ICLR 2024 Papers

More distinguishable and consistent pixel features for each category will benefit the semantic segmentation under various settings. Existing efforts to mine better pixel-level features attempt to explicitly model the categorical distribution, which fails to achieve optimal due to the significant pixel feature variance. Moreover, prior research endeavors have scarcely delved into the thorough analysis and meticulous handling of pixel-level variance, leaving semantic segmentation at a coarse granularity. In this work, we analyze the causes of pixel-level variance and introduce the concept of $\textbf{pixel learning}$ to concentrate on the tailored learning process of pixels to handle the pixel-level variance, enhancing the per-pixel recognition capability of segmentation models. Under the context of the pixel learning scheme, each image is viewed as a distribution of pixels, and pixel learning aims to pursue consistent pixel representation inside an image, continuously align pixels from different images (distributions), and eventually achieve consistent pixel representation for each category, even cross-domains. We proposed a pure pixel-level learning framework, namely PiXL, which consists of a pixel partition module to divide pixels into sub-domains, a prototype generation, a selection module to prepare targets for subsequent alignment, and a pixel alignment module to guarantee pixel feature consistency intra-/inter-images, and inter-domains. Extensive evaluations of multiple learning paradigms, including unsupervised domain adaptation and semi-/fully-supervised segmentation, show that PiXL outperforms state-of-the-art performances, especially when annotated images are scarce. Visualization of the embedding space further demonstrates that pixel learning attains a superior representation of pixel features. The code is available at https://github.com/ChenGan-JS/PiXL.

Similarity-based representation learning has shown impressive capabilities in both supervised (e.g., metric learning) and unsupervised (e.g., contrastive learning) scenarios. Existing approaches effectively constrained the representation difference (i.e., the disagreement between the embeddings of two instances) to fit the corresponding (pseudo) similarity supervision. However, most of them can hardly restrict the variation of representation difference, sometimes leading to overfitting results where the clusters are disordered by drastically changed differences. In this paper, we thus propose a novel difference alignment regularization (DAR) to encourage all representation differences between inter-class instances to be as close as possible, so that the learning algorithm can produce consistent differences to distinguish data points from each other. To this end, we construct a new cross-total-variation (CTV) norm to measure the divergence among representation differences, and we convert it into an equivalent stochastic form for easy optimization. Then, we integrate the proposed regularizer into the empirical loss for difference-aligned similarity learning (DASL), shrinking the hypothesis space and alleviating overfitting. Theoretically, we prove that our regularizer tightens the error bound of the traditional similarity learning. Experiments on multi-domain data demonstrate the superiority of DASL over existing approaches in both supervised metric learning and unsupervised contrastive learning tasks.

We investigate contrastive learning in the federated setting through the lens of Sim- CLR and multi-view mutual information maximization. In doing so, we uncover a connection between contrastive representation learning and user verification; by adding a user verification loss to each client’s local SimCLR loss we recover a lower bound to the global multi-view mutual information. To accommodate for the case of when some labelled data are available at the clients, we extend our SimCLR variant to the federated semi-supervised setting. We see that a supervised SimCLR objective can be obtained with two changes: a) the contrastive loss is computed between datapoints that share the same label and b) we require an additional auxiliary head that predicts the correct labels from either of the two views. Along with the proposed SimCLR extensions, we also study how different sources of non-i.i.d.-ness can impact the performance of federated unsupervised learning through global mutual information maximization; we find that a global objective is beneficial for some sources of non-i.i.d.-ness but can be detrimental for others. We empirically evaluate our proposed extensions in various tasks to validate our claims and furthermore demonstrate that our proposed modifications generalize to other pretraining methods.

Context-based fine-tuning methods, including prompting, in-context learning, soft prompting (also known as prompt tuning), and prefix-tuning, have gained popularity due to their ability to often match the performance of full fine-tuning with a fraction of the parameters. Despite their empirical successes, there is little theoretical understanding of how these techniques influence the internal computation of the model and their expressiveness limitations. We show that despite the continuous embedding space being more expressive than the discrete token space, soft-prompting and prefix-tuning are potentially less expressive than full fine-tuning, even with the same number of learnable parameters. Concretely, context-based fine-tuning cannot change the relative attention pattern over the content and can only bias the outputs of an attention layer in a fixed direction. This suggests that while techniques like prompting, in-context learning, soft prompting, and prefix-tuning can effectively elicit skills present in the pretrained model, they may not be able to learn novel tasks that require new attention patterns.

