Zesheng Ye

Yao Liu, Zesheng Ye, Rui Wang et al.

Tremendous efforts have been put forth on predicting pedestrian trajectory with generative models to accommodate uncertainty and multi-modality in human behaviors. An individual's inherent uncertainty, e.g., change of destination, can be masked by complex patterns resulting from the movements of interacting pedestrians. However, latent variable-based generative models often entangle such uncertainty with complexity, leading to limited either latent expressivity or predictive diversity. In this work, we propose to separately model these two factors by implicitly deriving a flexible latent representation to capture intricate pedestrian movements, while integrating predictive uncertainty of individuals with explicit bivariate Gaussian mixture densities over their future locations. More specifically, we present a model-agnostic uncertainty-aware pedestrian trajectory prediction framework, parameterizing sufficient statistics for the mixture of Gaussians that jointly comprise the multi-modal trajectories. We further estimate these parameters of interest by approximating a denoising process that progressively recovers pedestrian movements from noise. Unlike previous studies, we translate the predictive stochasticity to explicit distributions, allowing it to readily generate plausible future trajectories indicating individuals' self-uncertainty. Moreover, our framework is compatible with different neural net architectures. We empirically show the performance gains over state-of-the-art even with lighter backbones, across most scenes on two public benchmarks.

Jing Du, Zesheng Ye, Lina Yao et al.

Recent sequential recommendation models rely increasingly on consecutive short-term user-item interaction sequences to model user interests. These approaches have, however, raised concerns about both short- and long-term interests. (1) {\it short-term}: interaction sequences may not result from a monolithic interest, but rather from several intertwined interests, even within a short period of time, resulting in their failures to model skip behaviors; (2) {\it long-term}: interaction sequences are primarily observed sparsely at discrete intervals, other than consecutively over the long run. This renders difficulty in inferring long-term interests, since only discrete interest representations can be derived, without taking into account interest dynamics across sequences. In this study, we address these concerns by learning (1) multi-scale representations of short-term interests; and (2) dynamics-aware representations of long-term interests. To this end, we present an \textbf{I}nterest \textbf{D}ynamics modeling framework using generative \textbf{N}eural \textbf{P}rocesses, coined IDNP, to model user interests from a functional perspective. IDNP learns a global interest function family to define each user's long-term interest as a function instantiation, manifesting interest dynamics through function continuity. Specifically, IDNP first encodes each user's short-term interactions into multi-scale representations, which are then summarized as user context. By combining latent global interest with user context, IDNP then reconstructs long-term user interest functions and predicts interactions at upcoming query timestep. Moreover, IDNP can model such interest functions even when interaction sequences are limited and non-consecutive. Extensive experiments on four real-world datasets demonstrate that our model outperforms state-of-the-arts on various evaluation metrics.

Zesheng Ye, Jing Du, Lina Yao

Conditional Neural Processes~(CNPs) formulate distributions over functions and generate function observations with exact conditional likelihoods. CNPs, however, have limited expressivity for high-dimensional observations, since their predictive distribution is factorized into a product of unconstrained (typically) Gaussian outputs. Previously, this could be handled using latent variables or autoregressive likelihood, but at the expense of intractable training and quadratically increased complexity. Instead, we propose calibrating CNPs with an adversarial training scheme besides regular maximum likelihood estimates. Specifically, we train an energy-based model (EBM) with noise contrastive estimation, which enforces EBM to identify true observations from the generations of CNP. In this way, CNP must generate predictions closer to the ground-truth to fool EBM, instead of merely optimizing with respect to the fixed-form likelihood. From generative function reconstruction to downstream regression and classification tasks, we demonstrate that our method fits mainstream CNP members, showing effectiveness when unconstrained Gaussian likelihood is defined, requiring minimal computation overhead while preserving foundation properties of CNPs.

