ICLR 2024 Papers

Linear attentions have shown promise for improving Transformer efficiency, reducing attention's quadratic complexity to linear in sequence length. This holds exciting promise for (1) training linear Transformers from scratch, (2) `inetuned-conversion of task-specific Transformers into linear versions that recover task performance, and (3) pretrained-conversion of Transformers, such as language models, into linear versions readily finetunable on downstream tasks. However, linear attentions often underperform compared to standard softmax attention. To close this performance gap, we study the behaviors of softmax and linear attentions in various train-from-scratch and finetuned-conversion settings. We find prior linear attentions lack key properties of softmax attention tied to good performance: low-entropy (or spiky) weights and dot-product monotonicity. We further observe surprisingly simple feature maps that retain these properties match softmax performance, but are inefficient to compute in linear attention. We thus propose Hedgehog, a learnable linear attention that retains the spiky and monotonic properties of softmax attention while maintaining linear complexity. Hedgehog uses simple, trainable MLPs to produce attention weights mimicking softmax attention. Experiments show Hedgehog recovers over 99\% of standard Transformer performance in train-from-scratch and finetuned-conversion settings, outperforming prior linear attentions by up to 6 perplexity points on WikiText-103 when training causal GPT models from scratch, and up to 8.7 GLUE score points when converting finetuned bidirectional BERT models. Hedgehog also enables pretrained-conversion. Converting a pretrained GPT-2 into a linear attention variant achieves state-of-the-art 16.7 perplexity on WikiText-103 for 125M subquadratic decoder models. We finally turn a pretrained Llama-2 7B into a viable linear attention Llama. With low-rank adaptation, Hedgehog-Llama-2 7B achieves 28.1 higher ROUGE-1 points over the base standard attention model, where prior linear attentions lead to 16.5 point drops.

Recent advances in depthwise-separable convolutional neural networks (DS-CNNs) have led to novel architectures, that surpass the performance of classical CNNs, by a considerable scalability and accuracy margin. This paper reveals another striking property of DS-CNN architectures: discernible and explainable patterns emerge in their trained depthwise convolutional kernels in all layers. Through an extensive analysis of millions of trained filters, with different sizes and from various models, we employed unsupervised clustering with autoencoders, to categorize these filters. Astonishingly, the patterns converged into a few main clusters, each resembling the difference of Gaussian (DoG) functions, and their first and second-order derivatives. Notably, we classify over 95\% and 90\% of the filters from state-of-the-art ConvNeXtV2 and ConvNeXt models, respectively. This finding is not merely a technological curiosity; it echoes the foundational models neuroscientists have long proposed for the vision systems of mammals. Our results thus deepen our understanding of the emergent properties of trained DS-CNNs and provide a bridge between artificial and biological visual processing systems. More broadly, they pave the way for more interpretable and biologically-inspired neural network designs in the future.

Large-scale Vision Transformers have achieved promising performance on downstream tasks through feature pre-training. However, the performance of vanilla lightweight Vision Transformers (ViTs) is still far from satisfactory compared to that of recent lightweight CNNs or hybrid networks. In this paper, we aim to unlock the potential of vanilla lightweight ViTs by exploring the adaptation of the widely-used re-parameterization technology to ViTs for improving learning ability during training without increasing the inference cost. The main challenge comes from the fact that CNNs perfectly complement with re-parameterization over convolution and batch normalization, while vanilla Transformer architectures are mainly comprised of linear and layer normalization layers. We propose to incorporate the nonlinear ensemble into linear layers by expanding the depth of the linear layers with batch normalization and fusing multiple linear features with hierarchical representation ability through a pyramid structure. We also discover and solve a new transformer-specific distribution rectification problem caused by multi-branch re-parameterization. Finally, we propose our Two-Dimensional Re-parameterized Linear module (TDRL) for ViTs. Under the popular self-supervised pre-training and supervised fine-tuning strategy, our TDRL can be used in these two stages to enhance both generic and task-specific representation. Experiments demonstrate that our proposed method not only boosts the performance of vanilla Vit-Tiny on various vision tasks to new state-of-the-art (SOTA) but also shows promising generality ability on other networks. Code will be available.

