Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey

Large models represent a groundbreaking advancement in multiple application fields, enabling remarkable achievements across various tasks. However, their unprecedented scale comes with significant computational costs. These models, often consisting of billions of parameters, require vast amounts of computational resources for execution. Especially, the expansive scale and computational demands pose considerable challenges when customizing them for particular downstream tasks, particularly over the hardware platforms constrained by computational capabilities. Parameter Efficient Fine-Tuning (PEFT) provides a practical solution by efficiently adapt the large models over the various downstream tasks. In particular, PEFT refers to the process of adjusting the parameters of a pre-trained large models to adapt it to a specific task while minimizing the number of additional parameters introduced or computational resources required. This approach is particularly important when dealing with large language models with high parameter counts, as fine-tuning these models from scratch can be computationally expensive and resource-intensive, posing considerable challenges in the supporting system platform design. In this survey, we present comprehensive studies of various PEFT algorithms, examining their performance and computational overhead. Moreover, we provide an overview of applications developed using different PEFT algorithms and discuss common techniques employed to mitigate computation costs for PEFT. In addition to the algorithmic perspective, we overview various real-world system designs to investigate the implementation costs associated with different PEFT algorithms. This survey serves as an indispensable resource for researchers aiming to understand both the PEFT algorithm and its system implementation, offering detailed insights into recent advancements and practical applications.

Physics-Inspired Deep Learning Anti-Aliasing Framework in Efficient Channel State Feedback

Acquiring downlink channel state information (CSI) at the base station is vital for optimizing performance in massive Multiple input multiple output (MIMO) Frequency-Division Duplexing (FDD) systems. While deep learning architectures have been successful in facilitating UE-side CSI feedback and gNB-side recovery, the undersampling issue prior to CSI feedback is often overlooked. This issue, which arises from low density pilot placement in current standards, results in significant aliasing effects in outdoor channels and consequently limits CSI recovery performance. To this end, this work introduces a new CSI upsampling framework at the gNB as a post-processing solution to address the gaps caused by undersampling. Leveraging the physical principles of discrete Fourier transform shifting theorem and multipath reciprocity, our framework effectively uses uplink CSI to mitigate aliasing effects. We further develop a learning-based method that integrates the proposed algorithm with the Iterative Shrinkage-Thresholding Algorithm Net (ISTA-Net) architecture, enhancing our approach for non-uniform sampling recovery. Our numerical results show that both our rule-based and deep learning methods significantly outperform traditional interpolation techniques and current state-of-the-art approaches in terms of performance.

3D Beamforming Through Joint Phase-Time Arrays

High-frequency wide-bandwidth cellular communications over mmW and sub-THz offer the opportunity for high data rates, however, it also presents high pathloss, resulting in limited coverage. To mitigate the coverage limitations, high-gain beamforming is essential. Implementation of beamforming involves a large number of antennas, which introduces analog beam constraint, i.e., only one frequency-flat beam is generated per transceiver chain (TRx). Recently introduced joint phase-time array (JPTA) architecture, which utilizes both true time delay (TTD) units and phase shifters (PSs), alleviates analog beam constraint by creating multiple frequency-dependent beams per TRx, for scheduling multiple users at different directions in a frequency-division manner. One class of previous studies offered solutions with "rainbow" beams, which tend to allocate a small bandwidth per beam direction. Another class focused on uniform linear array (ULA) antenna architecture, whose frequency-dependent beams were designed along a single axis of either azimuth or elevation direction. In this paper, we present a novel 3D beamforming codebook design aimed at maximizing beamforming gain to steer radiation toward desired azimuth and elevation directions, as well as across sub-bands partitioned according to scheduled users' bandwidth requirements. We provide both analytical solutions and iterative algorithms to design the PSs and TTD units for a desired subband beam pattern. Through simulations of the beamforming gain, we observe that our proposed solutions outperform the state-of-the-art solutions reported elsewhere.

