NeurIPS 2024 Papers

The state-of-the-art feature attribution methods often neglect the influence of unobservable confounders, posing a risk of misinterpretation, especially when it is crucial for the interpretation to remain faithful to the data. To counteract this, we propose a new approach, data-faithful feature attribution, which trains a confounder-free model using instrumental variables. The cluttered effects of unobservable confounders in a model trained as such are decoupled from input features, thereby aligning the output of the model with the contribution of input features to the target feature in the data generation. Furthermore, feature attribution results produced by our method are more robust when focusing on attributions from the perspective of data generation. Our experiments on both synthetic and real-world datasets demonstrate the effectiveness of our approaches.

As Large Language Models (LLMs) demonstrate impressive capabilities, demystifying their internal mechanisms becomes increasingly vital. Neuron attribution, which attributes LLM outputs to specific neurons to reveal the semantic properties they learn, has emerged as a key interpretability approach. However, while neuron attribution has made significant progress in deciphering text-only LLMs, its application to Multimodal LLMs (MLLMs) remains less explored. To address this gap, we propose a novel Neuron Attribution method tailored for MLLMs, termed NAM. Specifically, NAM not only reveals the modality-specific semantic knowledge learned by neurons within MLLMs, but also highlights several intriguing properties of neurons, such as cross-modal invariance and semantic sensitivity. These properties collectively elucidate the inner workings mechanism of MLLMs, providing a deeper understanding of how MLLMs process and generate multi-modal content. Through theoretical analysis and empirical validation, we demonstrate the efficacy of NAM and the valuable insights it offers. Furthermore, leveraging NAM, we introduce a multi-modal knowledge editing paradigm, underscoring the practical significance of our approach for downstream applications of MLLMs.

Concept Bottleneck Models (CBMs) have emerged as a promising interpretable method whose final prediction is based on intermediate, human-understandable concepts rather than the raw input. Through time-consuming manual interventions, a user can correct wrongly predicted concept values to enhance the model's downstream performance. We propose *Stochastic Concept Bottleneck Models* (SCBMs), a novel approach that models concept dependencies. In SCBMs, a single-concept intervention affects all correlated concepts, thereby improving intervention effectiveness. Unlike previous approaches that model the concept relations via an autoregressive structure, we introduce an explicit, distributional parameterization that allows SCBMs to retain the CBMs' efficient training and inference procedure. Additionally, we leverage the parameterization to derive an effective intervention strategy based on the confidence region. We show empirically on synthetic tabular and natural image datasets that our approach improves intervention effectiveness significantly. Notably, we showcase the versatility and usability of SCBMs by examining a setting with CLIP-inferred concepts, alleviating the need for manual concept annotations.

Pre-trained large language models (LLMs) exhibit impressive mathematical reasoning capabilities, yet how they compute basic arithmetic, such as addition, remains unclear. This paper shows that pre-trained LLMs add numbers using Fourier features---dimensions in the hidden state that represent numbers via a set of features sparse in the frequency domain. Within the model, MLP and attention layers use Fourier features in complementary ways: MLP layers primarily approximate the magnitude of the answer using low-frequency features, while attention layers primarily perform modular addition (e.g., computing whether the answer is even or odd) using high-frequency features. Pre-training is crucial for this mechanism: models trained from scratch to add numbers only exploit low-frequency features, leading to lower accuracy. Introducing pre-trained token embeddings to a randomly initialized model rescues its performance. Overall, our analysis demonstrates that appropriate pre-trained representations (e.g., Fourier features) can unlock the ability of Transformers to learn precise mechanisms for algorithmic tasks.

We present ProtoViT, a method for interpretable image classification combining deep learning and case-based reasoning. This method classifies an image by comparing it to a set of learned prototypes, providing explanations of the form ``this looks like that.'' In our model, a prototype consists of **parts**, which can deform over irregular geometries to create a better comparison between images. Unlike existing models that rely on Convolutional Neural Network (CNN) backbones and spatially rigid prototypes, our model integrates Vision Transformer (ViT) backbones into prototype based models, while offering spatially deformed prototypes that not only accommodate geometric variations of objects but also provide coherent and clear prototypical feature representations with an adaptive number of prototypical parts. Our experiments show that our model can generally achieve higher performance than the existing prototype based models. Our comprehensive analyses ensure that the prototypes are consistent and the interpretations are faithful.

