NeurIPS 2024 Papers

Despite advances using low-rank adapters and quantization, pretraining of large models on consumer hardware has not been possible without model sharding, offloading during training, or per-layer gradient updates. To address these limitations, we propose Low-Rank Adapters for Quantized Training (LoQT), a method for efficiently training quantized models. LoQT uses gradient-based tensor factorization to initialize low-rank trainable weight matrices that are periodically merged into quantized full-rank weight matrices. Our approach is suitable for both pretraining and fine-tuning models. We demonstrate this for language modeling and downstream task adaptation, finding that LoQT enables efficient training of models up to 7B parameters on a 24GB GPU. We also demonstrate the feasibility of training a 13B model using per-layer gradient updates on the same hardware.

Recent advancements in text-to-image diffusion models have enabled the personalization of these models to generate custom images from textual prompts. This paper presents an efficient LoRA-based personalization approach for on-device subject-driven generation, where pre-trained diffusion models are fine-tuned with user-specific data on resource-constrained devices. Our method, termed Hollowed Net, enhances memory efficiency during fine-tuning by modifying the architecture of a diffusion U-Net to temporarily remove a fraction of its deep layers, creating a hollowed structure. This approach directly addresses on-device memory constraints and substantially reduces GPU memory requirements for training, in contrast to previous methods that primarily focus on minimizing training steps and reducing the number of parameters to update. Additionally, the personalized Hollowed Net can be transferred back into the original U-Net, enabling inference without additional memory overhead. Quantitative and qualitative analyses demonstrate that our approach not only reduces training memory to levels as low as those required for inference but also maintains or improves personalization performance compared to existing methods.

In many settings, machine learning models may be used to inform decisions that impact individuals or entities who interact with the model. Such entities, or *agents,* may *game* model decisions by manipulating their inputs to the model to obtain better outcomes and maximize some utility. We consider a multi-agent setting where the goal is to identify the “worst offenders:” agents that are gaming most aggressively. However, identifying such agents is difficult without knowledge of their utility function. Thus, we introduce a framework in which each agent’s tendency to game is parameterized via a scalar. We show that this gaming parameter is only partially identifiable. By recasting the problem as a causal effect estimation problem where different agents represent different “treatments,” we prove that a ranking of all agents by their gaming parameters is identifiable. We present empirical results in a synthetic data study validating the usage of causal effect estimation for gaming detection and show in a case study of diagnosis coding behavior in the U.S. that our approach highlights features associated with gaming.

Recently, a wide range of memory-efficient LLM training algorithms have gained substantial popularity. These methods leverage the low-rank structure of gradients to project optimizer states into a subspace using projection matrix found by singular value decomposition (SVD). However, convergence of these algorithms is highly dependent on the update rules of their projection matrix. In this work, we provide the \emph{first} convergence guarantee for arbitrary update rules of projection matrix. This guarantee is generally applicable to optimizers that can be analyzed with Hamiltonian Descent, including most common ones, such as LION, Adam. Inspired by our theoretical understanding, we propose Online Subspace Descent, a new family of subspace descent optimizer without SVD. Instead of updating projection matrix with eigenvectors, Online Subspace Descent updates projection matrix wtih online PCA. Online Subspace Descent is flexible and introduces only minimum overhead to training. We demonstrate that, for the task of pretraining LLaMA models ranging from 60M to 1B parameters on the C4 dataset, Online Subspace Descent achieves lower perplexity than state-of-the-art low-rank training methods across different settings and narrows the gap with full-rank baselines.