Joint Phase-Time Arrays: A Paradigm for Frequency-Dependent Analog Beamforming in 6G

Hybrid beamforming is an attractive solution to build cost-effective and energy-efficient transceivers for millimeter-wave and terahertz systems. However, conventional hybrid beamforming techniques rely on analog components that generate a frequency flat response such as phase-shifters and switches, which limits the flexibility of the achievable beam patterns. As a novel alternative, this paper proposes a new class of hybrid beamforming called Joint phase-time arrays (JPTA), that additionally use true-time delay elements in the analog beamforming to create frequency-dependent analog beams. Using as an example two important frequency-dependent beam behaviors, the numerous benefits of such flexibility are exemplified. Subsequently, the JPTA beamformer design problem to generate any desired beam behavior is formulated and near-optimal algorithms to the problem are proposed. Simulations show that the proposed algorithms can outperform heuristics solutions for JPTA beamformer update. Furthermore, it is shown that JPTA can achieve the two exemplified beam behaviors with one radio-frequency chain, while conventional hybrid beamforming requires the radio-frequency chains to scale with the number of antennas to achieve similar performance. Finally, a wide range of problems to further tap into the potential of JPTA are also listed as future directions.

Computing Approximate $\ell_p$ Sensitivities

Recent works in dimensionality reduction for regression tasks have introduced the notion of sensitivity, an estimate of the importance of a specific datapoint in a dataset, offering provable guarantees on the quality of the approximation after removing low-sensitivity datapoints via subsampling. However, fast algorithms for approximating $\ell_p$ sensitivities, which we show is equivalent to approximate $\ell_p$ regression, are known for only the $\ell_2$ setting, in which they are termed leverage scores. In this work, we provide efficient algorithms for approximating $\ell_p$ sensitivities and related summary statistics of a given matrix. In particular, for a given $n \times d$ matrix, we compute $\alpha$-approximation to its $\ell_1$ sensitivities at the cost of $O(n/\alpha)$ sensitivity computations. For estimating the total $\ell_p$ sensitivity (i.e. the sum of $\ell_p$ sensitivities), we provide an algorithm based on importance sampling of $\ell_p$ Lewis weights, which computes a constant factor approximation to the total sensitivity at the cost of roughly $O(\sqrt{d})$ sensitivity computations. Furthermore, we estimate the maximum $\ell_1$ sensitivity, up to a $\sqrt{d}$ factor, using $O(d)$ sensitivity computations. We generalize all these results to $\ell_p$ norms for $p > 1$. Lastly, we experimentally show that for a wide class of matrices in real-world datasets, the total sensitivity can be quickly approximated and is significantly smaller than the theoretical prediction, demonstrating that real-world datasets have low intrinsic effective dimensionality.

Hardness of Low Rank Approximation of Entrywise Transformed Matrix Products

Inspired by fast algorithms in natural language processing, we study low rank approximation in the entrywise transformed setting where we want to find a good rank $k$ approximation to $f(U \cdot V)$, where $U, V^\top \in \mathbb{R}^{n \times r}$ are given, $r = O(\log(n))$, and $f(x)$ is a general scalar function. Previous work in sublinear low rank approximation has shown that if both (1) $U = V^\top$ and (2) $f(x)$ is a PSD kernel function, then there is an $O(nk^{\omega-1})$ time constant relative error approximation algorithm, where $\omega \approx 2.376$ is the exponent of matrix multiplication. We give the first conditional time hardness results for this problem, demonstrating that both conditions (1) and (2) are in fact necessary for getting better than $n^{2-o(1)}$ time for a relative error low rank approximation for a wide class of functions. We give novel reductions from the Strong Exponential Time Hypothesis (SETH) that rely on lower bounding the leverage scores of flat sparse vectors and hold even when the rank of the transformed matrix $f(UV)$ and the target rank are $n^{o(1)}$, and when $U = V^\top$. Furthermore, even when $f(x) = x^p$ is a simple polynomial, we give runtime lower bounds in the case when $U \neq V^\top$ of the form $\Omega(\min(n^{2-o(1)}, \Omega(2^p)))$. Lastly, we demonstrate that our lower bounds are tight by giving an $O(n \cdot \text{poly}(k, 2^p, 1/\epsilon))$ time relative error approximation algorithm and a fast $O(n \cdot \text{poly}(k, p, 1/\epsilon))$ additive error approximation using fast tensor-based sketching. Additionally, since our low rank algorithms rely on matrix-vector product subroutines, our lower bounds extend to show that computing $f(UV)W$, for even a small matrix $W$, requires $\Omega(n^{2-o(1)})$ time.

