilizing zero-shot and few-shot chain-of-thoughts methods and we explored the approach of using retrieval-augmented LLMs when providing an external formula database. We also fine-tune on smaller models with size not exceeding 2B. Our empirical findings underscore the significant potential for improvement in existing models when applied to our complex, formula-driven FormulaQA.

System-level Safety Guard: Safe Tracking Control through Uncertain Neural Network Dynamics Models

The Neural Network (NN), as a black-box function approximator, has been considered in many control and robotics applications. However, difficulties in verifying the overall system safety in the presence of uncertainties hinder the modular deployment of NN in safety-critical systems. In this paper, we leverage the NNs as predictive models for trajectory tracking of unknown dynamical systems. We consider controller design in the presence of both intrinsic uncertainty and uncertainties from other system modules. In this setting, we formulate the constrained trajectory tracking problem and show that it can be solved using Mixed-integer Linear Programming (MILP). The proposed MILP-based solution enjoys a provable safety guarantee for the overall system, and the approach is empirically demonstrated in robot navigation and obstacle avoidance through simulations. The demonstration videos are available at https://xiaolisean.github.io/publication/2023-11-01-L4DC2024.

High Probability Guarantees for Random Reshuffling

We consider the stochastic gradient method with random reshuffling ($\mathsf{RR}$) for tackling smooth nonconvex optimization problems. $\mathsf{RR}$ finds broad applications in practice, notably in training neural networks. In this work, we first investigate the concentration property of $\mathsf{RR}$'s sampling procedure and establish a new high probability sample complexity guarantee for driving the gradient (without expectation) below $\varepsilon$, which effectively characterizes the efficiency of a single $\mathsf{RR}$ execution. Our derived complexity matches the best existing in-expectation one up to a logarithmic term while imposing no additional assumptions nor changing $\mathsf{RR}$'s updating rule. Furthermore, by leveraging our derived high probability descent property and bound on the stochastic error, we propose a simple and computable stopping criterion for $\mathsf{RR}$ (denoted as $\mathsf{RR}$-$\mathsf{sc}$). This criterion is guaranteed to be triggered after a finite number of iterations, and then $\mathsf{RR}$-$\mathsf{sc}$ returns an iterate with its gradient below $\varepsilon$ with high probability. Moreover, building on the proposed stopping criterion, we design a perturbed random reshuffling method ($\mathsf{p}$-$\mathsf{RR}$) that involves an additional randomized perturbation procedure near stationary points. We derive that $\mathsf{p}$-$\mathsf{RR}$ provably escapes strict saddle points and efficiently returns a second-order stationary point with high probability, without making any sub-Gaussian tail-type assumptions on the stochastic gradient errors. Finally, we conduct numerical experiments on neural network training to support our theoretical findings.

A New Random Reshuffling Method for Nonsmooth Nonconvex Finite-sum Optimization

In this work, we propose and study a novel stochastic optimization algorithm, termed the normal map-based proximal random reshuffling (norm-PRR) method, for nonsmooth nonconvex finite-sum problems. Random reshuffling techniques are prevalent and widely utilized in large-scale applications, e.g., in the training of neural networks. While the convergence behavior and advantageous acceleration effects of random reshuffling methods are fairly well understood in the smooth setting, much less seems to be known in the nonsmooth case and only few proximal-type random reshuffling approaches with provable guarantees exist. We establish the iteration complexity ${\cal O}(n^{-1/3}T^{-2/3})$ for norm-PRR, where $n$ is the number of component functions and $T$ counts the total number of iteration. We also provide novel asymptotic convergence results for norm-PRR. Specifically, under the Kurdyka-{\L}ojasiewicz (KL) inequality, we establish strong limit-point convergence, i.e., the iterates generated by norm-PRR converge to a single stationary point. Moreover, we derive last iterate convergence rates of the form ${\cal O}(k^{-p})$; here, $p \in [0, 1]$ depends on the KL exponent $\theta \in [0,1)$ and step size dynamics. Finally, we present preliminary numerical results on machine learning problems that demonstrate the efficiency of the proposed method.

Low-Complexity Joint Beamforming for RIS-Assisted MU-MISO Systems Based on Model-Driven Deep Learning

Reconfigurable intelligent surfaces (RIS) can improve signal propagation environments by adjusting the phase of the incident signal. However, optimizing the phase shifts jointly with the beamforming vector at the access point is challenging due to the non-convex objective function and constraints. In this study, we propose an algorithm based on weighted minimum mean square error optimization and power iteration to maximize the weighted sum rate (WSR) of a RIS-assisted downlink multi-user multiple-input single-output system. To further improve performance, a model-driven deep learning (DL) approach is designed, where trainable variables and graph neural networks are introduced to accelerate the convergence of the proposed algorithm. We also extend the proposed method to include beamforming with imperfect channel state information and derive a two-timescale stochastic optimization algorithm. Simulation results show that the proposed algorithm outperforms state-of-the-art algorithms in terms of complexity and WSR. Specifically, the model-driven DL approach has a runtime that is approximately 3% of the state-of-the-art algorithm to achieve the same performance. Additionally, the proposed algorithm with 2-bit phase shifters outperforms the compared algorithm with continuous phase shift.

Missing Value Imputation for Multi-attribute Sensor Data Streams via Message Propagation (Extended Version)

Sensor data streams occur widely in various real-time applications in the context of the Internet of Things (IoT). However, sensor data streams feature missing values due to factors such as sensor failures, communication errors, or depleted batteries. Missing values can compromise the quality of real-time analytics tasks and downstream applications. Existing imputation methods either make strong assumptions about streams or have low efficiency. In this study, we aim to accurately and efficiently impute missing values in data streams that satisfy only general characteristics in order to benefit real-time applications more widely. First, we propose a message propagation imputation network (MPIN) that is able to recover the missing values of data instances in a time window. We give a theoretical analysis of why MPIN is effective. Second, we present a continuous imputation framework that consists of data update and model update mechanisms to enable MPIN to perform continuous imputation both effectively and efficiently. Extensive experiments on multiple real datasets show that MPIN can outperform the existing data imputers by wide margins and that the continuous imputation framework is efficient and accurate.

Understanding Deep Representation Learning via Layerwise Feature Compression and Discrimination

Over the past decade, deep learning has proven to be a highly effective tool for learning meaningful features from raw data. However, it remains an open question how deep networks perform hierarchical feature learning across layers. In this work, we attempt to unveil this mystery by investigating the structures of intermediate features. Motivated by our empirical findings that linear layers mimic the roles of deep layers in nonlinear networks for feature learning, we explore how deep linear networks transform input data into output by investigating the output (i.e., features) of each layer after training in the context of multi-class classification problems. Toward this goal, we first define metrics to measure within-class compression and between-class discrimination of intermediate features, respectively. Through theoretical analysis of these two metrics, we show that the evolution of features follows a simple and quantitative pattern from shallow to deep layers when the input data is nearly orthogonal and the network weights are minimum-norm, balanced, and approximate low-rank: Each layer of the linear network progressively compresses within-class features at a geometric rate and discriminates between-class features at a linear rate with respect to the number of layers that data have passed through. To the best of our knowledge, this is the first quantitative characterization of feature evolution in hierarchical representations of deep linear networks. Empirically, our extensive experiments not only validate our theoretical results numerically but also reveal a similar pattern in deep nonlinear networks which aligns well with recent empirical studies. Moreover, we demonstrate the practical implications of our results in transfer learning. Our code is available at \url{https://github.com/Heimine/PNC_DLN}.