Multimodal patient representation learning aims to integrate information from multiple modalities and generate comprehensive patient representations for subsequent clinical predictive tasks. However, many existing approaches either presuppose the availability of all modalities and labels for each patient or only deal with missing modalities. In reality, patient data often comes with both missing modalities and labels for various reasons (i.e., the missing modality and label issue). Moreover, multimodal models might over-rely on certain modalities, causing sub-optimal performance when these modalities are absent (i.e., the modality collapse issue). To address these issues, we introduce MUSE: a mutual-consistent graph contrastive learning method. MUSE uses a flexible bipartite graph to represent the patient-modality relationship, which can adapt to various missing modality patterns. To tackle the modality collapse issue, MUSE learns to focus on modality-general and label-decisive features via a mutual-consistent contrastive learning loss. Notably, the unsupervised component of the contrastive objective only requires self-supervision signals, thereby broadening the training scope to incorporate patients with missing labels. We evaluate MUSE on three publicly available datasets: MIMIC-IV, eICU, and ADNI. Results show that MUSE outperforms all baselines, and MUSE+ further elevates the absolute improvement to ~4% by extending the training scope to patients with absent labels.

We study offline reinforcement learning (RL) with general function approximation. General function approximation is a powerful tool for algorithm design and analysis, but its adaptation to offline RL encounters several challenges due to varying approximation targets and assumptions that blur the real meanings of function assumptions. In this paper, we try to formulate and clarify the treatment of general function approximation in offline RL in two aspects: (1) analyzing different types of assumptions and their practical usage, and (2) understanding its role as a restriction on underlying MDPs from information-theoretic perspectives. Additionally, we introduce a new insight for lower bound establishing: one can exploit model-realizability to establish general-purpose lower bounds that can be generalized into other functions. Building upon this insight, we propose two generic lower bounds that contribute to a better understanding of offline RL with general function approximation.

The extrapolation capability of Large Language Models (LLMs) based on Rotary Position Embedding \citep{su2021roformer} is currently a topic of considerable interest. The mainstream approach to addressing extrapolation with LLMs involves modifying RoPE by replacing 10000, the rotary base of $\theta_n={10000}^{-2n/d}$ in the original RoPE, with a larger value and providing longer fine-tuning text. In this work, we first observe that fine-tuning a RoPE-based LLM with either a smaller or larger base in pre-training context length could significantly enhance its extrapolation performance. After that, we propose \textbf{\textit{Scaling Laws of RoPE-based Extrapolation}}, a unified framework from the periodic perspective, to describe the relationship between the extrapolation performance and base value as well as tuning context length. In this process, we also explain the origin of the RoPE-based extrapolation issue by \textbf{\textit{critical dimension for extrapolation}}. Besides these observations and analyses, we achieve extrapolation up to 1 million context length within only 16K training length on LLaMA2 7B and 13B \citep{touvron2023llama2}.

We study the estimation of a planted signal hidden in a recently introduced nested matrix-tensor model, which is an extension of the classical spiked rank-one tensor model, motivated by multi-view clustering. Prior work has theoretically examined the performance of a tensor-based approach, which relies on finding a best rank-one approximation, a problem known to be computationally hard. A tractable alternative approach consists in computing instead the best rank-one (matrix) approximation of an unfolding of the observed tensor data, but its performance was hitherto unknown. We quantify here the performance gap between these two approaches, in particular by deriving the precise algorithmic threshold of the unfolding approach and demonstrating that it exhibits a BBP-type transition behavior. This work is therefore in line with recent contributions which deepen our understanding of why tensor-based methods surpass matrix-based methods in handling structured tensor data.

Recently, semidefinite programming (SDP) techniques have shown great promise in providing accurate Lipschitz bounds for neural networks. Specifically, the LipSDP approach (Fazlyab et al., 2019) has received much attention and provides the least conservative Lipschitz upper bounds that can be computed with polynomial time guarantees. However, one main restriction of LipSDP is that its formulation requires the activation functions to be slope-restricted on $[0,1]$, preventing its further use for more general activation functions such as GroupSort, MaxMin, and Householder. One can rewrite MaxMin activations for example as residual ReLU networks. However, a direct application of LipSDP to the resultant residual ReLU networks is conservative and even fails in recovering the well-known fact that the MaxMin activation is 1-Lipschitz. Our paper bridges this gap and extends LipSDP beyond slope-restricted activation functions. To this end, we provide novel quadratic constraints for GroupSort, MaxMin, and Householder activations via leveraging their underlying properties such as sum preservation. Our proposed analysis is general and provides a unified approach for estimating $\ell_2$ and $\ell_\infty$ Lipschitz bounds for a rich class of neural network architectures, including non-residual and residual neural networks and implicit models, with GroupSort, MaxMin, and HouseHolder activations. Finally, we illustrate the utility of our approach with a variety of experiments and show that our proposed SDPs generate less conservative Lipschitz bounds in comparison to existing approaches.