Zesheng Ye, Lina Yao

Conditional Neural Processes~(CNPs) bridge neural networks with probabilistic inference to approximate functions of Stochastic Processes under meta-learning settings. Given a batch of non-{\it i.i.d} function instantiations, CNPs are jointly optimized for in-instantiation observation prediction and cross-instantiation meta-representation adaptation within a generative reconstruction pipeline. There can be a challenge in tying together such two targets when the distribution of function observations scales to high-dimensional and noisy spaces. Instead, noise contrastive estimation might be able to provide more robust representations by learning distributional matching objectives to combat such inherent limitation of generative models. In light of this, we propose to equip CNPs by 1) aligning prediction with encoded ground-truth observation, and 2) decoupling meta-representation adaptation from generative reconstruction. Specifically, two auxiliary contrastive branches are set up hierarchically, namely in-instantiation temporal contrastive learning~({\tt TCL}) and cross-instantiation function contrastive learning~({\tt FCL}), to facilitate local predictive alignment and global function consistency, respectively. We empirically show that {\tt TCL} captures high-level abstraction of observations, whereas {\tt FCL} helps identify underlying functions, which in turn provides more efficient representations. Our model outperforms other CNPs variants when evaluating function distribution reconstruction and parameter identification across 1D, 2D and high-dimensional time-series.

Zesheng Ye, Lina Yao, Yu Zhang et al.

Recent studies demonstrate the use of a two-stage supervised framework to generate images that depict human perception to visual stimuli from EEG, referring to EEG-visual reconstruction. They are, however, unable to reproduce the exact visual stimulus, since it is the human-specified annotation of images, not their data, that determines what the synthesized images are. Moreover, synthesized images often suffer from noisy EEG encodings and unstable training of generative models, making them hard to recognize. Instead, we present a single-stage EEG-visual retrieval paradigm where data of two modalities are correlated, as opposed to their annotations, allowing us to recover the exact visual stimulus for an EEG clip. We maximize the mutual information between the EEG encoding and associated visual stimulus through optimization of a contrastive self-supervised objective, leading to two additional benefits. One, it enables EEG encodings to handle visual classes beyond seen ones during training, since learning is not directed at class annotations. In addition, the model is no longer required to generate every detail of the visual stimulus, but rather focuses on cross-modal alignment and retrieves images at the instance level, ensuring distinguishable model output. Empirical studies are conducted on the largest single-subject EEG dataset that measures brain activities evoked by image stimuli. We demonstrate the proposed approach completes an instance-level EEG-visual retrieval task which existing methods cannot. We also examine the implications of a range of EEG and visual encoder structures. Furthermore, for a mostly studied semantic-level EEG-visual classification task, despite not using class annotations, the proposed method outperforms state-of-the-art supervised EEG-visual reconstruction approaches, particularly on the capability of open class recognition.

Chengyi Cai, Zesheng Ye, Jiangchao Yao et al.

In large language model (LLM) unlearning, private information is required to be removed. Task arithmetic unlearns by subtracting a specific task vector (TV)--defined as the parameter difference between a privacy-information-tuned model and the original model. While efficient, it can cause over-forgetting by disrupting parameters essential for retaining other information. Motivated by the observation that each parameter exhibits different importance for forgetting versus retention, we propose a per-parameter task arithmetic (PerTA) mechanism to rescale the TV, allowing per-parameter adjustment. These weights quantify the relative importance of each parameter for forgetting versus retention, estimated via gradients (i.e., PerTA-grad) or the diagonal Fisher information approximation (i.e., PerTA-fisher). Moreover, we discuss the effectiveness of PerTA, extend it to a more general form, and provide further analysis. Extensive experiments demonstrate that PerTA consistently improves upon standard TV, and in many cases surpasses widely used training-based unlearning methods in both forgetting effectiveness and overall model utility. By retaining the efficiency of task arithmetic while mitigating over-forgetting, PerTA offers a principled and practical framework for LLM unlearning.

Chengyi Cai, Zesheng Ye, Lei Feng et al.

Visual reprogramming (VR) leverages the intrinsic capabilities of pretrained vision models by adapting their input or output interfaces to solve downstream tasks whose labels (i.e., downstream labels) might be totally different from the labels associated with the pretrained models (i.e., pretrained labels). When adapting the output interface, label mapping methods transform the pretrained labels to downstream labels by establishing a gradient-free one-to-one correspondence between the two sets of labels. However, in this paper, we reveal that one-to-one mappings may overlook the complex relationship between pretrained and downstream labels. Motivated by this observation, we propose a Bayesian-guided Label Mapping (BLM) method. BLM constructs an iteratively-updated probabilistic label mapping matrix, with each element quantifying a pairwise relationship between pretrained and downstream labels. The assignment of values to the constructed matrix is guided by Bayesian conditional probability, considering the joint distribution of the downstream labels and the labels predicted by the pretrained model on downstream samples. Experiments conducted on both pretrained vision models (e.g., ResNeXt) and vision-language models (e.g., CLIP) demonstrate the superior performance of BLM over existing label mapping methods. The success of BLM also offers a probabilistic lens through which to understand and analyze the effectiveness of VR. Our code is available at https://github.com/tmlr-group/BayesianLM.

