ICML 2024 Papers

Despite the widespread adoption of prompting, prompt tuning and prefix-tuning of transformer models, our theoretical understanding of these fine-tuning methods remains limited. A key question is whether one can arbitrarily modify the behavior of a pretrained model by prompting or prefix-tuning it. Formally, whether prompting and prefix-tuning a pretrained model can universally approximate sequence-to-sequence functions. This paper answers in the affirmative and demonstrates that much smaller pretrained models than previously thought can be universal approximators when prefixed. In fact, prefix-tuning a single attention head is sufficient to approximate any continuous function making the attention mechanism uniquely suited for universal approximation. Moreover, any sequence-to-sequence function can be approximated by prefixing a transformer with depth linear in the sequence length. Beyond these density-type results, we also offer Jackson-type bounds on the length of the prefix needed to approximate a function to a desired precision.

Among the widely used parameter-efficient fine-tuning (PEFT) methods, LoRA and its variants have gained considerable popularity because of avoiding additional inference costs. However, there still often exists an accuracy gap between these methods and full fine-tuning (FT). In this work, we first introduce a novel weight decomposition analysis to investigate the inherent differences between FT and LoRA. Aiming to resemble the learning capacity of FT from the findings, we propose Weight-Decomposed Low-Rank Adaptation (DoRA). DoRA decomposes the pre-trained weight into two components, magnitude and direction, for fine-tuning, specifically employing LoRA for directional updates to efficiently minimize the number of trainable parameters. By employing DoRA, we enhance both the learning capacity and training stability of LoRA while avoiding any additional inference overhead. DoRA consistently outperforms LoRA on fine-tuning LLaMA, LLaVA, and VL-BART on various downstream tasks, such as commonsense reasoning, visual instruction tuning, and image/video-text understanding. The code is available at https://github.com/NVlabs/DoRA.

Visual prompt tuning (VPT) is a promising solution incorporating learnable prompt tokens to customize pre-trained models for downstream tasks. However, VPT and its variants often encounter challenges like prompt initialization, prompt length, and subpar performance in self-supervised pretraining, hindering successful contextual adaptation. This study commences by exploring the correlation evolvement between prompts and patch tokens during proficient training. Inspired by the observation that the prompt tokens tend to share high mutual information with patch tokens, we propose initializing prompts with downstream token prototypes. The strategic initialization, a stand-in for the previous initialization, substantially improves performance. To refine further, we optimize token construction with a streamlined pipeline that maintains excellent performance with almost no increase in computational expenses compared to VPT. Exhaustive experiments show our proposed approach outperforms existing methods by a remarkable margin. For instance, after MAE pre-training, our method improves accuracy by up to 10%$\sim$30% compared to VPT, and outperforms Full fine-tuning 19 out of 24 cases while using less than 0.4% of learnable parameters. Besides, the experimental results demonstrate the proposed SPT is robust to prompt lengths and scales well with model capacity and training data size. We finally provide an insightful exploration into the amount of target data facilitating the adaptation of pre-trained models to downstream tasks. The code is available at https://github.com/WangYZ1608/Self-Prompt-Tuning.

With the increasingly powerful performances and enormous scales of pretrained models, promoting parameter efficiency in fine-tuning has become a crucial need for effective and efficient adaptation to various downstream tasks. One representative line of fine-tuning methods is Orthogonal Fine-tuning (OFT), which rigorously preserves the angular distances within the parameter space to preserve the pretrained knowledge. Despite the empirical effectiveness, OFT still suffers low parameter efficiency at $\mathcal{O}(d^2)$ and limited capability of downstream adaptation. Inspired by Givens rotation, in this paper, we proposed quasi-Givens Orthogonal Fine-Tuning (qGOFT) to address the problems. We first use $\mathcal{O}(d)$ Givens rotations to accomplish arbitrary orthogonal transformation in $SO(d)$ with provable equivalence, reducing parameter complexity from $\mathcal{O}(d^2)$ to $\mathcal{O}(d)$. Then we introduce flexible norm and relative angular adjustments under soft orthogonality regularization to enhance the adaptation capability of downstream semantic deviations. Extensive experiments on various tasks and pretrained models validate the effectiveness of our methods.

We introduce a fine-grained framework for uncertainty quantification of predictive models under distributional shifts. This framework distinguishes the shift in covariate distributions from that in the conditional relationship between the outcome ($Y$) and the covariates ($X$). We propose to reweight the training samples to adjust for an identifiable shift in covariate distribution while protecting against the worst-case conditional distribution shift bounded in an $f$-divergence ball. Based on ideas from conformal inference and distributionally robust learning, we present an algorithm that outputs (approximately) valid and efficient prediction intervals in the presence of distributional shifts. As a use case, we apply the framework to sensitivity analysis of individual treatment effects with hidden confounding. The proposed methods are evaluated in simulations and four real data applications, demonstrating superior robustness and efficiency compared with existing benchmarks.

