Recent advances in diffusion models have demonstrated their remarkable ability to capture complex image distributions, but the geometric properties of the learned data manifold remain poorly understood. We address this gap by introducing a score-based Riemannian metric that leverages the Stein score function from diffusion models to characterize the intrinsic geometry of the data manifold without requiring explicit parameterization. Our approach defines a metric tensor in the ambient space that stretches distances perpendicular to the manifold while preserving them along tangential directions, effectively creating a geometry where geodesics naturally follow the manifold's contours. We develop efficient algorithms for computing these geodesics and demonstrate their utility for both interpolation between data points and extrapolation beyond the observed data distribution. Through experiments on synthetic data with known geometry, Rotated MNIST, and complex natural images via Stable Diffusion, we show that our score-based geodesics capture meaningful transformations that respect the underlying data distribution. Our method consistently outperforms baseline approaches on perceptual metrics (LPIPS) and distribution-level metrics (FID, KID), producing smoother, more realistic image transitions. These results reveal the implicit geometric structure learned by diffusion models and provide a principled way to navigate the manifold of natural images through the lens of Riemannian geometry.
In this paper, we propose a new method called Gradual Domain Osmosis, which aims to solve the problem of smooth knowledge migration from source domain to target domain in Gradual Domain Adaptation (GDA). Traditional Gradual Domain Adaptation methods mitigate domain bias by introducing intermediate domains and self-training strategies, but often face the challenges of inefficient knowledge migration or missing data in intermediate domains. In this paper, we design an optimisation framework based on the hyperparameter $\lambda$ by dynamically balancing the loss weights of the source and target domains, which enables the model to progressively adjust the strength of knowledge migration ($\lambda$ incrementing from 0 to 1) during the training process, thus achieving cross-domain generalisation more efficiently. Specifically, the method incorporates self-training to generate pseudo-labels and iteratively updates the model by minimising a weighted loss function to ensure stability and robustness during progressive adaptation in the intermediate domain. The experimental part validates the effectiveness of the method on rotated MNIST, colour-shifted MNIST, portrait dataset and forest cover type dataset, and the results show that it outperforms existing baseline methods. The paper further analyses the impact of the dynamic tuning strategy of the hyperparameter $\lambda$ on the performance through ablation experiments, confirming the advantages of progressive domain penetration in mitigating the domain bias and enhancing the model generalisation capability. The study provides a theoretical support and practical framework for asymptotic domain adaptation and expands its application potential in dynamic environments.
Symmetries (transformations by group actions) are present in many datasets, and leveraging them holds significant promise for improving predictions in machine learning. In this work, we aim to understand when and how deep networks can learn symmetries from data. We focus on a supervised classification paradigm where data symmetries are only partially observed during training: some classes include all transformations of a cyclic group, while others include only a subset. We ask: can deep networks generalize symmetry invariance to the partially sampled classes? In the infinite-width limit, where kernel analogies apply, we derive a neural kernel theory of symmetry learning to address this question. The group-cyclic nature of the dataset allows us to analyze the spectrum of neural kernels in the Fourier domain; here we find a simple characterization of the generalization error as a function of the interaction between class separation (signal) and class-orbit density (noise). We observe that generalization can only be successful when the local structure of the data prevails over its non-local, symmetric, structure, in the kernel space defined by the architecture. This occurs when (1) classes are sufficiently distinct and (2) class orbits are sufficiently dense. Our framework also applies to equivariant architectures (e.g., CNNs), and recovers their success in the special case where the architecture matches the inherent symmetry of the data. Empirically, our theory reproduces the generalization failure of finite-width networks (MLP, CNN, ViT) trained on partially observed versions of rotated-MNIST. We conclude that conventional networks trained with supervision lack a mechanism to learn symmetries that have not been explicitly embedded in their architecture a priori. Our framework could be extended to guide the design of architectures and training procedures able to learn symmetries from data.