Recent analysis on the training dynamics of Transformers has unveiled an interesting characteristic: the training loss plateaus for a significant number of training steps, and then suddenly (and sharply) drops to near--optimal values. To understand this phenomenon in depth, we formulate the low-rank matrix completion problem as a masked language modeling (MLM) task, and show that it is possible to train a BERT model to solve this task to low error. Furthermore, the loss curve shows a plateau early in training followed by a sudden drop to near-optimal values, despite no changes in the training procedure or hyper-parameters. To gain interpretability insights into this sudden drop, we examine the model's predictions, attention heads, and hidden states before and after this transition. Concretely, we observe that (a) the model transitions from simply copying the masked input to accurately predicting the masked entries; (b) the attention heads transition to interpretable patterns relevant to the task; and (c) the embeddings and hidden states encode information relevant to the problem. We also analyze the training dynamics of individual model components to understand the sudden drop in loss.

Recently, a series of diffusion-aware distillation algorithms have emerged to alleviate the computational overhead associated with the multi-step inference process of Diffusion Models (DMs). Current distillation techniques often dichotomize into two distinct aspects: i) ODE Trajectory Preservation; and ii) ODE Trajectory Reformulation. However, these approaches suffer from severe performance degradation or domain shifts. To address these limitations, we propose Hyper-SD, a novel framework that synergistically amalgamates the advantages of ODE Trajectory Preservation and Reformulation, while maintaining near-lossless performance during step compression. Firstly, we introduce Trajectory Segmented Consistency Distillation to progressively perform consistent distillation within pre-defined time-step segments, which facilitates the preservation of the original ODE trajectory from a higher-order perspective. Secondly, we incorporate human feedback learning to boost the performance of the model in a low-step regime and mitigate the performance loss incurred by the distillation process. Thirdly, we integrate score distillation to further improve the low-step generation capability of the model and offer the first attempt to leverage a unified LoRA to support the inference process at all steps. Extensive experiments and user studies demonstrate that Hyper-SD achieves SOTA performance from 1 to 8 inference steps for both SDXL and SD1.5. For example, Hyper-SDXL surpasses SDXL-Lightning by +0.68 in CLIP Score and +0.51 in Aes Score in the 1-step inference.

In this paper, we introduce YOLA, a novel framework for object detection in low-light scenarios. Unlike previous works, we propose to tackle this challenging problem from the perspective of feature learning. Specifically, we propose to learn illumination-invariant features through the Lambertian image formation model. We observe that, under the Lambertian assumption, it is feasible to approximate illumination-invariant feature maps by exploiting the interrelationships between neighboring color channels and spatially adjacent pixels. By incorporating additional constraints, these relationships can be characterized in the form of convolutional kernels, which can be trained in a detection-driven manner within a network. Towards this end, we introduce a novel module dedicated to the extraction of illumination-invariant features from low-light images, which can be easily integrated into existing object detection frameworks. Our empirical findings reveal significant improvements in low-light object detection tasks, as well as promising results in both well-lit and over-lit scenarios.

Meta learning is a promising paradigm in the era of large models and task distributional robustness has become an indispensable consideration in real-world scenarios. Recent advances have examined the effectiveness of tail task risk minimization in fast adaptation robustness improvement \citep{wang2023simple}. This work contributes to more theoretical investigations and practical enhancements in the field. Specifically, we reduce the distributionally robust strategy to a max-min optimization problem, constitute the Stackelberg equilibrium as the solution concept, and estimate the convergence rate. In the presence of tail risk, we further derive the generalization bound, establish connections with estimated quantiles, and practically improve the studied strategy. Accordingly, extensive evaluations demonstrate the significance of our proposal in boosting robustness.

Large Language Models (LLMs) often suffer from overconfidence during inference, particularly when adapted to downstream domain-specific tasks with limited data. Previous work addresses this issue by employing approximate Bayesian estimation after the LLMs are trained, enabling them to quantify uncertainty. However, such post-training approaches' performance is severely limited by the parameters learned during training. In this paper, we go beyond post-training Bayesianization and propose Bayesian Low-Rank Adaptation by Backpropagation (BLoB), an algorithm that continuously and jointly adjusts both the mean and covariance of LLM parameters throughout the whole fine-tuning process. Our empirical results verify the effectiveness of BLoB in terms of generalization and uncertainty estimation, when evaluated on both in-distribution and out-of-distribution data.