Chengyi Cai, Zesheng Ye, Lei Feng et al.

Visual reprogramming (VR) reuses pre-trained vision models for downstream image classification tasks by adding trainable noise patterns to inputs. When applied to vision-language models (e.g., CLIP), existing VR approaches follow the same pipeline used in vision models (e.g., ResNet, ViT), where ground-truth class labels are inserted into fixed text templates to guide the optimization of VR patterns. This label-based approach, however, overlooks the rich information and diverse attribute-guided textual representations that CLIP can exploit, which may lead to the misclassification of samples. In this paper, we propose Attribute-based Visual Reprogramming (AttrVR) for CLIP, utilizing descriptive attributes (DesAttrs) and distinctive attributes (DistAttrs), which respectively represent common and unique feature descriptions for different classes. Besides, as images of the same class may reflect different attributes after VR, AttrVR iteratively refines patterns using the $k$-nearest DesAttrs and DistAttrs for each image sample, enabling more dynamic and sample-specific optimization. Theoretically, AttrVR is shown to reduce intra-class variance and increase inter-class separation. Empirically, it achieves superior performance in 12 downstream tasks for both ViT-based and ResNet-based CLIP. The success of AttrVR facilitates more effective integration of VR from unimodal vision models into vision-language models. Our code is available at https://github.com/tmlr-group/AttrVR.

Jing Du, Zesheng Ye, Congbo Ma et al.

Multi-modal recommendation (MMR) enriches item representations by introducing item content, e.g., visual and textual descriptions, to improve upon interaction-only recommenders. The success of MMR hinges on aligning these content modalities with user preferences derived from interaction data, yet dominant practices based on disentangling modality-invariant preference-driving signals from modality-specific preference-irrelevant noises are flawed. First, they assume a one-size-fits-all relevance of item content to user preferences for all users, which contradicts the user-conditional fact of preferences. Second, they optimize pairwise contrastive losses separately toward cross-modal alignment, systematically ignoring higher-order dependencies inherent when multiple content modalities jointly influence user choices. In this paper, we introduce GTC, a conditional Generative Total Correlation learning framework. We employ an interaction-guided diffusion model to perform user-aware content feature filtering, preserving only personalized features relevant to each individual user. Furthermore, to capture complete cross-modal dependencies, we optimize a tractable lower bound of the total correlation of item representations across all modalities. Experiments on standard MMR benchmarks show GTC consistently outperforms state-of-the-art, with gains of up to 28.30% in NDCG@5. Ablation studies validate both conditional preference-driven feature filtering and total correlation optimization, confirming the ability of GTC to model user-conditional relationships in MMR tasks. The code is available at: https://github.com/jingdu-cs/GTC.

Chengyi Cai, Zesheng Ye, Lei Feng et al.

Model reprogramming adapts pretrained models to downstream tasks by modifying only the input and output spaces. Visual reprogramming (VR) is one instance for vision tasks that adds a trainable noise pattern (i.e., a visual prompt) to input images to facilitate downstream classification. The existing VR approaches for CLIP train a single visual prompt using all descriptions of different downstream classes. However, the limited learning capacity may result in (1) a failure to capture diverse aspects of the descriptions (e.g., shape, color, and texture), and (2) a possible bias toward less informative attributes that do not help distinguish between classes. In this paper, we introduce a decoupling-and-reweighting framework. Our decoupled visual prompts (DVP) are optimized using descriptions grouped by explicit causes (DVP-cse) or unsupervised clusters (DVP-cls). Then, we integrate the outputs of these visual prompts with a probabilistic reweighting matrix (PRM) that measures their contributions to each downstream class. Theoretically, DVP lowers the empirical risk bound. Experimentally, DVP outperforms baselines on average across 11 downstream datasets. Notably, the DVP-PRM integration enables insights into how individual visual prompts influence classification decisions, providing a probabilistic framework for understanding reprogramming. Our code is available at https://github.com/tmlr-group/DecoupledVP.

Chengyi Cai, Zesheng Ye, Lei Feng et al.

