Ziqi Wen

Ziqi Wen, Tianqin Li, Zhi Jing et al. · cmu

Deep learning models are known to exhibit a strong texture bias, while human tends to rely heavily on global shape structure for object recognition. The current benchmark for evaluating a model's global shape bias is a set of style-transferred images with the assumption that resistance to the attack of style transfer is related to the development of global structure sensitivity in the model. In this work, we show that networks trained with style-transfer images indeed learn to ignore style, but its shape bias arises primarily from local detail. We provide a \textbf{Disrupted Structure Testbench (DiST)} as a direct measurement of global structure sensitivity. Our test includes 2400 original images from ImageNet-1K, each of which is accompanied by two images with the global shapes of the original image disrupted while preserving its texture via the texture synthesis program. We found that \textcolor{black}{(1) models that performed well on the previous cue-conflict dataset do not fare well in the proposed DiST; (2) the supervised trained Vision Transformer (ViT) lose its global spatial information from positional embedding, leading to no significant advantages over Convolutional Neural Networks (CNNs) on DiST. While self-supervised learning methods, especially mask autoencoder significantly improves the global structure sensitivity of ViT. (3) Improving the global structure sensitivity is orthogonal to resistance to style-transfer, indicating that the relationship between global shape structure and local texture detail is not an either/or relationship. Training with DiST images and style-transferred images are complementary, and can be combined to train network together to enhance the global shape sensitivity and robustness of local features.} Our code will be hosted in github: https://github.com/leelabcnbc/DiST

Tianqin Li, Ziqi Wen, Yangfan Li et al.

Current deep-learning models for object recognition are known to be heavily biased toward texture. In contrast, human visual systems are known to be biased toward shape and structure. What could be the design principles in human visual systems that led to this difference? How could we introduce more shape bias into the deep learning models? In this paper, we report that sparse coding, a ubiquitous principle in the brain, can in itself introduce shape bias into the network. We found that enforcing the sparse coding constraint using a non-differential Top-K operation can lead to the emergence of structural encoding in neurons in convolutional neural networks, resulting in a smooth decomposition of objects into parts and subparts and endowing the networks with shape bias. We demonstrated this emergence of shape bias and its functional benefits for different network structures with various datasets. For object recognition convolutional neural networks, the shape bias leads to greater robustness against style and pattern change distraction. For the image synthesis generative adversary networks, the emerged shape bias leads to more coherent and decomposable structures in the synthesized images. Ablation studies suggest that sparse codes tend to encode structures, whereas the more distributed codes tend to favor texture. Our code is host at the github repository: \url{https://github.com/Crazy-Jack/nips2023_shape_vs_texture}

Zizheng Huang, Haoxing Chen, Ziqi Wen et al.

Contrastive learning (CL) continuously achieves significant breakthroughs across multiple domains. However, the most common InfoNCE-based methods suffer from some dilemmas, such as \textit{uniformity-tolerance dilemma} (UTD) and \textit{gradient reduction}, both of which are related to a $\mathcal{P}_{ij}$ term. It has been identified that UTD can lead to unexpected performance degradation. We argue that the fixity of temperature is to blame for UTD. To tackle this challenge, we enrich the CL loss family by presenting a Model-Aware Contrastive Learning (MACL) strategy, whose temperature is adaptive to the magnitude of alignment that reflects the basic confidence of the instance discrimination task, then enables CL loss to adjust the penalty strength for hard negatives adaptively. Regarding another dilemma, the gradient reduction issue, we derive the limits of an involved gradient scaling factor, which allows us to explain from a unified perspective why some recent approaches are effective with fewer negative samples, and summarily present a gradient reweighting to escape this dilemma. Extensive remarkable empirical results in vision, sentence, and graph modality validate our approach's general improvement for representation learning and downstream tasks.

Ziqi Wen, Parsa Madinei, Miguel P. Eckstein

Evaluating whether large vision-language models (VLMs) align with human perception for high-level semantic scene comprehension remains a challenge. Traditional white-box interpretability methods are inapplicable to closed-source architectures and passive metrics fail to isolate causal features. We introduce Counterfactual Semantic Saliency (CSS). This black-box, model-agnostic framework quantifies the importance of objects by measuring the semantic shift induced by their causal ablation from a scene. To evaluate AI-human semantic alignment, we tested prominent VLMs against a human psychophysics baseline comprising 16,289 valid responses across 307 complex natural scenes and 1,306 high-fidelity counterfactual variants. Our analysis reveals a pervasive scene comprehension gap: models exhibit an overreliance (relative to humans) on large objects (size bias), objects at the center of the image (center bias), and high saliency objects. In contrast, models rely less on people in the scenes than our human participants to describe the images. A model's size bias is a primary driver explaining variations in model-human semantic divergence. Code and data will be available at https://github.com/starsky77/Counterfactual-Semantic-Saliency.

Shravan Murlidaran, Ziqi Wen, Sana Shehabi et al.