Modern applications often leverage multiple views of a subject of study. Within neuroscience, there is growing interest in large-scale simultaneous recordings across multiple brain regions. Understanding the relationship between views (e.g., the neural activity in each region recorded) can reveal fundamental principles about the characteristics of each representation and about the system. However, existing methods to characterize such relationships either lack the expressivity required to capture complex nonlinearities, describe only sources of variance that are shared between views, or discard geometric information that is crucial to interpreting the data. Here, we develop a nonlinear neural network-based method that, given paired samples of high-dimensional views, disentangles low-dimensional shared and private latent variables underlying these views while preserving intrinsic data geometry. Across multiple simulated and real datasets, we demonstrate that our method outperforms competing methods. Using simulated populations of lateral geniculate nucleus (LGN) and V1 neurons we demonstrate our model's ability to discover interpretable shared and private structure across different noise conditions. On a dataset of unrotated and corresponding but randomly rotated MNIST digits, we recover private latents for the rotated view that encode rotation angle regardless of digit class, and places the angle representation on a 1-d manifold, while shared latents encode digit class but not rotation angle. Applying our method to simultaneous Neuropixels recordings of hippocampus and prefrontal cortex while mice run on a linear track, we discover a low-dimensional shared latent space that encodes the animal's position. We propose our approach as a general-purpose method for finding succinct and interpretable descriptions of paired data sets in terms of disentangled shared and private latent variables.
Normalizing flow-based generative models have been widely used in applications where the exact density estimation is of major importance. Recent research proposes numerous methods to improve their expressivity. However, conditioning on a context is largely overlooked area in the bijective flow research. Conventional conditioning with the vector concatenation is limited to only a few flow types. More importantly, this approach cannot support a practical setup where a set of context-conditioned (specialist) models are trained with the fixed pretrained general-knowledge (generalist) model. We propose ContextFlow++ approach to overcome these limitations using an additive conditioning with explicit generalist-specialist knowledge decoupling. Furthermore, we support discrete contexts by the proposed mixed-variable architecture with context encoders. Particularly, our context encoder for discrete variables is a surjective flow from which the context-conditioned continuous variables are sampled. Our experiments on rotated MNIST-R, corrupted CIFAR-10C, real-world ATM predictive maintenance and SMAP unsupervised anomaly detection benchmarks show that the proposed ContextFlow++ offers faster stable training and achieves higher performance metrics. Our code is publicly available at https://github.com/gudovskiy/contextflow.
Conventional domain adaptation typically transfers knowledge from a source domain to a stationary target domain. However, in many real-world cases, target data usually emerge sequentially and have continuously evolving distributions. Restoring and adapting to such target data results in escalating computational and resource consumption over time. Hence, it is vital to devise algorithms to address the evolving domain adaptation (EDA) problem, \emph{i.e.,} adapting models to evolving target domains without access to historic target domains. To achieve this goal, we propose a simple yet effective approach, termed progressive conservative adaptation (PCAda). To manage new target data that diverges from previous distributions, we fine-tune the classifier head based on the progressively updated class prototypes. Moreover, as adjusting to the most recent target domain can interfere with the features learned from previous target domains, we develop a conservative sparse attention mechanism. This mechanism restricts feature adaptation within essential dimensions, thus easing the inference related to historical knowledge. The proposed PCAda is implemented with a meta-learning framework, which achieves the fast adaptation of the classifier with the help of the progressively updated class prototypes in the inner loop and learns a generalized feature without severely interfering with the historic knowledge via the conservative sparse attention in the outer loop. Experiments on Rotated MNIST, Caltran, and Portraits datasets demonstrate the effectiveness of our method.