It is often advantageous to train models on a subset of the available train examples, because the examples are of variable quality or because one would like to train with fewer examples, without sacrificing performance. We present Gradient Information Optimization (GIO), a scalable, task-agnostic approach to this data selection problem that requires only a small set of (unlabeled) examples representing a target distribution. GIO begins from a natural, information-theoretic objective that is intractable in practice. Our contribution is in showing that it can be made highly scalable through a simple relaxation of the objective and a highly efficient implementation. In experiments with machine translation, spelling correction, and image recognition, we show that GIO delivers outstanding results with very small train sets. These findings are robust to different representation models and hyperparameters for GIO itself. GIO is task- and domain-agnostic and can be applied out-of-the-box to new datasets and domains. We open source a pip-installable implementation of the algorithm as "pip install grad-info-opt".

Multilingual large language models trained on non-parallel data yield impressive translation capabilities. Existing studies demonstrate that incidental sentence-level bilingualism within pre-training data contributes to the LLM's translation abilities. However, it has also been observed that LLM's translation capabilities persist even when incidental sentence-level bilingualism are excluded from the training corpus. In this study, we comprehensively investigate the unreasonable effectiveness and the underlying mechanism for LLM's translation abilities, specifically addressing the question why large language models learn to translate without parallel data, using the BLOOM model series as a representative example. Through extensive experiments, our findings suggest the existence of unintentional bilingualism in the pre-training corpus, especially word alignment data significantly contributes to the large language model's acquisition of translation ability. Moreover, the translation signal derived from word alignment data is comparable to that from sentence-level bilingualism. Additionally, we study the effects of monolingual data and parameter-sharing in assisting large language model to learn to translate. Together, these findings present another piece of the broader puzzle of trying to understand how large language models acquire translation capability.

Deep neural networks (DNNs) on Riemannian manifolds have garnered increasing interest in various applied areas. For instance, DNNs on spherical and hyperbolic manifolds have been designed to solve a wide range of computer vision and nature language processing tasks. One of the key factors that contribute to the success of these networks is that spherical and hyperbolic manifolds have the rich algebraic structures of gyrogroups and gyrovector spaces. This enables principled and effective generalizations of the most successful DNNs to these manifolds. Recently, some works have shown that many concepts in the theory of gyrogroups and gyrovector spaces can also be generalized to matrix manifolds such as Symmetric Positive Definite (SPD) and Grassmann manifolds. As a result, some building blocks for SPD and Grassmann neural networks, e.g., isometric models and multinomial logistic regression (MLR) can be derived in a way that is fully analogous to their spherical and hyperbolic counterparts. Building upon these works, in this paper, we design fully-connected (FC) and convolutional layers for SPD neural networks. We also develop MLR on Symmetric Positive Semi-definite (SPSD) manifolds, and propose a method for performing backpropagation with the Grassmann logarithmic map in the projector perspective. We demonstrate the effectiveness of the proposed approach in the human action recognition and node classification tasks.

Human visual recognition is a sparse process, where only a few salient visual cues are attended to rather than every detail being traversed uniformly. However, most current vision networks follow a dense paradigm, processing every single visual unit (such as pixels or patches) in a uniform manner. In this paper, we challenge this dense convention and present a new vision transformer, coined SparseFormer, to explicitly imitate human's sparse visual recognition in an end-to-end manner. SparseFormer learns to represent images using a highly limited number of tokens (e.g., down to $9$) in the latent space with sparse feature sampling procedure instead of processing dense units in the original image space. Therefore, SparseFormer circumvents most of dense operations on the image space and has much lower computational costs. Experiments on the ImageNet-1K classification show that SparseFormer delivers performance on par with canonical or well-established models while offering more favorable accuracy-throughput tradeoff. Moreover, the design of our network can be easily extended to the video classification task with promising performance with lower compute. We hope our work can provide an alternative way for visual modeling and inspire further research on sparse vision architectures. Code and weights are available at https://github.com/showlab/sparseformer.