Model overconfidence and poor calibration are common in machine learning and difficult to account for when applying standard empirical risk minimization. In this work, we propose a novel method to alleviate these problems that we call odd-$k$-out learning (OKO), which minimizes the cross-entropy error for sets rather than for single examples. This naturally allows the model to capture correlations across data examples and achieves both better accuracy and calibration, especially in limited training data and class-imbalanced regimes. Perhaps surprisingly, OKO often yields better calibration even when training with hard labels and dropping any additional calibration parameter tuning, such as temperature scaling. We demonstrate this in extensive experimental analyses and provide a mathematical theory to interpret our findings. We emphasize that OKO is a general framework that can be easily adapted to many settings and a trained model can be applied to single examples at inference time, without significant run-time overhead or architecture changes.

The Language of Thought Hypothesis suggests that human cognition operates on a structured, language-like system of mental representations. While neural language models can naturally benefit from the compositional structure inherently and explicitly expressed in language data, learning such representations from non-linguistic general observations, like images, remains a challenge. In this work, we introduce the Neural Language of Thought Model (NLoTM), a novel approach for unsupervised learning of LoTH-inspired representation and generation. NLoTM comprises two key components: (1) the Semantic Vector-Quantized Variational Autoencoder, which learns hierarchical, composable discrete representations aligned with objects and their properties, and (2) the Autoregressive LoT Prior, an autoregressive transformer that learns to generate semantic concept tokens compositionally, capturing the underlying data distribution. We evaluate NLoTM on several 2D and 3D image datasets, demonstrating superior performance in downstream tasks, out-of-distribution generalization, and image generation quality compared to patch-based VQ-VAE and continuous object-centric representations. Our work presents a significant step towards creating neural networks exhibiting more human-like understanding by developing LoT-like representations and offers insights into the intersection of cognitive science and machine learning.

Learning compositional representation is a key aspect of object-centric learning as it enables flexible systematic generalization and supports complex visual reasoning. However, most of the existing approaches rely on auto-encoding objective, while the compositionality is implicitly imposed by the architectural or algorithmic bias in the encoder. This misalignment between auto-encoding objective and learning compositionality often results in failure of capturing meaningful object representations. In this study, we propose a novel objective that explicitly encourages compositionality of the representations. Built upon the existing object-centric learning framework (e.g., slot attention), our method incorporates additional constraints that an arbitrary mixture of object representations from two images should be valid by maximizing the likelihood of the composite data. We demonstrate that incorporating our objective to the existing framework consistently improves the objective-centric learning and enhances the robustness to the architectural choices.

Transformers have demonstrated effectiveness in in-context solving data-fitting problems from various (latent) models, as reported by Garg et al. (2022). However, the absence of an inherent iterative structure in the transformer architecture presents a challenge in emulating the iterative algorithms, which are commonly employed in traditional machine learning methods. To address this, we propose the utilization of looped transformer architecture and its associated training methodology, with the aim of incorporating iterative characteristics into the transformer architectures. Experimental results suggest that the looped transformer achieves performance comparable to the standard transformer in solving various data-fitting problems, while utilizing less than 10% of the parameter count.

Classification models are ubiquitously deployed in society and necessitate high utility, fairness, and robustness performance. Current research efforts mainly focus on improving model architectures and learning algorithms on fixed datasets to achieve this goal. In contrast, in this paper, we address an orthogonal yet crucial problem: given a fixed convex learning model (or a convex surrogate for a non-convex model) and a function of interest, we assess what data benefits the model by interpreting the feature space, and then aim to improve performance as measured by this function. To this end, we propose the use of influence estimation models for interpreting the classifier's performance from the perspective of the data feature space. Additionally, we propose data selection approaches based on influence that enhance model utility, fairness, and robustness. Through extensive experiments on synthetic and real-world datasets, we validate and demonstrate the effectiveness of our approaches not only for conventional classification scenarios, but also under more challenging scenarios such as distribution shifts, fairness poisoning attacks, utility evasion attacks, online learning, and active learning.

As a sub-discipline of evolutionary and computational linguistics, emergent communication (EC) studies communication protocols, called emergent languages, arising in simulations where agents communicate. A key goal of EC is to give rise to languages that share statistical properties with natural languages. In this paper, we reinterpret Lewis's signaling game, a frequently used setting in EC, as beta-VAE and reformulate its objective function as ELBO. Consequently, we clarify the existence of prior distributions of emergent languages and show that the choice of the priors can influence their statistical properties. Specifically, we address the properties of word lengths and segmentation, known as Zipf's law of abbreviation (ZLA) and Harris's articulation scheme (HAS), respectively. It has been reported that the emergent languages do not follow them when using the conventional objective. We experimentally demonstrate that by selecting an appropriate prior distribution, more natural segments emerge, while suggesting that the conventional one prevents the languages from following ZLA and HAS.

