NeurIPS 2024 论文列表

Studies show that safety-tuned models may nevertheless divulge harmful information. In this work, we show that whether they do so depends significantly on who they are talking to, which we refer to as *user persona*. In fact, we find manipulating user persona to be more effective for eliciting harmful content than certain more direct attempts to control model refusal. We study both natural language prompting and activation steering as intervention methods and show that activation steering is significantly more effective at bypassing safety filters. We shed light on the mechanics of this phenomenon by showing that even when model generations are safe, harmful content can persist in hidden representations and can be extracted by decoding from earlier layers. We also show we can predict a persona’s effect on refusal given only the geometry of its steering vector. Finally, we show that certain user personas induce the model to form more charitable interpretations of otherwise dangerous queries.

An important line of research in the field of explainability is to extract a small subset of crucial rationales from the full input. The most widely used criterion for rationale extraction is the maximum mutual information (MMI) criterion. However, in certain datasets, there are spurious features non-causally correlated with the label and also get high mutual information, complicating the loss landscape of MMI. Although some penalty-based methods have been developed to penalize the spurious features (e.g., invariance penalty, intervention penalty, etc) to help MMI work better, these are merely remedial measures. In the optimization objectives of these methods, spurious features are still distinguished from plain noise, which hinders the discovery of causal rationales. This paper aims to develop a new criterion that treats spurious features as plain noise, allowing the model to work on datasets rich in spurious features as if it were working on clean datasets, thereby making rationale extraction easier. We theoretically observe that removing either plain noise or spurious features from the input does not alter the conditional distribution of the remaining components relative to the task label. However, significant changes in the conditional distribution occur only when causal features are eliminated. Based on this discovery, the paper proposes a criterion for \textbf{M}aximizing the \textbf{R}emaining \textbf{D}iscrepancy (MRD). Experiments on six widely used datasets show that our MRD criterion improves rationale quality (measured by the overlap with human-annotated rationales) by up to $10.4\%$ as compared to several recent competitive MMI variants. Code: \url{https://github.com/jugechengzi/Rationalization-MRD}.

Transformer-based large language models (LLMs) have successfully handled various tasks. As one fundamental module in Transformers, position encoding encodes the positional information of tokens in a sequence. Specifically, rotary position embedding (RoPE), one of the most widely used techniques, encodes the positional information by dividing the query or key value with $d$ elements into $d/2$ pairs and rotating the 2d vectors corresponding to each pair of elements. Therefore, the direction of each pair and the position-related rotation jointly determine the attention score. In this paper, we show that the direction of the 2d pair is largely affected by the angle between the corresponding weight vector pair. We theoretically show that non-orthogonal weight vector pairs lead to great attention on tokens at a certain relative position and are less sensitive to the input which may correspond to basic syntactic information. Meanwhile, the orthogonal weight vector pairs are more flexible regarding the relative position, which may correspond to high-level syntactic information. Empirical evidence supports the hypothesis that shallow layers of LLMs focus more on local syntax and deep layers focus more on high-level semantics. Furthermore, we show that LLMs fine-tuning mainly changes the pairs of weight vectors that are nearly orthogonal, i.e., the weight corresponding to high-level semantics, which enables the reduction of the number of trainable parameters during fine-tuning without sacrificing performance. We propose a method namely Angle-based Weight Selection (AWS) to reduce the fine-tuning overhead and verify the effectiveness of the proposed method on widely used Alpaca fine-tuned Llama-2.

We introduce QuaRot, a new Quantization scheme based on Rotations, which is able to quantize LLMs end-to-end, including all weights, activations, and KV cache in 4 bits. QuaRot rotates LLMs in a way that removes outliers from the hidden state without changing the output, making quantization easier. This computational invariance is applied to the hidden state (residual) of the LLM, as well as to the activations of the feed-forward components, aspects of the attention mechanism, and to the KV cache. The result is a quantized model where all matrix multiplications are performed in 4 bits, without any channels identified for retention in higher precision. Our 4-bit quantized LLAMA2-70B model has losses of at most 0.47 WikiText-2 perplexity and retains 99% of the zero-shot performance. We also show that QuaRot can provide lossless 6 and 8 bit LLAMA-2 models without any calibration data using round-to-nearest quantization. Code is available at github.com/spcl/QuaRot.