Noise poses a widespread challenge in signal processing, particularly when it comes to denoising images. Although convolutional neural networks (CNNs) have exhibited remarkable success in this field, they are predicated upon the belief that noise follows established distributions, which restricts their practicality when dealing with real-world noise. To overcome this limitation, several efforts have been taken to collect noisy image datasets from the real world. Generative methods, employing techniques such as generative adversarial networks (GANs) and normalizing flows (NFs), have emerged as a solution for generating realistic noisy images. Recent works model noise using camera metadata, however requiring metadata even for sampling phase. In contrast, in this work, we aim to estimate the underlying camera settings, enabling us to improve noise modeling and generate diverse noise distributions. To this end, we introduce a new NF framework that allows us to both classify noise based on camera settings and generate various noisy images. Through experimental results, our model demonstrates exceptional noise quality and leads in denoising performance on benchmark datasets.

Large-scale contrastive vision-language pre-trained models provide the zero-shot model achieving competitive performance across a range of image classification tasks without requiring training on downstream data. Recent works have confirmed that while additional fine-tuning of the zero-shot model on the reference data results in enhanced downstream performance, it compromises the model's robustness against distribution shifts. Our investigation begins by examining the conditions required to achieve the goals of robust fine-tuning, employing descriptions based on feature distortion theory and joint energy-based models. Subsequently, we propose a novel robust fine-tuning algorithm, Lipsum-FT, that effectively utilizes the language modeling aspect of the vision-language pre-trained models. Extensive experiments conducted on distribution shift scenarios in DomainNet and ImageNet confirm the superiority of our proposed Lipsum-FT approach over existing robust fine-tuning methods.

The recent GPT-4 has demonstrated extraordinary multi-modal abilities, such as directly generating websites from handwritten text and identifying humorous elements within images. These features are rarely observed in previous vision-language models. However, the technical details behind GPT-4 continue to remain undisclosed. We believe that the enhanced multi-modal generation capabilities of GPT-4 stem from the utilization of sophisticated large language models (LLM). To examine this phenomenon, we present MiniGPT-4, which aligns a frozen visual encoder with a frozen advanced LLM, Vicuna, using one projection layer. Our work, for the first time, uncovers that properly aligning the visual features with an advanced large language model can possess numerous advanced multi-modal abilities demonstrated by GPT-4, such as detailed image description generation and website creation from hand-drawn drafts. Furthermore, we also observe other emerging capabilities in MiniGPT-4, including writing stories and poems inspired by given images, teaching users how to cook based on food photos, and so on. In our experiment, we found that the model trained on short image caption pairs could produce unnatural language outputs (e.g., repetition and fragmentation). To address this problem, we curate a detailed image description dataset in the second stage to finetune the model, which consequently improves the model's generation reliability and overall usability.

The popularity of LLaMA (Touvron et al., 2023a;b) and other recently emerged moderate-sized large language models (LLMs) highlights the potential of building smaller yet powerful LLMs. Regardless, the cost of training such models from scratch on trillions of tokens remains high. In this work, we study structured pruning as an effective means to develop smaller LLMs from pre-trained, larger models. Our approach employs two key techniques: (1) targeted structured pruning, which prunes a larger model to a specified target shape by removing layers, heads, and intermediate and hidden dimensions in an end-to-end manner, and (2) dynamic batch loading, which dynamically updates the composition of sampled data in each training batch based on varying losses across different domains. We demonstrate the efficacy of our approach by presenting the Sheared-LLaMA series, pruning the LLaMA2-7B model down to 1.3B and 2.7B parameters. Sheared-LLaMA models outperform state-of-the-art open-source models of equivalent sizes, such as Pythia, INCITE, OpenLLaMA and the concurrent TinyLlama models, on a wide range of downstream and instruction tuning evaluations, while requiring only 3% of compute compared to training such models from scratch. This work provides compelling evidence that leveraging existing LLMs with structured pruning is a far more cost-effective approach for building competitive small-scale LLMs