AI models are increasingly prevalent in high-stakes environments, necessitating thorough assessment of their capabilities and risks. Benchmarks are popular for measuring these attributes and for comparing model performance, tracking progress, and identifying weaknesses in foundation and non-foundation models. They can inform model selection for downstream tasks and influence policy initiatives. However, not all benchmarks are the same: their quality depends on their design and usability. In this paper, we develop an assessment framework considering 40 best practices across a benchmark's life cycle and evaluate 25 AI benchmarks against it. We find that there exist large quality differences and that commonly used benchmarks suffer from significant issues. We further find that most benchmarks do not report statistical significance of their results nor can results be easily replicated. To support benchmark developers in aligning with best practices, we provide a checklist for minimum quality assurance based on our assessment. We also develop a living repository of benchmark assessments to support benchmark comparability.

As new machine learning methods demand larger training datasets, researchers and developers face significant challenges in dataset management. Although ethics reviews, documentation, and checklists have been established, it remains uncertain whether consistent dataset management practices exist across the community. This lack of a comprehensive overview hinders our ability to diagnose and address fundamental tensions and ethical issues related to managing large datasets. We present a systematic review of datasets published at the NeurIPS Datasets and Benchmarks track, focusing on four key aspects: provenance, distribution, ethical disclosure, and licensing. Our findings reveal that dataset provenance is often unclear due to ambiguous filtering and curation processes. Additionally, a variety of sites are used for dataset hosting, but only a few offer structured metadata and version control. These inconsistencies underscore the urgent need for standardized data infrastructures for the publication and management of datasets.

In current deep learning tasks, Adam-style optimizers—such as Adam, Adagrad, RMSprop, Adafactor, and Lion—have been widely used as alternatives to SGD-style optimizers. These optimizers typically update model parameters using the sign of gradients, resulting in more stable convergence curves. The learning rate and the batch size are the most critical hyperparameters for optimizers, which require careful tuning to enable effective convergence. Previous research has shown that the optimal learning rate increases linearly (or follows similar rules) with batch size for SGD-style optimizers. However, this conclusion is not applicable to Adam-style optimizers. In this paper, we elucidate the connection between optimal learning rates and batch sizes for Adam-style optimizers through both theoretical analysis and extensive experiments. First, we raise the scaling law between batch sizes and optimal learning rates in the “sign of gradient” case, in which we prove that the optimal learning rate first rises and then falls as the batch size increases. Moreover, the peak value of the surge will gradually move toward the larger batch size as training progresses. Second, we conduct experiments on various CV and NLP tasks and verify the correctness of the scaling law.

We study the problem of dynamic regret minimization in online convex optimization, in which the objective is to minimize the difference between the cumulative loss of an algorithm and that of an arbitrary sequence of comparators. While the literature on this topic is very rich, a unifying framework for the analysis and design of these algorithms is still missing. In this paper we show that *for linear losses, dynamic regret minimization is equivalent to static regret minimization in an extended decision space*. Using this simple observation, we show that there is a frontier of lower bounds trading off penalties due to the variance of the losses and penalties due to variability of the comparator sequence, and provide a framework for achieving any of the guarantees along this frontier. As a result, we also prove for the first time that adapting to the squared path-length of an arbitrary sequence of comparators to achieve regret $R_{T}(u_{1},\dots,u_{T})\le O(\sqrt{T\sum_{t} \\|u_{t}-u_{t+1}\\|^{2}})$ is impossible. However, using our framework we introduce an alternative notion of variability based on a locally-smoothed comparator sequence $\bar u_{1}, \dots, \bar u_{T}$, and provide an algorithm guaranteeing dynamic regret of the form $R_{T}(u_{1},\dots,u_{T})\le \tilde O(\sqrt{T\sum_{i}\\|\bar u_{i}-\bar u_{i+1}\\|^{2}})$, while still matching in the worst case the usual path-length dependencies up to polylogarithmic terms.

Transformers and linear state space models can be evaluated in parallel on modern hardware, but evaluating nonlinear RNNs appears to be an inherently sequential problem. Recently, however, Lim et al. '24 developed an approach called DEER, which evaluates nonlinear RNNs in parallel by posing the states as the solution to a fixed-point problem. They derived a parallel form of Newton's method to solve the fixed-point problem and achieved significant speedups over sequential evaluation. However, the computational complexity of DEER is cubic in the state size, and the algorithm can suffer from numerical instability. We address these limitations with two novel contributions. To reduce the computational complexity, we apply quasi-Newton approximations and show they converge comparably to Newton, use less memory, and are faster. To stabilize DEER, we leverage a connection between the Levenberg-Marquardt algorithm and Kalman smoothing, which we call ELK. This connection allows us to stabilize Newton's method while using efficient parallelized Kalman smoothing algorithms to retain performance. Through several experiments, we show that these innovations allow for parallel evaluation of nonlinear RNNs at larger scales and with greater stability.