Self-attention in vision transformers is often thought to perform perceptual grouping where tokens attend to other tokens with similar embeddings, which could correspond to semantically similar features of an object. However, attending to dissimilar tokens can be beneficial by providing contextual information. We propose to analyze the query-key interaction by the singular value decomposition of the interaction matrix (i.e. ${\textbf{W}_q}^\top\textbf{W}_k$). We find that in many ViTs, especially those with classification training objectives, early layers attend more to similar tokens, while late layers show increased attention to dissimilar tokens, providing evidence corresponding to perceptual grouping and contextualization, respectively. Many of these interactions between features represented by singular vectors are interpretable and semantic, such as attention between relevant objects, between parts of an object, or between the foreground and background. This offers a novel perspective on interpreting the attention mechanism, which contributes to understanding how transformer models utilize context and salient features when processing images.

The reliable recovery and uncertainty quantification of a fixed effect function $\mu$ in a functional mixed model, for modeling population- and object-level variability in noisily observed functional data, is a notoriously challenging task: variations along the $x$ and $y$ axes are confounded with additive measurement error, and cannot in general be disentangled. The question then as to what properties of $\mu$ may be reliably recovered becomes important. We demonstrate that it is possible to recover the size-and-shape of a square-integrable $\mu$ under a Bayesian functional mixed model. The size-and-shape of $\mu$ is a geometric property invariant to a family of space-time unitary transformations, viewed as rotations of the Hilbert space, that jointly transform the $x$ and $y$ axes. A random object-level unitary transformation then captures size-and-shape preserving deviations of $\mu$ from an individual function, while a random linear term and measurement error capture size-and-shape altering deviations. The model is regularized by appropriate priors on the unitary transformations, posterior summaries of which may then be suitably interpreted as optimal data-driven rotations of a fixed orthonormal basis for the Hilbert space. Our numerical experiments demonstrate utility of the proposed model, and superiority over the current state-of-the-art.

Over half of cancer patients experience long-term pain management challenges. Recently, interest has grown in systems for cancer pain treatment effectiveness assessment (TEA) and medication recommendation (MR) to optimize pharmacological care. These systems aim to improve treatment effectiveness by recommending personalized medication plans based on comprehensive patient information. Despite progress, current systems lack multidisciplinary treatment (MDT) team assessments of treatment and the patient's perception of medication, crucial for effective cancer pain management. Moreover, managing cancer pain medication requires multiple adjustments to the treatment plan based on the patient's evolving condition, a detail often missing in existing datasets. To tackle these issues, we designed the PEACE dataset specifically for cancer pain medication research. It includes detailed pharmacological care records for over 38,000 patients, covering demographics, clinical examination, treatment outcomes, medication plans, and patient self-perceptions. Unlike existing datasets, PEACE records not only long-term and multiple follow-ups both inside and outside hospitals but also includes patients' self-assessments of medication effects and the impact on their lives. We conducted a proof-of-concept study with 13 machine learning algorithms on the PEACE dataset for the TEA (classification task) and MR (regression task). These experiments provide valuable insights into the potential of the PEACE dataset for advancing personalized cancer pain management. The dataset is accessible at: [https://github.com/YTYTYD/PEACE].