Getting aligned on representational alignment

Biological and artificial information processing systems form representations of the world that they can use to categorize, reason, plan, navigate, and make decisions. To what extent do the representations formed by these diverse systems agree? Can diverging representations still lead to the same behaviors? And how can systems modify their representations to better match those of another system? These questions pertaining to the study of \textbf{\emph{representational alignment}} are at the heart of some of the most active research areas in contemporary cognitive science, neuroscience, and machine learning. Unfortunately, there is limited knowledge-transfer between research communities interested in representational alignment, and much of the progress in one field ends up being rediscovered independently in another, when greater cross-field communication would be advantageous. To improve communication between fields, we propose a unifying framework that can serve as a common language between researchers studying representational alignment. We survey the literature from the fields of cognitive science, neuroscience, and machine learning, and demonstrate how prior work fits into this framework. Finally, we lay out open problems in representational alignment where progress can benefit all three fields. We hope that our work can catalyze cross-disciplinary collaboration and accelerate progress for all communities studying and developing information processing systems. We note that this is a working paper and encourage readers to reach out with their suggestions for future revisions.

An Active Noise Control System Based on Soundfield Interpolation Using a Physics-informed Neural Network

Conventional multiple-point active noise control (ANC) systems require placing error microphones within the region of interest (ROI), inconveniencing users. This paper designs a feasible monitoring microphone arrangement placed outside the ROI, providing a user with more freedom of movement. The soundfield within the ROI is interpolated from the microphone signals using a physics-informed neural network (PINN). PINN exploits the acoustic wave equation to assist soundfield interpolation under a limited number of monitoring microphones, and demonstrates better interpolation performance than the spherical harmonic method in simulations. An ANC system is designed to take advantage of the interpolated signal to reduce noise signal within the ROI. The PINN-assisted ANC system reduces noise more than that of the multiple-point ANC system in simulations.

Optimal preprocessing of WiFi CSI for sensing applications

Due to its ubiquitous and contact-free nature, the use of WiFi infrastructure for performing sensing tasks has tremendous potential. However, the channel state information (CSI) measured by a WiFi receiver suffers from errors in both its gain and phase, which can significantly hinder sensing tasks. By analyzing these errors from different WiFi receivers, a mathematical model for these gain and phase errors is developed in this work. Based on these models, several theoretically justified preprocessing algorithms for correcting such errors at a receiver and, thus, obtaining clean CSI are presented. Simulation results show that at typical system parameters, the developed algorithms for cleaning CSI can reduce noise by $40$% and $200$%, respectively, compared to baseline methods for gain correction and phase correction, without significantly impacting computational cost. The superiority of the proposed methods is also validated in a real-world test bed for respiration rate monitoring (an exemplary sensing task), where they improve the estimation signal-to-noise ratio by $20$% compared to baseline methods.

Leveraging Contextual Counterfactuals Toward Belief Calibration

Beliefs and values are increasingly being incorporated into our AI systems through alignment processes, such as carefully curating data collection principles or regularizing the loss function used for training. However, the meta-alignment problem is that these human beliefs are diverse and not aligned across populations; furthermore, the implicit strength of each belief may not be well calibrated even among humans, especially when trying to generalize across contexts. Specifically, in high regret situations, we observe that contextual counterfactuals and recourse costs are particularly important in updating a decision maker's beliefs and the strengths to which such beliefs are held. Therefore, we argue that including counterfactuals is key to an accurate calibration of beliefs during alignment. To do this, we first segment belief diversity into two categories: subjectivity (across individuals within a population) and epistemic uncertainty (within an individual across different contexts). By leveraging our notion of epistemic uncertainty, we introduce `the belief calibration cycle' framework to more holistically calibrate this diversity of beliefs with context-driven counterfactual reasoning by using a multi-objective optimization. We empirically apply our framework for finding a Pareto frontier of clustered optimal belief strengths that generalize across different contexts, demonstrating its efficacy on a toy dataset for credit decisions.