As a popular paradigm for juggling data privacy and collaborative training, federated learning (FL) is flourishing to distributively process the large scale of heterogeneous datasets on edged clients. Due to bandwidth limitations and security considerations, it ingeniously splits the original problem into multiple subproblems to be solved in parallel, which empowers primal dual solutions to great application values in FL. In this paper, we review the recent development of classical federated primal dual methods and point out a serious common defect of such methods in non-convex scenarios, which we say is a ``dual drift'' caused by dual hysteresis of those longstanding inactive clients under partial participation training. To further address this problem, we propose a novel Aligned Federated Primal Dual (A-FedPD) method, which constructs virtual dual updates to align global consensus and local dual variables for those protracted unparticipated local clients. Meanwhile, we provide a comprehensive analysis of the optimization and generalization efficiency for the A-FedPD method on smooth non-convex objectives, which confirms its high efficiency and practicality. Extensive experiments are conducted on several classical FL setups to validate the effectiveness of our proposed method.

Large Language Models (LLMs) have exhibited an impressive ability to perform In-Context Learning (ICL) from only a few examples. Recent works have indicated that the functions learned by ICL can be represented through compressed vectors derived from the transformer. However, the working mechanisms and optimization of these vectors are yet to be thoroughly explored. In this paper, we address this gap by presenting a comprehensive analysis of these compressed vectors, drawing parallels to the parameters trained with gradient descent, and introducing the concept of state vector. Inspired by the works on model soup and momentum-based gradient descent, we propose inner and momentum optimization methods that are applied to refine the state vector progressively as test-time adaptation. Moreover, we simulate state vector aggregation in the multiple example setting, where demonstrations comprising numerous examples are usually too lengthy for regular ICL, and further propose a divide-and-conquer aggregation method to address this challenge. We conduct extensive experiments using Llama-2 and GPT-J in both zero-shot setting and few-shot setting. The experimental results show that our optimization method effectively enhances the state vector and achieves the state-of-the-art performance on diverse tasks.

One of the key challenges in machine learning is to find interpretable representations of learned functions. The Möbius transform is essential for this purpose, as its coefficients correspond to unique *importance scores* for *sets of input variables*. This transform is closely related to widely used game-theoretic notions of importance like the *Shapley* and *Bhanzaf value*, but it also captures crucial higher-order interactions. Although computing the Möbius Transform of a function with $n$ inputs involves $2^n$ coefficients, it becomes tractable when the function is *sparse* and of *low-degree* as we show is the case for many real-world functions. Under these conditions, the complexity of the transform computation is significantly reduced. When there are $K$ non-zero coefficients, our algorithm recovers the Möbius transform in $O(Kn)$ samples and $O(Kn^2)$ time asymptotically under certain assumptions, the first non-adaptive algorithm to do so. We also uncover a surprising connection between group testing and the Möbius transform. For functions where all interactions involve at most $t$ inputs, we use group testing results to compute the Möbius transform with $O(Kt\log n)$ sample complexity and $O(K\mathrm{poly}(n))$ time. A robust version of this algorithm withstands noise and maintains this complexity. This marks the first $n$ sub-linear query complexity, noise-tolerant algorithm for the Möbius transform. While our algorithms are conceptualized in an idealized setting, they indicate that the Möbius transform is a potent tool for interpreting deep learning models.

Exploring the integration of if-then logic rules within neural network architectures presents an intriguing area. This integration seamlessly transforms the rule learning task into neural network training using backpropagation and stochastic gradient descent. From a well-trained sparse and shallow neural network, one can interpret each layer and neuron through the language of logic rules, and a global explanatory rule set can be directly extracted. However, ensuring interpretability may impose constraints on the flexibility, depth, and width of neural networks. In this paper, we propose HyperLogic: a novel framework leveraging hypernetworks to generate weights of the main network. HyperLogic can unveil multiple diverse rule sets, each capable of capturing heterogeneous patterns in data. This provides a simple yet effective method to increase model flexibility and preserve interpretability. We theoretically analyzed the benefits of the HyperLogic by examining the approximation error and generalization capabilities under two types of regularization terms: sparsity and diversity regularizations. Experiments on real data demonstrate that our method can learn more diverse, accurate, and concise rules.

