Tai Sing Lee

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}

Tianye Wang, Haoxuan Yao, Tai Sing Lee et al.

The visual system evolved to process natural scenes, yet most of our understanding of the topology and function of visual cortex derives from studies using artificial stimuli. To gain deeper insights into visual processing of natural scenes, we utilized widefield calcium-imaging of primate V4 in response to many natural images, generating a large dataset of columnar-scale responses. We used this dataset to build a digital twin of V4 via deep learning, generating a detailed topographical map of natural image preferences at each cortical position. The map revealed clustered functional domains for specific classes of natural image features. These ranged from surface-related attributes like color and texture to shape-related features such as edges, curvature, and facial features. We validated the model-predicted domains with additional widefield calcium-imaging and single-cell resolution two-photon imaging. Our study illuminates the detailed topological organization and neural codes in V4 that represent natural scenes.

Nikolas McNeal, Jennifer Huang, Aniekan Umoren et al.

We explore the use of a neural network inspired by predictive coding for modeling human music perception. This network was developed based on the computational neuroscience theory of recurrent interactions in the hierarchical visual cortex. When trained with video data using self-supervised learning, the model manifests behaviors consistent with human visual illusions. Here, we adapt this network to model the hierarchical auditory system and investigate whether it will make similar choices to humans regarding the musicality of a set of random pitch sequences. When the model is trained with a large corpus of instrumental classical music and popular melodies rendered as mel spectrograms, it exhibits greater prediction errors for random pitch sequences that are rated less musical by human subjects. We found that the prediction error depends on the amount of information regarding the subsequent note, the pitch interval, and the temporal context. Our findings suggest that predictability is correlated with human perception of musicality and that a predictive coding neural network trained on music can be used to characterize the features and motifs contributing to human perception of music.

Jiayuan Liu, Tianqin Li, Shiyi Du et al.

Context window expansion is often treated as a straightforward capability upgrade for LLMs, but we find it systematically fails in multi-agent social dilemmas. Across 7 LLMs and 4 games over 500 rounds, expanding accessible history degrades cooperation in 18 of 28 model--game settings, a pattern we term the memory curse. We isolate the underlying mechanism through three analyses. First, lexical analysis of 378,000 reasoning traces associates this breakdown with eroding forward-looking intent rather than rising paranoia. We validate this using targeted fine-tuning as a cognitive probe: a LoRA adapter trained exclusively on forward-looking traces mitigates the decay and transfers zero-shot to distinct games. Second, memory sanitization holds prompt length fixed while replacing visible history with synthetic cooperative records, which restores cooperation substantially, proving the trigger is memory content, not length alone. Finally, ablating explicit Chain-of-Thought reasoning often reduces the collapse, showing that deliberation paradoxically amplifies the memory curse. Together, these results recast memory as an active determinant of multi-agent behavior: longer recall can either destabilize or support cooperation depending on the reasoning patterns it elicits.

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.

Tianqin Li, Junru Zhao, Dunhan Jiang et al.

David Marr's seminal theory of human perception stipulates that visual processing is a multi-stage process, prioritizing the derivation of boundary and surface properties before forming semantic object representations. In contrast, contrastive representation learning frameworks typically bypass this explicit multi-stage approach, defining their objective as the direct learning of a semantic representation space for objects. While effective in general contexts, this approach sacrifices the inductive biases of vision, leading to slower convergence speed and learning shortcut resulting in texture bias. In this work, we demonstrate that leveraging Marr's multi-stage theory-by first constructing boundary and surface-level representations using perceptual constructs from early visual processing stages and subsequently training for object semantics-leads to 2x faster convergence on ResNet18, improved final representations on semantic segmentation, depth estimation, and object recognition, and enhanced robustness and out-of-distribution capability. Together, we propose a pretraining stage before the general contrastive representation pretraining to further enhance the final representation quality and reduce the overall convergence time via inductive bias from human vision systems.

Tianqin Li, George Liu, Tai Sing Lee

Despite remarkable progress in computer vision, modern recognition systems remain fundamentally limited by their dependence on rich, redundant visual inputs. In contrast, humans can effortlessly understand sparse, minimal representations like line drawings, suggesting that structure, rather than appearance, underlies efficient visual understanding. In this work, we propose a novel structure-first learning paradigm that uses line drawings as an initial training modality to induce more compact and generalizable visual representations. We demonstrate that models trained with this approach develop a stronger shape bias, more focused attention, and greater data efficiency across classification, detection, and segmentation tasks. Notably, these models also exhibit lower intrinsic dimensionality, requiring significantly fewer principal components to capture representational variance, which mirrors observations of low-dimensional, efficient representations in the human brain. Beyond performance improvements, structure-first learning produces more compressible representations, enabling better distillation into lightweight student models. Students distilled from teachers trained on line drawings consistently outperform those trained from color-supervised teachers, highlighting the benefits of structurally compact knowledge. Together, our results support the view that structure-first visual learning fosters efficiency, generalization, and human-aligned inductive biases, offering a simple yet powerful strategy for building more robust and adaptable vision systems.