The weight-sharing mechanism of convolutional kernels ensures translation-equivariance of convolution neural networks (CNNs). Recently, rotation-equivariance has been investigated. However, research on scale-equivariance or simultaneous scale-rotation-equivariance is insufficient. This study proposes a Lie group-CNN, which can keep scale-rotation-equivariance for image classification tasks. The Lie group-CNN includes a lifting module, a series of group convolution modules, a global pooling layer, and a classification layer. The lifting module transfers the input image from Euclidean space to Lie group space, and the group convolution is parameterized through a fully connected network using Lie-algebra of Lie-group elements as inputs to achieve scale-rotation-equivariance. The Lie group SIM(2) is utilized to establish the Lie group-CNN with scale-rotation-equivariance. Scale-rotation-equivariance of Lie group-CNN is verified and achieves the best recognition accuracy on the blood cell dataset (97.50%) and the HAM10000 dataset (77.90%) superior to Lie algebra convolution network, dilation convolution, spatial transformer network, and scale-equivariant steerable network. In addition, the generalization ability of the Lie group-CNN on SIM(2) on rotation-equivariance is verified on rotated-MNIST and rotated-CIFAR10, and the robustness of the network is verified on SO(2) and SE(2). Therefore, the Lie group-CNN can successfully extract geometric features and performs equivariant recognition on images with rotation and scale transformations.
Much work has been devoted to devising architectures that build group-equivariant representations, while invariance is often induced using simple global pooling mechanisms. Little work has been done on creating expressive layers that are invariant to given symmetries, despite the success of permutation invariant pooling in various molecular tasks. In this work, we present Group Invariant Global Pooling (GIGP), an invariant pooling layer that is provably sufficiently expressive to represent a large class of invariant functions. We validate GIGP on rotated MNIST and QM9, showing improvements for the latter while attaining identical results for the former. By making the pooling process group orbit-aware, this invariant aggregation method leads to improved performance, while performing well-principled group aggregation.
Biological neural networks are capable of recruiting different sets of neurons to encode different memories. However, when training artificial neural networks on a set of tasks, typically, no mechanism is employed for selectively producing anything analogous to these neuronal ensembles. Further, artificial neural networks suffer from catastrophic forgetting, where the network's performance rapidly deteriorates as tasks are learned sequentially. By contrast, sequential learning is possible for a range of biological organisms. We introduce Learned Context Dependent Gating (LXDG), a method to flexibly allocate and recall `artificial neuronal ensembles', using a particular network structure and a new set of regularization terms. Activities in the hidden layers of the network are modulated by gates, which are dynamically produced during training. The gates are outputs of networks themselves, trained with a sigmoid output activation. The regularization terms we have introduced correspond to properties exhibited by biological neuronal ensembles. The first term penalizes low gate sparsity, ensuring that only a specified fraction of the network is used. The second term ensures that previously learned gates are recalled when the network is presented with input from previously learned tasks. Finally, there is a regularization term responsible for ensuring that new tasks are encoded in gates that are as orthogonal as possible from previously used ones. We demonstrate the ability of this method to alleviate catastrophic forgetting on continual learning benchmarks. When the new regularization terms are included in the model along with Elastic Weight Consolidation (EWC) it achieves better performance on the benchmark `permuted MNIST' than with EWC alone. The benchmark `rotated MNIST' demonstrates how similar tasks recruit similar neurons to the artificial neuronal ensemble.
Deep neural networks lack straightforward ways to incorporate domain knowledge and are notoriously considered black boxes. Prior works attempted to inject domain knowledge into architectures implicitly through data augmentation. Building on recent advances on equivariant neural networks, we propose networks that explicitly encode domain knowledge, specifically equivariance with respect to rotations. By using unfolded architectures, a rich framework that originated from sparse coding and has theoretical guarantees, we present interpretable networks with sparse activations. The equivariant unfolded networks compete favorably with baselines, with only a fraction of their parameters, as showcased on (rotated) MNIST and CIFAR-10.