Feature attribution methods explain black-box machine learning (ML) models by assigning importance scores to input features. These methods can be computationally expensive for large ML models. To address this challenge, there have been increasing efforts to develop amortized explainers, where a ML model is trained to efficiently approximate computationally expensive feature attribution scores. Despite their efficiency, amortized explainers can produce misleading explanations. In this paper, we propose selective explanations to (i) detect when amortized explainers generate inaccurate explanations and (ii) improve the approximation of the explanation using a technique we call explanations with initial guess. Selective explanations allow practitioners to specify the fraction of samples that receive explanations with initial guess, offering a principled way to bridge the gap between amortized explainers (one inference) and more computationally costly approximations (multiple inferences). Our experiments on various models and datasets demonstrate that feature attributions via selective explanations strike a favorable balance between explanation quality and computational efficiency.

Parameter-efficient finetuning (PEFT) methods seek to adapt large neural models via updates to a small number of *weights*. However, much prior interpretability work has shown that *representations* encode rich semantic information, suggesting that editing representations might be a more powerful alternative. We pursue this hypothesis by developing a family of **Representation Finetuning (ReFT)** methods. ReFT methods operate on a frozen base model and learn task-specific interventions on hidden representations. We define a strong instance of the ReFT family, Low-rank Linear Subspace ReFT (LoReFT), and we identify an ablation of this method that trades some performance for increased efficiency. Both are drop-in replacements for existing PEFTs and learn interventions that are 15x--65x more parameter-efficient than LoRA. We showcase LoReFT on eight commonsense reasoning tasks, four arithmetic reasoning tasks, instruction-tuning, and GLUE. In all these evaluations, our ReFTs deliver the best balance of efficiency and performance, and almost always outperform state-of-the-art PEFTs. Upon publication, we will publicly release our generic ReFT training library.

Today's best language models still struggle with "hallucinations", factually incorrect generations, which impede their ability to reliably retrieve information seen during training. The *reversal curse*, where models cannot recall information when probed in a different order than was encountered during training, exemplifies limitations in information retrieval. To better understand these limitations, we reframe the reversal curse as a *factorization curse* --- a failure of models to learn the same joint distribution under different factorizations. We more closely simulate finetuning workflows which train pretrained models on specialized knowledge by introducing *WikiReversal*, a realistic testbed based on Wikipedia knowledge graphs. Through a series of controlled experiments with increasing levels of realism, including non-reciprocal relations, we find that reliable information retrieval is an inherent failure of the next-token prediction objective used in popular large language models. Moreover, we demonstrate reliable information retrieval cannot be solved with scale, reversed tokens, or even naive bidirectional-attention training. Consequently, various approaches to finetuning on specialized data would necessarily provide mixed results on downstream tasks, unless the model has already seen the right sequence of tokens. Across five tasks of varying levels of complexity, our results uncover a promising path forward: factorization-agnostic objectives can significantly mitigate the reversal curse and hint at improved knowledge storage and planning capabilities.

Graph regression is a fundamental task that has gained significant attention in various graph learning tasks. However, the inference process is often not easily interpretable. Current explanation techniques are limited to understanding Graph Neural Network (GNN) behaviors in classification tasks, leaving an explanation gap for graph regression models. In this work, we propose a novel explanation method to interpret the graph regression models (XAIG-R). Our method addresses the distribution shifting problem and continuously ordered decision boundary issues that hinder existing methods away from being applied in regression tasks. We introduce a novel objective based on the graph information bottleneck theory (GIB) and a new mix-up framework, which can support various GNNs and explainers in a model-agnostic manner. Additionally, we present a self-supervised learning strategy to tackle the continuously ordered labels in regression tasks. We evaluate our proposed method on three benchmark datasets and a real-life dataset introduced by us, and extensive experiments demonstrate its effectiveness in interpreting GNN models in regression tasks.

Parametric dimensionality reduction methods have gained prominence for their ability to generalize to unseen datasets, an advantage that traditional non-parametric approaches typically lack. Despite their growing popularity, there remains a prevalent misconception among practitioners about the equivalence in performance between parametric and non-parametric methods. Here, we show that these methods are not equivalent -- parametric methods retain global structure but lose significant local details. To explain this, we provide evidence that parameterized approaches lack the ability to repulse negative samples, and the choice of loss function also has an impact. Addressing these issues, we developed a new parametric method, ParamRepulsor, that incorporates Hard Negative Mining and a loss function that applies a strong repulsive force. This new method achieves state-of-the-art performance on local structure preservation for parametric methods without sacrificing the fidelity of global structural representation. Our code is available at https://github.com/hyhuang00/ParamRepulsor.