Yue Li, Weifan Wang, Tai Sing Lee

Recent neurophysiological studies have revealed that the early visual cortex can rapidly learn global image context, as evidenced by a sparsification of population responses and a reduction in mean activity when exposed to familiar versus novel image contexts. This phenomenon has been attributed primarily to local recurrent interactions, rather than changes in feedforward or feedback pathways, supported by both empirical findings and circuit-level modeling. Recurrent neural circuits capable of simulating these effects have been shown to reshape the geometry of neural manifolds, enhancing robustness and invariance to irrelevant variations. In this study, we employ a Vision Transformer (ViT)-based autoencoder to investigate, from a functional perspective, how familiarity training can induce sensitivity to global context in the early layers of a deep neural network. We hypothesize that rapid learning operates via fast weights, which encode transient or short-term memory traces, and we explore the use of Low-Rank Adaptation (LoRA) to implement such fast weights within each Transformer layer. Our results show that (1) The proposed ViT-based autoencoder's self-attention circuit performs a manifold transform similar to a neural circuit model of the familiarity effect. (2) Familiarity training aligns latent representations in early layers with those in the top layer that contains global context information. (3) Familiarity training broadens the self-attention scope within the remembered image context. (4) These effects are significantly amplified by LoRA-based fast weights. Together, these findings suggest that familiarity training introduces global sensitivity to earlier layers in a hierarchical network, and that a hybrid fast-and-slow weight architecture may provide a viable computational model for studying rapid global context learning in the brain.

Isaac Lin, Tianye Wang, Shang Gao et al.

Convolutional neural networks (CNNs) have been shown to be state-of-the-art models for visual cortical neurons. Cortical neurons in the primary visual cortex are sensitive to contextual information mediated by extensive horizontal and feedback connections. Standard CNNs integrate global contextual information to model contextual modulation via two mechanisms: successive convolutions and a fully connected readout layer. In this paper, we find that self-attention (SA), an implementation of non-local network mechanisms, can improve neural response predictions over parameter-matched CNNs in two key metrics: tuning curve correlation and peak tuning. We introduce peak tuning as a metric to evaluate a model's ability to capture a neuron's top feature preference. We factorize networks to assess each context mechanism, revealing that information in the local receptive field is most important for modeling overall tuning, but surround information is critically necessary for characterizing the tuning peak. We find that self-attention can replace posterior spatial-integration convolutions when learned incrementally, and is further enhanced in the presence of a fully connected readout layer, suggesting that the two context mechanisms are complementary. Finally, we find that decomposing receptive field learning and contextual modulation learning in an incremental manner may be an effective and robust mechanism for learning surround-center interactions.

Andrew Luo, Tianqin Li, Wen-Hao Zhang et al.

Recent advances in deep generative models have led to immense progress in 3D shape synthesis. While existing models are able to synthesize shapes represented as voxels, point-clouds, or implicit functions, these methods only indirectly enforce the plausibility of the final 3D shape surface. Here we present a 3D shape synthesis framework (SurfGen) that directly applies adversarial training to the object surface. Our approach uses a differentiable spherical projection layer to capture and represent the explicit zero isosurface of an implicit 3D generator as functions defined on the unit sphere. By processing the spherical representation of 3D object surfaces with a spherical CNN in an adversarial setting, our generator can better learn the statistics of natural shape surfaces. We evaluate our model on large-scale shape datasets, and demonstrate that the end-to-end trained model is capable of generating high fidelity 3D shapes with diverse topology.

Yimeng Zhang, Harold Rockwell, Sicheng Dai et al.