The classical Canonical Correlation Analysis (CCA) identifies the correlations between two sets of multivariate variables based on their covariance, which has been widely applied in diverse fields such as computer vision, natural language processing, and speech analysis. Despite its popularity, CCA can encounter challenges in explaining correlations between two variable sets within high-dimensional data contexts. Thus, this paper studies Sparse Canonical Correlation Analysis (SCCA) that enhances the interpretability of CCA. We first show that SCCA generalizes three well-known sparse optimization problems, sparse PCA, sparse SVD, and sparse regression, which are all classified as NP-hard problems. This result motivates us to develop strong formulations and efficient algorithms. Our main contributions include (i) the introduction of a combinatorial formulation that captures the essence of SCCA and allows the development of exact and approximation algorithms; (ii) the establishment of the complexity results for two low-rank special cases of SCCA; and (iii) the derivation of an equivalent mixed-integer semidefinite programming model that facilitates a specialized branch-and-cut algorithm with analytical cuts. The effectiveness of our proposed formulations and algorithms is validated through numerical experiments.

Large language models (LLMs) have shown impressive capabilities across various tasks. However, training LLMs from scratch requires significant computational power and extensive memory capacity. Recent studies have explored low-rank structures on weights for efficient fine-tuning in terms of parameters and memory, either through low-rank adaptation or factorization. While effective for fine-tuning, low-rank structures are generally less suitable for pretraining because they restrict parameters to a low-dimensional subspace. In this work, we propose to parameterize the weights as a sum of low-rank and sparse matrices for pretraining, which we call SLTrain. The low-rank component is learned via matrix factorization, while for the sparse component, we employ a simple strategy of uniformly selecting the sparsity support at random and learning only the non-zero entries with the fixed support. While being simple, the random fixed-support sparse learning strategy significantly enhances pretraining when combined with low-rank learning. Our results show that SLTrain adds minimal extra parameters and memory costs compared to pretraining with low-rank parameterization, yet achieves substantially better performance, which is comparable to full-rank training. Remarkably, when combined with quantization and per-layer updates, SLTrain can reduce memory requirements by up to 73% when pretraining the LLaMA 7B model.

Recent efforts to scale Transformer models have demonstrated rapid progress across a wide range of tasks (Wei at. al 2022). However, fine-tuning these models for downstream tasks is quite expensive due to their large parameter counts. Parameter-efficient fine-tuning (PEFT) approaches have emerged as a viable alternative, allowing us to fine-tune models by updating only a small number of parameters. In this work, we propose a general framework for parameter efficient fine-tuning (PEFT), based on *structured unrestricted-rank matrices* (SURM) which can serve as a drop-in replacement for popular approaches such as Adapters and LoRA. Unlike other methods like LoRA, SURMs give us more flexibility in finding the right balance between compactness and expressiveness. This is achieved by using *low displacement rank matrices* (LDRMs), which hasn't been used in this context before. SURMs remain competitive with baselines, often providing significant quality improvements while using a smaller parameter budget. SURMs achieve: **5**-**7**% accuracy gains on various image classification tasks while replacing low-rank matrices in LoRA and: up to **12x** reduction of the number of parameters in adapters (with virtually no loss in quality) on the GLUE benchmark.

Data selection is essential for training deep learning models. An effective data sampler assigns proper sampling probability for training data and helps the model converge to a good local minimum with high performance. Previous studies in data sampling are mainly based on heuristic rules or learning through a huge amount of time-consuming trials. In this paper, we propose an automatic swift sampler search algorithm, SS, to explore automatically learning effective samplers efficiently. In particular, SS utilizes a novel formulation to map a sampler to a low dimension of hyper-parameters and uses an approximated local minimum to quickly examine the quality of a sampler. Benefiting from its low computational expense, SS can be applied on large-scale data sets with high efficiency. Comprehensive experiments on various tasks demonstrate that SS powered sampling can achieve obvious improvements (e.g., 1.5% on ImageNet) and transfer among different neural networks. Project page: https://github.com/Alexander-Yao/Swift-Sampler.