Pre-trained large language models based on Transformers have demonstrated remarkable in-context learning (ICL) abilities. With just a few demonstration examples, the models can implement new tasks without any parameter updates. However, it is still an open question to understand the mechanism of ICL. In this paper, we attempt to explore the ICL process in Transformers through a lens of representation learning. Initially, leveraging kernel methods, we figure out a dual model for one softmax attention layer. The ICL inference process of the attention layer aligns with the training procedure of its dual model, generating token representation predictions that are equivalent to the dual model's test outputs. We delve into the training process of this dual model from a representation learning standpoint and further derive a generalization error bound related to the quantity of demonstration tokens. Subsequently, we extend our theoretical conclusions to more complicated scenarios, including one Transformer layer and multiple attention layers. Furthermore, drawing inspiration from existing representation learning methods especially contrastive learning, we propose potential modifications for the attention layer. Finally, experiments are designed to support our findings.

The inner workings of neural networks can be better understood if we can fully decipher the information encoded in neural activations. In this paper, we argue that this information is embodied by the subset of inputs that give rise to similar activations. We propose InversionView, which allows us to practically inspect this subset by sampling from a trained decoder model conditioned on activations. This helps uncover the information content of activation vectors, and facilitates understanding of the algorithms implemented by transformer models. We present four case studies where we investigate models ranging from small transformers to GPT-2. In these studies, we show that InversionView can reveal clear information contained in activations, including basic information about tokens appearing in the context, as well as more complex information, such as the count of certain tokens, their relative positions, and abstract knowledge about the subject. We also provide causally verified circuits to confirm the decoded information.

Due to its powerful representational capabilities, Transformers have gradually become the mainstream model in the field of machine vision. However, the vast and complex parameters of Transformers impede researchers from gaining a deep understanding of their internal mechanisms, especially error mechanisms. Existing methods for interpreting Transformers mainly focus on understanding them from the perspectives of the importance of input tokens or internal modules, as well as the formation and meaning of features. In contrast, inspired by research on information integration mechanisms and conjunctive errors in the biological visual system, this paper conducts an in-depth exploration of the internal error mechanisms of Transformers. We first propose an information integration hypothesis for Transformers in the machine vision domain and provide substantial experimental evidence to support this hypothesis. This includes the dynamic integration of information among tokens and the static integration of information within tokens in Transformers, as well as the presence of conjunctive errors therein. Addressing these errors, we further propose heuristic dynamic integration constraint methods and rule-based static integration constraint methods to rectify errors and ultimately improve model performance. The entire methodology framework is termed as Transformer Doctor, designed for diagnosing and treating internal errors within transformers. Through a plethora of quantitative and qualitative experiments, it has been demonstrated that Transformer Doctor can effectively address internal errors in transformers, thereby enhancing model performance.

Distributional regression aims at estimating the conditional distribution of a target variable given explanatory co-variates. It is a crucial tool for forecasting when a precise uncertainty quantification is required. A popular methodology consists in fitting a parametric model via empirical risk minimization where the risk is measured by the Continuous Rank Probability Score (CRPS). For independent and identically distributed observations, we provide a concentration result for the estimation error and an upper bound for its expectation. Furthermore, we consider model selection performed by minimization of the validation error and provide a concentration bound for the regret. A similar result is proved for convex aggregation of models. Finally, we show that our results may be applied to various models such as EMOS, distributional regression networks, distributional nearest neighbours or distributional random forests and we illustrate our findings on two data sets (QSAR aquatic toxicity and Airfoil self-noise).

Deep neural networks (DNNs) are widely used models for investigating biological visual representations. However, existing DNNs are mostly designed to analyze neural responses to static images, relying on feedforward structures and lacking physiological neuronal mechanisms. There is limited insight into how the visual cortex represents natural movie stimuli that contain context-rich information. To address these problems, this work proposes the long-range feedback spiking network (LoRaFB-SNet), which mimics top-down connections between cortical regions and incorporates spike information processing mechanisms inherent to biological neurons. Taking into account the temporal dependence of representations under movie stimuli, we present Time-Series Representational Similarity Analysis (TSRSA) to measure the similarity between model representations and visual cortical representations of mice. LoRaFB-SNet exhibits the highest level of representational similarity, outperforming other well-known and leading alternatives across various experimental paradigms, especially when representing long movie stimuli. We further conduct experiments to quantify how temporal structures (dynamic information) and static textures (static information) of the movie stimuli influence representational similarity, suggesting that our model benefits from long-range feedback to encode context-dependent representations just like the brain. Altogether, LoRaFB-SNet is highly competent in capturing both dynamic and static representations of the mouse visual cortex and contributes to the understanding of movie processing mechanisms of the visual system. Our codes are available at https://github.com/Grasshlw/SNN-Neural-Similarity-Movie.