Feedforward CNN models have proven themselves in recent years as state-of-the-art models for predicting single-neuron responses to natural images in early visual cortical neurons. In this paper, we extend these models with recurrent convolutional layers, reflecting the well-known massive recurrence in the cortex, and show robust increases in predictive performance over feedforward models across thousands of hyperparameter combinations in three datasets of macaque V1 and V2 single-neuron responses. We propose the recurrent circuit can be conceptualized as a form of ensemble computing, with each iteration generating more effective feedforward paths of various path lengths to allow a combination of solutions in the final approximation. The statistics of the paths in the ensemble provide insights to the differential performance increases among our recurrent models. We also assess whether the recurrent circuits learned for neural response prediction can be related to cortical circuits. We find that the hidden units in the recurrent circuits of the appropriate models, when trained on long-duration wide-field image presentations, exhibit similar temporal response dynamics and classical contextual modulations as observed in V1 neurons. This work provides insights to the computational rationale of recurrent circuits and suggests that neural response prediction could be useful for characterizing the recurrent neural circuits in the visual cortex.

Siming Yan, Xuyang Fang, Bowen Xiao et al.

The abundant recurrent horizontal and feedback connections in the primate visual cortex are thought to play an important role in bringing global and semantic contextual information to early visual areas during perceptual inference, helping to resolve local ambiguity and fill in missing details. In this study, we find that introducing feedback loops and horizontal recurrent connections to a deep convolution neural network (VGG16) allows the network to become more robust against noise and occlusion during inference, even in the initial feedforward pass. This suggests that recurrent feedback and contextual modulation transform the feedforward representations of the network in a meaningful and interesting way. We study the population codes of neurons in the network, before and after learning with feedback, and find that learning with feedback yielded an increase in discriminability (measured by d-prime) between the different object classes in the population codes of the neurons in the feedforward path, even at the earliest layer that receives feedback. We find that recurrent feedback, by injecting top-down semantic meaning to the population activities, helps the network learn better feedforward paths to robustly map noisy image patches to the latent representations corresponding to important visual concepts of each object class, resulting in greater robustness of the network against noises and occlusion as well as better fine-grained recognition.

Jielin Qiu, Ge Huang, Tai Sing Lee

In this paper we developed a hierarchical network model, called Hierarchical Prediction Network (HPNet), to understand how spatiotemporal memories might be learned and encoded in the recurrent circuits in the visual cortical hierarchy for predicting future video frames. This neurally inspired model operates in the analysis-by-synthesis framework. It contains a feed-forward path that computes and encodes spatiotemporal features of successive complexity and a feedback path for the successive levels to project their interpretations to the level below. Within each level, the feed-forward path and the feedback path intersect in a recurrent gated circuit, instantiated in a LSTM module, to generate a prediction or explanation of the incoming signals. The network learns its internal model of the world by minimizing the errors of its prediction of the incoming signals at each level of the hierarchy. We found that hierarchical interaction in the network increases semantic clustering of global movement patterns in the population codes of the units along the hierarchy, even in the earliest module. This facilitates the learning of relationships among movement patterns, yielding state-of-the-art performance in long range video sequence predictions in the benchmark datasets. The network model automatically reproduces a variety of prediction suppression and familiarity suppression neurophysiological phenomena observed in the visual cortex, suggesting that hierarchical prediction might indeed be an important principle for representational learning in the visual cortex.

Heqing Ya, Haonan Sun, Jeffrey Helt et al.

We develop an approach for unsupervised learning of associations between co-occurring perceptual events using a large graph. We applied this approach to successfully solve the image captcha of China's railroad system. The approach is based on the principle of suspicious coincidence. In this particular problem, a user is presented with a deformed picture of a Chinese phrase and eight low-resolution images. They must quickly select the relevant images in order to purchase their train tickets. This problem presents several challenges: (1) the teaching labels for both the Chinese phrases and the images were not available for supervised learning, (2) no pre-trained deep convolutional neural networks are available for recognizing these Chinese phrases or the presented images, and (3) each captcha must be solved within a few seconds. We collected 2.6 million captchas, with 2.6 million deformed Chinese phrases and over 21 million images. From these data, we constructed an association graph, composed of over 6 million vertices, and linked these vertices based on co-occurrence information and feature similarity between pairs of images. We then trained a deep convolutional neural network to learn a projection of the Chinese phrases onto a 230-dimensional latent space. Using label propagation, we computed the likelihood of each of the eight images conditioned on the latent space projection of the deformed phrase for each captcha. The resulting system solved captchas with 77% accuracy in 2 seconds on average. Our work, in answering this practical challenge, illustrates the power of this class of unsupervised association learning techniques, which may be related to the brain's general strategy for associating language stimuli with visual objects on the principle of suspicious coincidence.

Hao Wang, Xingyu Lin, Yimeng Zhang et al.