When humans view scenes without a specific task (free-viewing), they initially direct their eye movements toward the scene center and then fixate on people, text, objects being gazed at or grasped, and semantically meaningful regions. What these signature fixation patterns reflect and whether they optimize an underlying perceptual task remain unknown. We show that a computational agent with simulated foveation, trained to optimize scene comprehension, exhibits emergent human fixation signature patterns. In contrast, versions of the agent trained to search or classify scenes, or equipped with peripheral vision that was better or worse than human vision, predicted human fixation patterns less accurately. Thus, human free-viewing fixation patterns may emerge as a functional byproduct of optimizing scene comprehension under the biological constraints of foveated vision.

Parsa Madinei, Ryan Solgi, Ziqi Wen et al.

We introduce INTERLACE, a novel framework that prunes redundant layers in VLMs while maintaining performance through sample-efficient finetuning. Existing layer pruning methods lead to significant performance drop when applied to VLMs. Instead, we analyze triplets of consecutive layers to identify local redundancy, removing the most redundant of the first two layers, finetune the remaining layer to compensate for the lost capacity, and freeze the third layer to serve as a stable anchor during finetuning. We found that this interleaved finetune-freeze design enables rapid convergence with minimal data after pruning. By finetuning only a subset of layers on just 1% of the FineVision dataset for one epoch, Interlace achieves 88.9% average performance retention after dropping 25% of the network, achieving SOTA performance. Our code is available at: https://github.com/pmadinei/Interlace.git

Tianqin Li, Ziqi Wen, Leiran Song et al.

Human vision organizes local cues into coherent global forms using Gestalt principles like closure, proximity, and figure-ground assignment -- functions reliant on global spatial structure. We investigate whether modern vision models show similar behaviors, and under what training conditions these emerge. We find that Vision Transformers (ViTs) trained with Masked Autoencoding (MAE) exhibit activation patterns consistent with Gestalt laws, including illusory contour completion, convexity preference, and dynamic figure-ground segregation. To probe the computational basis, we hypothesize that modeling global dependencies is necessary for Gestalt-like organization. We introduce the Distorted Spatial Relationship Testbench (DiSRT), which evaluates sensitivity to global spatial perturbations while preserving local textures. Using DiSRT, we show that self-supervised models (e.g., MAE, CLIP) outperform supervised baselines and sometimes even exceed human performance. ConvNeXt models trained with MAE also exhibit Gestalt-compatible representations, suggesting such sensitivity can arise without attention architectures. However, classification finetuning degrades this ability. Inspired by biological vision, we show that a Top-K activation sparsity mechanism can restore global sensitivity. Our findings identify training conditions that promote or suppress Gestalt-like perception and establish DiSRT as a diagnostic for global structure sensitivity across models.

Jonathan Skaza, Parsa Madinei, Ziqi Wen et al.

Visual complexity prediction is a fundamental problem in computer vision with applications in image compression, retrieval, and classification. Understanding what makes humans perceive an image as complex is also a long-standing question in cognitive science. Recent approaches have leveraged multimodal models that combine visual and linguistic representations, but it remains unclear whether language information is necessary for this task. We propose DReX (DINO-ResNet Fusion), a vision-only model that fuses self-supervised and convolutional representations through a learnable attention mechanism to predict image complexity. Our architecture integrates multi-scale hierarchical features from ResNet-50 with semantically rich representations from DINOv3 ViT-S/16, enabling the model to capture both low-level texture patterns and high-level semantic structure. DReX achieves state-of-the-art performance on the IC9600 benchmark (Pearson r = 0.9581), surpassing previous methods--including those trained on multimodal image-text data--while using approximately 21.5x fewer learnable parameters. Furthermore, DReX generalizes robustly across multiple datasets and metrics, achieving superior results on Pearson and Spearman correlation, Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). Ablation and attention analyses confirm that DReX leverages complementary cues from both backbones, with the DINOv3 [CLS] token enhancing sensitivity to visual complexity. Our findings suggest that visual features alone can be sufficient for human-aligned complexity prediction and that, when properly fused, self-supervised transformers and supervised deep convolutional neural networks offer complementary and synergistic benefits for this task.

Ziqi Wen, Jonathan Skaza, Shravan Murlidaran et al.

Although models exist that predict human response times (RTs) in tasks such as target search and visual discrimination, the development of image-computable predictors for scene understanding time remains an open challenge. Recent advances in vision-language models (VLMs), which can generate scene descriptions for arbitrary images, combined with the availability of quantitative metrics for comparing linguistic descriptions, offer a new opportunity to model human scene understanding. We hypothesize that the primary bottleneck in human scene understanding and the driving source of variability in response times across scenes is the interaction between the foveated nature of the human visual system and the spatial distribution of task-relevant visual information within an image. Based on this assumption, we propose a novel image-computable model that integrates foveated vision with VLMs to produce a spatially resolved map of scene understanding as a function of fixation location (Foveated Scene Understanding Map, or F-SUM), along with an aggregate F-SUM score. This metric correlates with average (N=17) human RTs (r=0.47) and number of saccades (r=0.51) required to comprehend a scene (across 277 scenes). The F-SUM score also correlates with average (N=16) human description accuracy (r=-0.56) in time-limited presentations. These correlations significantly exceed those of standard image-based metrics such as clutter, visual complexity, and scene ambiguity based on language entropy. Together, our work introduces a new image-computable metric for predicting human response times in scene understanding and demonstrates the importance of foveated visual processing in shaping comprehension difficulty.