While following different technical routes, both low-rank and orthogonal adaptation techniques can efficiently adapt large-scale pre-training models in specific tasks or domains based on a small piece of trainable parameters. In this study, we bridge the gap between these two techniques, proposing a simple but effective adaptation method based on Householder reflections. Given a pre-trained model, our method fine-tunes its layers by multiplying each frozen weight matrix with an orthogonal matrix constructed by a chain of learnable Householder reflections (HRs). This HR-based orthogonal fine-tuning is equivalent to an adaptive low-rank adaptation. Moreover, we show that the orthogonality of the reflection planes corresponding to the HRs impacts the model capacity and regularity. The analysis motivates us to regularize the orthogonality of the HRs, leading to different implementations of the proposed Householder reflection adaptation (HRA) method. Compared with state-of-the-art methods, HRA achieves superior performance with fewer learnable parameters when adapting large language models and conditional image generators. The code of the experiments is available at https://github.com/DaShenZi721/HRA, and the method has been merged into the [PEFT](https://github.com/huggingface/peft) package.

The machine learning community has witnessed impressive advancements since large language models (LLMs) first appeared. Yet, their massive memory consumption has become a significant roadblock to large-scale training. For instance, a 7B model typically requires at least 60 GB of GPU memory with full parameter training, which presents challenges for researchers without access to high-resource environments. Parameter Efficient Fine-Tuning techniques such as Low-Rank Adaptation (LoRA) have been proposed to alleviate this problem. However, in most large-scale fine-tuning settings, their performance does not reach the level of full parameter training because they confine the parameter search to a low-rank subspace. Attempting to complement this deficiency, we investigate the layerwise properties of LoRA on fine-tuning tasks and observe an unexpected but consistent skewness of weight norms across different layers. Utilizing this key observation, a surprisingly simple training strategy is discovered, which outperforms both LoRA and full parameter training in a wide range of settings with memory costs as low as LoRA. We name it Layerwise Importance Sampled AdamW (LISA), a promising alternative for LoRA, which applies the idea of importance sampling to different layers in LLMs and randomly freeze most middle layers during optimization. Experimental results show that with similar or less GPU memory consumption, LISA surpasses LoRA or even full parameter tuning in downstream fine-tuning tasks, where LISA consistently outperforms LoRA by over 10%-35% in terms of MT-Bench score while achieving on-par or better performance in MMLU, AGIEval and WinoGrande. On large models, specifically LLaMA-2-70B, LISA surpasses LoRA on MT-Bench, GSM8K, and PubMedQA, demonstrating its effectiveness across different domains.

While open Large Language Models (LLMs) have made significant progress, they still fall short of matching the performance of their closed, proprietary counterparts, making the latter attractive even for the use on highly *private* data. Recently, various new methods have been proposed to adapt closed LLMs to private data without leaking private information to third parties and/or the LLM provider. In this work, we analyze the privacy protection and performance of the four most recent methods for private adaptation of closed LLMs. By examining their threat models and thoroughly comparing their performance under different privacy levels according to differential privacy (DP), various LLM architectures, and multiple datasets for classification and generation tasks, we find that: (1) all the methods leak query data, i.e., the (potentially sensitive) user data that is queried at inference time, to the LLM provider, (2) three out of four methods also leak large fractions of private training data to the LLM provider while the method that protects private data requires a local open LLM, (3) all the methods exhibit lower performance compared to three private gradient-based adaptation methods for *local open LLMs*, and (4) the private adaptation methods for closed LLMs incur higher monetary training and query costs than running the alternative methods on local open LLMs. This yields the conclusion that, to achieve truly *privacy-preserving LLM adaptations* that yield high performance and more privacy at lower costs, taking into account current methods and models, one should use open LLMs.