The convolutional layers of standard convolutional neural networks (CNNs) are equivariant to translation. However, the convolution and fully-connected layers are not equivariant or invariant to other affine geometric transformations. Recently, a new class of CNNs is proposed in which the conventional layers of CNNs are replaced with equivariant convolution, pooling, and batch-normalization layers. The final classification layer in equivariant neural networks is invariant to different affine geometric transformations such as rotation, reflection and translation, and the scalar value is obtained by either eliminating the spatial dimensions of filter responses using convolution and down-sampling throughout the network or average is taken over the filter responses. In this work, we propose to integrate the orthogonal moments which gives the high-order statistics of the function as an effective means for encoding global invariance with respect to rotation, reflection and translation in fully-connected layers. As a result, the intermediate layers of the network become equivariant while the classification layer becomes invariant. The most widely used Zernike, pseudo-Zernike and orthogonal Fourier-Mellin moments are considered for this purpose. The effectiveness of the proposed work is evaluated by integrating the invariant transition and fully-connected layer in the architecture of group-equivariant CNNs (G-CNNs) on rotated MNIST and CIFAR10 datasets.
Machine learning algorithms minimizing the average training loss usually suffer from poor generalization performance due to the greedy exploitation of correlations among the training data, which are not stable under distributional shifts. It inspires various works for domain generalization (DG), where a series of methods, such as Causal Matching and FISH, work by pairwise domain operations. They would need $O(n^2)$ pairwise domain operations with $n$ domains, where each one is often highly expensive. Moreover, while a common objective in the DG literature is to learn invariant representations against domain-induced spurious correlations, we highlight the importance of mitigating spurious correlations caused by objects. Based on the observation that diversity helps mitigate spurious correlations, we propose a Diversity boosted twO-level saMplIng framework (DOMI) utilizing Determinantal Point Processes (DPPs) to efficiently sample the most informative ones among large number of domains. We show that DOMI helps train robust models against spurious correlations from both domain-side and object-side, substantially enhancing the performance of the backbone DG algorithms on rotated MNIST, rotated Fashion MNIST, and iwildcam datasets.
Domain generalization aims at learning a universal model that performs well on unseen target domains, incorporating knowledge from multiple source domains. In this research, we consider the scenario where different domain shifts occur among conditional distributions of different classes across domains. When labeled samples in the source domains are limited, existing approaches are not sufficiently robust. To address this problem, we propose a novel domain generalization framework called {Wasserstein Distributionally Robust Domain Generalization} (WDRDG), inspired by the concept of distributionally robust optimization. We encourage robustness over conditional distributions within class-specific Wasserstein uncertainty sets and optimize the worst-case performance of a classifier over these uncertainty sets. We further develop a test-time adaptation module leveraging optimal transport to quantify the relationship between the unseen target domain and source domains to make adaptive inference for target data. Experiments on the Rotated MNIST, PACS and the VLCS datasets demonstrate that our method could effectively balance the robustness and discriminability in challenging generalization scenarios.
This paper demonstrates that a simple modification of the variational autoencoder (VAE) formalism enables the method to identify and classify rotated and distorted digits. In particular, the conventional objective (cost) function employed during the training process of a VAE both quantifies the agreement between the input and output data records and ensures that the latent space representation of the input data record is statistically generated with an appropriate mean and standard deviation. After training, simulated data realizations are generated by decoding appropriate latent space points. Since, however, standard VAE:s trained on randomly rotated MNIST digits cannot reliably distinguish between different digit classes since the rotated input data is effectively compared to a similarly rotated output data record. In contrast, an alternative implementation in which the objective function compares the output associated with each rotated digit to a corresponding fixed unreferenced reference digit is shown here to discriminate accurately among the rotated digits in latent space even when the dimension of the latent space is 2 or 3.