The sparsely gated mixture of experts (MoE) architecture sends different inputs to different subnetworks (experts), through trainable routers. MoE reduces the training computation significantly for large models, but its deployment can be still memory/computation expensive for some downstream tasks. Model pruning is a popular approach to reduce inference computation, but its application in MoE architecture is largely unexplored. To the best of our knowledge, this paper provides the first provably efficient technique for pruning experts in fine-tuned MoE models. We theoretically prove that prioritizing the pruning of the experts with a smaller change of the router’s $l_2$ norm from the pre-trained model guarantees the preservation of test accuracy, while significantly reducing the model size and the computational requirements. Although our theoretical analysis is centered on binary classification tasks on simplified MoE architecture, our expert pruning method is verified on large vision MoE models such as V-MoE and $\text{E}^3$-MoE fine-tuned on benchmark datasets such as CIFAR-10, CIFAR-100, and ImageNet.

Domain incremental learning (DIL) poses a significant challenge in real-world scenarios, as models need to be sequentially trained on diverse domains over time, all the while avoiding catastrophic forgetting. Mitigating representation drift, which refers to the phenomenon of learned representations undergoing changes as the model adapts to new tasks, can help alleviate catastrophic forgetting. In this study, we propose a novel DIL method named *DARE*, featuring a three-stage training process: Divergence, Adaptation, and REfinement. This process gradually adapts the representations associated with new tasks into the feature space spanned by samples from previous tasks, simultaneously integrating task-specific decision boundaries. Additionally, we introduce a novel strategy for buffer sampling and demonstrate the effectiveness of our proposed method, combined with this sampling strategy, in reducing representation drift within the feature encoder. This contribution effectively alleviates catastrophic forgetting across multiple DIL benchmarks. Furthermore, our approach prevents sudden representation drift at task boundaries, resulting in a well-calibrated DIL model that maintains the performance on previous tasks.

With the prevalence of pre-training-fine-tuning paradigm, how to efficiently adapt the pre-trained model to the downstream tasks has been an intriguing issue. $\textbf{P}$arameter-$\textbf{E}$fficient $\textbf{F}$ine-$\textbf{T}$uning(PEFT) methods have been proposed for low-cost adaptation. Although PEFT has demonstrated effectiveness and been widely applied, the underlying principles are still unclear. In this paper, we adopt the PAC-Bayesian generalization error bound, viewing pre-training as a shift of prior distribution which leads to a tighter bound for generalization error. We validate this shift from the perspectives of oscillations in the loss landscape and the quasi-sparsity in gradient distribution. Based on this, we propose a gradient-based sparse fine-tuning algorithm, named $\textbf{S}$parse $\textbf{I}$ncrement $\textbf{F}$ine-$\textbf{T}$uning(SIFT), and validate its effectiveness on a range of tasks including the GLUE Benchmark and Instruction-tuning. The code is accessible at https://github.com/song-wx/SIFT/.

There is a consensus that instruction fine-tuning of LLMs requires high-quality data, but what are they? LIMA (NeurIPS 2023) and AlpaGasus (ICLR 2024) are state-of-the-art methods for selecting such high-quality examples, either via manual curation or using GPT-3.5-Turbo as a quality scorer. We show that the extremely simple baseline of selecting the 1,000 instructions with longest responses---that intuitively contain more learnable information and are harder to overfit---from standard datasets can consistently outperform these sophisticated methods according to GPT-4 and PaLM-2 as judges, while remaining competitive on the Open LLM benchmarks that test factual knowledge. We demonstrate this for several LLMs (Llama-2-7B, Llama-2-13B, Mistral-7B-v0.1) and datasets (Alpaca-52k, Evol-Instruct-70k). In addition, a lightweight refinement of such long instructions can further improve the abilities of the fine-tuned LLMs, and allows us to obtain competitive results on MT-Bench and the 2nd highest-ranked Llama-2-7B-based model on AlpacaEval 2.0, while training on only 1,000 examples and no extra preference data. We also conduct a thorough analysis of our models to ensure that their enhanced performance is not simply due to GPT-4's preference for longer responses. Overall, our findings suggest that fine-tuning on the longest responses should be the default baseline for any work on instruction fine-tuning. We provide our code in this GitHub repository.