The softmax gating function is arguably the most popular choice in mixture of experts modeling. Despite its widespread use in practice, the softmax gating may lead to unnecessary competition among experts, potentially causing the undesirable phenomenon of representation collapse due to its inherent structure. In response, the sigmoid gating function has been recently proposed as an alternative and has been demonstrated empirically to achieve superior performance. However, a rigorous examination of the sigmoid gating function is lacking in current literature. In this paper, we verify theoretically that the sigmoid gating, in fact, enjoys a higher sample efficiency than the softmax gating for the statistical task of expert estimation. Towards that goal, we consider a regression framework in which the unknown regression function is modeled as a mixture of experts, and study the rates of convergence of the least squares estimator under the over-specified case in which the number of fitted experts is larger than the true value. We show that two gating regimes naturally arise and, in each of them, we formulate an identifiability condition for the expert functions and derive the corresponding convergence rates. In both cases, we find that experts formulated as feed-forward networks with commonly used activation such as $\mathrm{ReLU}$ and $\mathrm{GELU}$ enjoy faster convergence rates under the sigmoid gating than those under softmax gating. Furthermore, given the same choice of experts, we demonstrate that the sigmoid gating function requires a smaller sample size than its softmax counterpart to attain the same error of expert estimation and, therefore, is more sample efficient.

Aligning neural dynamics with movements is a fundamental goal in neuroscience and brain-machine interfaces. However, there is still a lack of dimensionality reduction methods that can effectively align low-dimensional latent dynamics with movements. To address this gap, we propose Neural Embeddings Rank (NER), a technique that embeds neural dynamics into a 3D latent space and contrasts the embeddings based on movement ranks. NER learns to regress continuous representations of neural dynamics (i.e., embeddings) on continuous movements. We apply NER and six other dimensionality reduction techniques to neurons in the primary motor cortex (M1), dorsal premotor cortex (PMd), and primary somatosensory cortex (S1) as monkeys perform reaching tasks. Only NER aligns latent dynamics with both hand position and direction, visualizable in 3D. NER reveals consistent latent dynamics in M1 and PMd across sixteen sessions over a year. Using a linear regression decoder, NER explains 86\% and 97\% of the variance in velocity and position, respectively. Linear models trained on data from one session successfully decode velocity, position, and direction in held-out test data from different dates and cortical areas (64\%, 88\%, and 90\%). NER also reveals distinct latent dynamics in S1 during consistent movements and in M1 during curved reaching tasks. The code is available at https://github.com/NeuroscienceAI/NER.

Multi-task offline reinforcement learning aims to develop a unified policy for diverse tasks without requiring real-time interaction with the environment. Recent work explores sequence modeling, leveraging the scalability of the transformer architecture as a foundation for multi-task learning. Given the variations in task content and complexity, formulating policies becomes a challenging endeavor, requiring careful parameter sharing and adept management of conflicting gradients to extract rich cross-task knowledge from multiple tasks and transfer it to unseen tasks. In this paper, we propose the Decomposed Prompt Decision Transformer (DPDT) that adopts a two-stage paradigm to efficiently learn prompts for unseen tasks in a parameter-efficient manner. We incorporate parameters from pre-trained language models (PLMs) to initialize DPDT, thereby providing rich prior knowledge encoded in language models. During the decomposed prompt tuning phase, we learn both cross-task and task-specific prompts on training tasks to achieve prompt decomposition. In the test time adaptation phase, the cross-task prompt, serving as a good initialization, were further optimized on unseen tasks through test time adaptation, enhancing the model's performance on these tasks. Empirical evaluation on a series of Meta-RL benchmarks demonstrates the superiority of our approach. The project is available at https://github.com/ruthless-man/DPDT.