Many complex tasks can be decomposed into simpler, independent parts. Discovering such underlying compositional structure has the potential to enable compositional generalization. Despite progress, our most powerful systems struggle to compose flexibly. It therefore seems natural to make models more modular to help capture the compositional nature of many tasks. However, it is unclear under which circumstances modular systems can discover hidden compositional structure. To shed light on this question, we study a teacher-student setting with a modular teacher where we have full control over the composition of ground truth modules. This allows us to relate the problem of compositional generalization to that of identification of the underlying modules. In particular we study modularity in hypernetworks representing a general class of multiplicative interactions. We show theoretically that identification up to linear transformation purely from demonstrations is possible without having to learn an exponential number of module combinations. We further demonstrate empirically that under the theoretically identified conditions, meta-learning from finite data can discover modular policies that generalize compositionally in a number of complex environments.

An effective technique for obtaining high-quality representations is adding a projection head on top of the encoder during training, then discarding it and using the pre-projection representations. Despite its proven practical effectiveness, the reason behind the success of this technique is poorly understood. The pre-projection representations are not directly optimized by the loss function, raising the question: what makes them better? In this work, we provide a rigorous theoretical answer to this question. We start by examining linear models trained with self-supervised contrastive loss. We reveal that the implicit bias of training algorithms leads to layer-wise progressive feature weighting, where features become increasingly unequal as we go deeper into the layers. Consequently, lower layers tend to have more normalized and less specialized representations. We theoretically characterize scenarios where such representations are more beneficial, highlighting the intricate interplay between data augmentation and input features. Additionally, we demonstrate that introducing non-linearity into the network allows lower layers to learn features that are completely absent in higher layers. Finally, we show how this mechanism improves the robustness in supervised contrastive learning and supervised learning. We empirically validate our results through various experiments on CIFAR-10/100, UrbanCars and shifted versions of ImageNet. We also introduce a potential alternative to projection head, which offers a more interpretable and controllable design.

Tabular data have been playing a mostly important role in diverse real-world fields, such as healthcare, engineering, finance, etc. With the recent success of deep learning, many tabular machine learning (ML) methods based on deep networks (e.g., Transformer, ResNet) have achieved competitive performance on tabular benchmarks. However, existing deep tabular ML methods suffer from the representation entanglement and localization, which largely hinders their prediction performance and leads to performance inconsistency on tabular tasks. To overcome these problems, we explore a novel direction of applying prototype learning for tabular ML and propose a prototype-based tabular representation learning framework, PTaRL, for tabular prediction tasks. The core idea of PTaRL is to construct prototype-based projection space (P-Space) and learn the disentangled representation around global data prototypes. Specifically, PTaRL mainly involves two stages: (i) Prototype Generating, that constructs global prototypes as the basis vectors of P-Space for representation, and (ii) Prototype Projecting, that projects the data samples into P-Space and keeps the core global data information via Optimal Transport. Then, to further acquire the disentangled representations, we constrain PTaRL with two strategies: (i) to diversify the coordinates towards global prototypes of different representations within P-Space, we bring up a diversifying constraint for representation calibration; (ii) to avoid prototype entanglement in P-Space, we introduce a matrix orthogonalization constraint to ensure the independence of global prototypes. Finally, we conduct extensive experiments in PTaRL coupled with state-of-the-art deep tabular ML models on various tabular benchmarks and the results have shown our consistent superiority.

Large language models (LLMs) are often augmented with tools to solve complex tasks. By generating code snippets and executing them through task-specific Application Programming Interfaces (APIs), they can offload certain functions to dedicated external modules, such as image encoding and performing calculations. However, most existing approaches to augment LLMs with tools are constrained by general-purpose APIs and lack the flexibility for tailoring them to specific tasks. In this work, we present CRAFT, a general tool creation and retrieval framework for LLMs. It creates toolsets specifically curated for the tasks and equips LLMs with a component that retrieves tools from these sets to enhance their capability to solve complex tasks. For each task, we collect specific code solutions by prompting GPT-4 to solve the training examples. Following a validation step ensuring the correctness, these solutions are abstracted into code snippets to enhance reusability, and deduplicated for higher quality. At inference time, the language model retrieves snippets from the toolsets and then executes them or generates the output conditioning on the retrieved snippets. Our method is designed to be flexible and offers a plug-and-play approach to adapt off-the-shelf LLMs to unseen domains and modalities, without any finetuning. Experiments on vision-language, tabular processing, and mathematical reasoning tasks show that our approach achieves substantial improvements compared to strong baselines. In addition, our in-depth analysis reveals that: (1) consistent performance improvement can be achieved by scaling up the number of tools and the capability of the backbone models; (2) each component of our approach contributes to the performance gains; (3) the created tools are well-structured and reliable with low complexity and atomicity.