The explainability of deep neural networks (DNNs) is critical for trust and reliability in AI systems. Path-based attribution methods, such as integrated gradients (IG), aim to explain predictions by accumulating gradients along a path from a baseline to the target image. However, noise accumulated during this process can significantly distort the explanation. While existing methods primarily concentrate on finding alternative paths to circumvent noise, they overlook a critical issue: intermediate-step images frequently diverge from the distribution of training data, further intensifying the impact of noise. This work presents a novel Denoising Diffusion Path (DDPath) to tackle this challenge by harnessing the power of diffusionmodels for denoising. By exploiting the inherent ability of diffusion models to progressively remove noise from an image, DDPath constructs a piece-wise linear path. Each segment of this path ensures that samples drawn from a Gaussian distribution are centered around the target image. This approach facilitates a gradual reduction of noise along the path. We further demonstrate that DDPath adheres to essential axiomatic properties for attribution methods and can be seamlessly integrated with existing methods such as IG. Extensive experimental results demonstrate that DDPath can significantly reduce noise in the attributions—resulting in clearer explanations—and achieves better quantitative results than traditional path-based methods.

The Mixture of Experts (MoE) paradigm provides a powerful way to decompose dense layers into smaller, modular computations often more amenable to human interpretation, debugging, and editability. However, a major challenge lies in the computational cost of scaling the number of experts high enough to achieve fine-grained specialization. In this paper, we propose the Multilinear Mixture of Experts (μMoE) layer to address this, focusing on vision models. μMoE layers enable scalable expert specialization by performing an implicit computation on prohibitively large weight tensors entirely in factorized form. Consequently, μMoEs (1) avoid the restrictively high inference-time costs of dense MoEs, yet (2) do not inherit the training issues of the popular sparse MoEs' discrete (non-differentiable) expert routing. We present both qualitative and quantitative evidence that scaling μMoE layers when fine-tuning foundation models for vision tasks leads to more specialized experts at the class-level, further enabling manual bias correction in CelebA attribute classification. Finally, we show qualitative results demonstrating the expert specialism achieved when pre-training large GPT2 and MLP-Mixer models with parameter-matched μMoE blocks at every layer, maintaining comparable accuracy. Our code is available at: https://github.com/james-oldfield/muMoE.

Asynchronous stochastic gradient descent (ASGD) has evolved into an indispensable optimization algorithm for training modern large-scale distributed machine learning tasks. Therefore, it is imperative to explore the generalization performance of the ASGD algorithm. However, the existing results are either pessimistic and vacuous or restricted by strict assumptions that fail to reveal the intrinsic impact of asynchronous training on generalization. In this study, we establish sharper stability and generalization bounds for ASGD under much weaker assumptions. Firstly, this paper studies the on-average model stability of ASGD and provides a non-vacuous upper bound on the generalization error, without relying on the Lipschitz assumption. Furthermore, we investigate the excess generalization error of the ASGD algorithm, revealing the effects of asynchronous delay, model initialization, number of training samples and iterations on generalization performance. Secondly, for the first time, this study explores the generalization performance of ASGD in the non-smooth case. We replace smoothness with the much weaker Hölder continuous assumption and achieve similar generalization results as in the smooth case. Finally, we validate our theoretical findings by training numerous machine learning models, including convex problems and non-convex tasks in computer vision and natural language processing.