Visual reprogramming (VR) is a prompting technique that aims to re-purpose a pre-trained model (e.g., a classifier on ImageNet) to target tasks (e.g., medical data prediction) by learning a small-scale pattern added into input images instead of tuning considerable parameters within the model. The location of the pattern within input samples is usually determined by a pre-defined mask shared across all samples. In this paper, we show that the shared mask potentially limits VR's generalization and increases its approximation error due to the lack of sample-level adaptation. Motivated by this finding, we design a new framework for VR called sample-specific multi-channel masks (SMM). Specifically, SMM employs a lightweight ConvNet and patch-wise interpolation to generate sample-specific three-channel masks instead of a shared and pre-defined mask. Since we generate different masks for individual samples, SMM is theoretically shown to reduce approximation error for the target tasks compared with existing state-of-the-art VR methods. We also empirically demonstrate its performance gain on both ResNet and ViT. The success of SMM further highlights the broader applicability of VR in leveraging the latent knowledge of pre-trained models for various target tasks. Our code is available at https://github.com/tmlr-group/SMM.

Yuhao Sun, Chengyi Cai, Jiacheng Zhang et al.

Recent research has shown that aligning fine-grained text descriptions with localized image patches can significantly improve the zero-shot performance of pre-trained vision-language models (e.g., CLIP). However, we find that both fine-grained text descriptions and localized image patches often contain redundant information, making text-visual alignment less effective. In this paper, we tackle this issue from two perspectives: \emph{View Refinement} and \emph{Description refinement}, termed as \textit{\textbf{Bi}-refinement for \textbf{F}ine-grained \textbf{T}ext-visual \textbf{A}lignment} (BiFTA). \emph{View refinement} removes redundant image patches with high \emph{Intersection over Union} (IoU) ratios, resulting in more distinctive visual samples. \emph{Description refinement} removes redundant text descriptions with high pairwise cosine similarity, ensuring greater diversity in the remaining descriptions. BiFTA achieves superior zero-shot performance on 6 benchmark datasets for both ViT-based and ResNet-based CLIP, justifying the necessity to remove redundant information in visual-text alignment.

Chengyi Cai, Zesheng Ye, Peike Li et al.

Unlearning in Multimodal Large Language Models (MLLMs) prevents the model from revealing private information when queried about target images. Existing MLLM unlearning methods largely adopt approaches developed for LLMs. They treat all answer tokens uniformly, disregarding their varying importance in the unlearning process. Moreover, these methods focus exclusively on the language modality, disregarding visual cues that indicate key tokens in answers. In this paper, after formulating the problem of unlearning in multimodal question answering for MLLMs, we propose Visual-Guided Key-Token Regularization (ViKeR). We leverage irrelevant visual inputs to predict ideal post-unlearning token-level distributions and use these distributions to regularize the unlearning process, thereby prioritizing key tokens. Further, we define key tokens in unlearning via information entropy and discuss ViKeR's effectiveness through token-level gradient reweighting, which amplifies updates on key tokens. Experiments on MLLMU and CLEAR benchmarks demonstrate that our method effectively performs unlearning while mitigating forgetting and maintaining response coherence.

Muxing Li, Zesheng Ye, Sharon Li et al.

Data rights owners can detect unauthorized data use in large language model (LLM) training by querying with proprietary samples. Often, superior performance (e.g., higher confidence or lower loss) on a sample relative to the untrained data implies it was part of the training corpus, as LLMs tend to perform better on data they have seen during training. However, this detection becomes fragile under data laundering, a practice of transforming the stylistic form of proprietary data, while preserving critical information to obfuscate data provenance. When an LLM is trained exclusively on such laundered variants, it no longer performs better on originals, erasing the signals that standard detections rely on. We counter this by inferring the unknown laundering transformation from black-box access to the target LLM and, via an auxiliary LLM, synthesizing queries that mimic the laundered data, even if rights owners have only the originals. As the search space of finding true laundering transformations is infinite, we abstract such a process into a high-level transformation goal (e.g., "lyrical rewriting") and concrete details (e.g., "with vivid imagery"), and introduce synthesis data reversion (SDR) that instantiates this abstraction. SDR first identifies the most probable goal for synthesis to narrow the search; it then iteratively refines details so that synthesized queries gradually elicit stronger detection signals from the target LLM. Evaluated on the MIMIR benchmark against diverse laundering practices and target LLM families (Pythia, Llama2, and Falcon), SDR consistently strengthens data misuse detection, providing a practical countermeasure to data laundering.

Muxing Li, Zesheng Ye, Sharon Li et al.