Recurrent feedback connections in the mammalian visual system have been hypothesized to play a role in synthesizing input in the theoretical framework of analysis by synthesis. The comparison of internally synthesized representation with that of the input provides a validation mechanism during perceptual inference and learning. Inspired by these ideas, we proposed that the synthesis machinery can compose new, unobserved images by imagination to train the network itself so as to increase the robustness of the system in novel scenarios. As a proof of concept, we investigated whether images composed by imagination could help an object recognition system to deal with occlusion, which is challenging for the current state-of-the-art deep convolutional neural networks. We fine-tuned a network on images containing objects in various occlusion scenarios, that are imagined or self-generated through a deep generator network. Trained on imagined occluded scenarios under the object persistence constraint, our network discovered more subtle and localized image features that were neglected by the original network for object classification, obtaining better separability of different object classes in the feature space. This leads to significant improvement of object recognition under occlusion for our network relative to the original network trained only on un-occluded images. In addition to providing practical benefits in object recognition under occlusion, this work demonstrates the use of self-generated composition of visual scenes through the synthesis loop, combined with the object persistence constraint, can provide opportunities for neural networks to discover new relevant patterns in the data, and become more flexible in dealing with novel situations.

Xingyu Lin, Hao Wang, Zhihao Li et al.

We develop a model of perceptual similarity judgment based on re-training a deep convolution neural network (DCNN) that learns to associate different views of each 3D object to capture the notion of object persistence and continuity in our visual experience. The re-training process effectively performs distance metric learning under the object persistency constraints, to modify the view-manifold of object representations. It reduces the effective distance between the representations of different views of the same object without compromising the distance between those of the views of different objects, resulting in the untangling of the view-manifolds between individual objects within the same category and across categories. This untangling enables the model to discriminate and recognize objects within the same category, independent of viewpoints. We found that this ability is not limited to the trained objects, but transfers to novel objects in both trained and untrained categories, as well as to a variety of completely novel artificial synthetic objects. This transfer in learning suggests the modification of distance metrics in view- manifolds is more general and abstract, likely at the levels of parts, and independent of the specific objects or categories experienced during training. Interestingly, the resulting transformation of feature representation in the deep networks is found to significantly better match human perceptual similarity judgment than AlexNet, suggesting that object persistence could be an important constraint in the development of perceptual similarity judgment in biological neural networks.

Mingmin Zhao, Chengxu Zhuang, Yizhou Wang et al.

We propose a new neurally-inspired model that can learn to encode the global relationship context of visual events across time and space and to use the contextual information to modulate the analysis by synthesis process in a predictive coding framework. The model learns latent contextual representations by maximizing the predictability of visual events based on local and global contextual information through both top-down and bottom-up processes. In contrast to standard predictive coding models, the prediction error in this model is used to update the contextual representation but does not alter the feedforward input for the next layer, and is thus more consistent with neurophysiological observations. We establish the computational feasibility of this model by demonstrating its ability in several aspects. We show that our model can outperform state-of-art performances of gated Boltzmann machines (GBM) in estimation of contextual information. Our model can also interpolate missing events or predict future events in image sequences while simultaneously estimating contextual information. We show it achieves state-of-art performances in terms of prediction accuracy in a variety of tasks and possesses the ability to interpolate missing frames, a function that is lacking in GBM.

Xiong Li, Tai Sing Lee, Yuncai Liu

Probabilistic generative modeling of data distributions can potentially exploit hidden information which is useful for discriminative classification. This observation has motivated the development of approaches that couple generative and discriminative models for classification. In this paper, we propose a new approach to couple generative and discriminative models in an unified framework based on PAC-Bayes risk theory. We first derive the model-parameter-independent stochastic feature mapping from a practical MAP classifier operating on generative models. Then we construct a linear stochastic classifier equipped with the feature mapping, and derive the explicit PAC-Bayes risk bounds for such classifier for both supervised and semi-supervised learning. Minimizing the risk bound, using an EM-like iterative procedure, results in a new posterior over hidden variables (E-step) and the update rules of model parameters (M-step). The derivation of the posterior is always feasible due to the way of equipping feature mapping and the explicit form of bounding risk. The derived posterior allows the tuning of generative models and subsequently the feature mappings for better classification. The derived update rules of the model parameters are same to those of the uncoupled models as the feature mapping is model-parameter-independent. Our experiments show that the coupling between data modeling generative model and the discriminative classifier via a stochastic feature mapping in this framework leads to a general classification tool with state-of-the-art performance.