Continual learning of partially similar tasks poses a challenge for artificial neural networks, as task similarity presents both an opportunity for knowledge transfer and a risk of interference and catastrophic forgetting. However, it remains unclear how task similarity in input features and readout patterns influences knowledge transfer and forgetting, as well as how they interact with common algorithms for continual learning. Here, we develop a linear teacher-student model with latent structure and show analytically that high input feature similarity coupled with low readout similarity is catastrophic for both knowledge transfer and retention. Conversely, the opposite scenario is relatively benign. Our analysis further reveals that task-dependent activity gating improves knowledge retention at the expense of transfer, while task-dependent plasticity gating does not affect either retention or transfer performance at the over-parameterized limit. In contrast, weight regularization based on the Fisher information metric significantly improves retention, regardless of task similarity, without compromising transfer performance. Nevertheless, its diagonal approximation and regularization in the Euclidean space are much less robust against task similarity. We demonstrate consistent results in a permuted MNIST task with latent variables. Overall, this work provides insights into when continual learning is difficult and how to mitigate it.

Hands are the primary means through which humans interact with the world. Reliable and always-available hand pose inference could yield new and intuitive control schemes for human-computer interactions, particularly in virtual and augmented reality. Computer vision is effective but requires one or multiple cameras and can struggle with occlusions, limited field of view, and poor lighting. Wearable wrist-based surface electromyography (sEMG) presents a promising alternative as an always-available modality sensing muscle activities that drive hand motion. However, sEMG signals are strongly dependent on user anatomy and sensor placement; existing sEMG models have thus required hundreds of users and device placements to effectively generalize for tasks other than pose inference. To facilitate progress on sEMG pose inference, we introduce the emg2pose benchmark, which is to our knowledge the first publicly available dataset of high-quality hand pose labels and wrist sEMG recordings. emg2pose contains 2kHz, 16 channel sEMG and pose labels from a 26-camera motion capture rig for 193 users, 370 hours, and 29 stages with diverse gestures - a scale comparable to vision-based hand pose datasets. We provide competitive baselines and challenging tasks evaluating real-world generalization scenarios: held-out users, sensor placements, and stages. This benchmark provides the machine learning community a platform for exploring complex generalization problems, holding potential to significantly enhance the development of sEMG-based human-computer interactions.

Noise in data significantly influences decision-making in the data science process. In fact, it has been shown that noise in data generation processes leads practitioners to find simpler models. However, an open question still remains: what is the degree of model simplification we can expect under different noise levels? In this work, we address this question by investigating the relationship between the amount of noise and model simplicity across various hypothesis spaces, focusing on decision trees and linear models. We formally show that noise acts as an implicit regularizer for several different noise models. Furthermore, we prove that Rashomon sets (sets of near-optimal models) constructed with noisy data tend to contain simpler models than corresponding Rashomon sets with non-noisy data. Additionally, we show that noise expands the set of ``good'' features and consequently enlarges the set of models that use at least one good feature. Our work offers theoretical guarantees and practical insights for practitioners and policymakers on whether simple-yet-accurate machine learning models are likely to exist, based on knowledge of noise levels in the data generation process.

We study fairness in social influence maximization, whereby one seeks to select seeds that spread a given information throughout a network, ensuring balanced outreach among different communities (e.g. demographic groups). In the literature, fairness is often quantified in terms of the expected outreach within individual communities. In this paper, we demonstrate that such fairness metrics can be misleading since they overlook the stochastic nature of information diffusion processes. When information diffusion occurs in a probabilistic manner, multiple outreach scenarios can occur. As such, outcomes such as “In 50% of the cases, no one in group 1 gets the information, while everyone in group 2 does, and in the other 50%, it is the opposite”, which always results in largely unfair outcomes, are classified as fair by a variety of fairness metrics in the literature. We tackle this problem by designing a new fairness metric, mutual fairness, that captures variability in outreach through optimal transport theory. We propose a new seed- selection algorithm that optimizes both outreach and mutual fairness, and we show its efficacy on several real datasets. We find that our algorithm increases fairness with only a minor decrease (and at times, even an increase) in efficiency.