To detect unauthorized data usage in training large-scale generative models (e.g., ChatGPT or Midjourney), membership inference attacks (MIA) have proven effective in distinguishing a single training instance (a member) from a single non-training instance (a non-member). This success is mainly credited to a memorization effect: models tend to perform better on a member than a non-member. However, we find that standard MIAs fail against distilled generative models (i.e., student models) that are increasingly deployed in practice for efficiency (e.g., ChatGPT 4o-mini). Trained exclusively on data generated from a large-scale model (a teacher model), the student model lacks direct exposure to any members (teacher's training data), nullifying the memorization effect that standard MIAs rely on. This finding reveals a serious privacy loophole, where generation-service providers could deploy a student model whose teacher was potentially trained on unauthorized data, yet claim the deployed model is clean because it was not directly trained on such data. Hence, are distilled models inherently unauditable for upstream privacy violations, and should we discard them when we care about privacy? We contend no, as we uncover a memory chain connecting the student and teacher's member data: the distribution of student-generated data aligns more closely with the distribution of the teacher's members than with non-members, thus we can detect unauthorized data usage even when direct instance-level memorization is absent. This leads us to posit that MIAs on distilled generative models should shift from instance-level scores to distribution-level statistics. We further propose three principles of distribution-based MIAs for detecting unauthorized training data through distilled generative models, and validate our position through an exemplar framework. We lastly discuss the implications of our position.

Zesheng Ye, Chengyi Cai, Ruijiang Dong et al.

As large-scale pre-trained foundation models continue to expand in size and capability, efficiently adapting them to specific downstream tasks has become increasingly critical. Despite substantial progress, existing adaptation approaches have evolved largely in isolation, without a clear understanding of their interrelationships. This survey introduces neural network reprogrammability as a unifying framework that bridges mainstream model adaptation techniques--model reprogramming, prompt tuning, and prompt instruction--previously fragmented research areas yet converges on a shared principle: repurposing a pre-trained model by manipulating information at the interfaces while keeping the model parameters frozen. These methods exploit neural networks' sensitivity to manipulation on different interfaces, be it through perturbing inputs, inserting tokens into intermediate layers, or providing task-specific examples in context, to redirect model behaviors towards desired outcomes. We then present a taxonomy that categorizes such information manipulation-based adaptation approaches across four key dimensions: manipulation format (fixed or learnable), location (interfaces where manipulations occur), operator (how they are applied), and output alignment requirement (post-processing needed to align outputs with downstream tasks). Notably, this framework applies consistently across data modalities, independent of specific model architectures. Moreover, viewing established techniques like in-context learning and chain-of-thought prompting through this lens reveals both their theoretical connections and practical distinctions. We further analyze remaining technical challenges and ethical considerations, positioning neural network reprogrammability as a fundamental paradigm for efficient model adaptation. We lastly identify promising research directions emerging from this integrative viewpoint.

Jing Du, Zesheng Ye, Bin Guo et al.

Cross-Domain Recommendation (CDR) seeks to enable effective knowledge transfer across domains. Existing works rely on either representation alignment or transformation bridges, but they struggle on identifying domain-shared from domain-specific latent factors. Specifically, while CDR describes user representations as a joint distribution over two domains, these methods fail to account for its joint identifiability as they primarily fixate on the marginal distribution within a particular domain. Such a failure may overlook the conditionality between two domains and how it contributes to latent factor disentanglement, leading to negative transfer when domains are weakly correlated. In this study, we explore what should and should not be transferred in cross-domain user representations from a causality perspective. We propose a Hierarchical subspace disentanglement approach to explore the Joint IDentifiability of cross-domain joint distribution, termed HJID, to preserve domain-specific behaviors from domain-shared factors. HJID organizes user representations into layers: generic shallow subspaces and domain-oriented deep subspaces. We first encode the generic pattern in the shallow subspace by minimizing the Maximum Mean Discrepancy of initial layer activation. Then, to dissect how domain-oriented latent factors are encoded in deeper layers activation, we construct a cross-domain causality-based data generation graph, which identifies cross-domain consistent and domain-specific components, adhering to the Minimal Change principle. This allows HJID to maintain stability whilst discovering unique factors for different domains, all within a generative framework of invertible transformations that guarantee the joint identifiability. With experiments on real-world datasets, we show that HJID outperforms SOTA methods on a range of strongly and weakly correlated CDR tasks.