Leveraging prior knowledge on intraclass variance due to transformations is a powerful method to improve the sample complexity of deep neural networks. This makes them applicable to practically important use-cases where training data is scarce. Rather than being learned, this knowledge can be embedded by enforcing invariance to those transformations. Invariance can be imposed using group-equivariant convolutions followed by a pooling operation. For rotation-invariance, previous work investigated replacing the spatial pooling operation with invariant integration which explicitly constructs invariant representations. Invariant integration uses monomials which are selected using an iterative approach requiring expensive pre-training. We propose a novel monomial selection algorithm based on pruning methods to allow an application to more complex problems. Additionally, we replace monomials with different functions such as weighted sums, multi-layer perceptrons and self-attention, thereby streamlining the training of invariant-integration-based architectures. We demonstrate the improved sample complexity on the Rotated-MNIST, SVHN and CIFAR-10 datasets where rotation-invariant-integration-based Wide-ResNet architectures using monomials and weighted sums outperform the respective baselines in the limited sample regime. We achieve state-of-the-art results using full data on Rotated-MNIST and SVHN where rotation is a main source of intraclass variation. On STL-10 we outperform a standard and a rotation-equivariant convolutional neural network using pooling.
We propose a new \emph{Transformed Risk Minimization} (TRM) framework as an extension of classical risk minimization. In TRM, we optimize not only over predictive models, but also over data transformations; specifically over distributions thereof. As a key application, we focus on learning augmentations; for instance appropriate rotations of images, to improve classification performance with a given class of predictors. Our TRM method (1) jointly learns transformations and models in a \emph{single training loop}, (2) works with any training algorithm applicable to standard risk minimization, and (3) handles any transforms, such as discrete and continuous classes of augmentations. To avoid overfitting when implementing empirical transformed risk minimization, we propose a novel regularizer based on PAC-Bayes theory. For learning augmentations of images, we propose a new parametrization of the space of augmentations via a stochastic composition of blocks of geometric transforms. This leads to the new \emph{Stochastic Compositional Augmentation Learning} (SCALE) algorithm. The performance of TRM with SCALE compares favorably to prior methods on CIFAR10/100. Additionally, we show empirically that SCALE can correctly learn certain symmetries in the data distribution (recovering rotations on rotated MNIST) and can also improve calibration of the learned model.
Group convolutional neural networks (G-CNNs) have been shown to increase parameter efficiency and model accuracy by incorporating geometric inductive biases. In this work, we investigate the properties of representations learned by regular G-CNNs, and show considerable parameter redundancy in group convolution kernels. This finding motivates further weight-tying by sharing convolution kernels over subgroups. To this end, we introduce convolution kernels that are separable over the subgroup and channel dimensions. In order to obtain equivariance to arbitrary affine Lie groups we provide a continuous parameterisation of separable convolution kernels. We evaluate our approach across several vision datasets, and show that our weight sharing leads to improved performance and computational efficiency. In many settings, separable G-CNNs outperform their non-separable counterpart, while only using a fraction of their training time. In addition, thanks to the increase in computational efficiency, we are able to implement G-CNNs equivariant to the $\mathrm{Sim(2)}$ group; the group of dilations, rotations and translations. $\mathrm{Sim(2)}$-equivariance further improves performance on all tasks considered.
Group Convolutional Neural Networks (G-CNNs) constrain learned features to respect the symmetries in the selected group, and lead to better generalization when these symmetries appear in the data. If this is not the case, however, equivariance leads to overly constrained models and worse performance. Frequently, transformations occurring in data can be better represented by a subset of a group than by a group as a whole, e.g., rotations in $[-90^{\circ}, 90^{\circ}]$. In such cases, a model that respects equivariance $\textit{partially}$ is better suited to represent the data. In addition, relevant transformations may differ for low and high-level features. For instance, full rotation equivariance is useful to describe edge orientations in a face, but partial rotation equivariance is better suited to describe face poses relative to the camera. In other words, the optimal level of equivariance may differ per layer. In this work, we introduce $\textit{Partial G-CNNs}$: G-CNNs able to learn layer-wise levels of partial and full equivariance to discrete, continuous groups and combinations thereof as part of training. Partial G-CNNs retain full equivariance when beneficial, e.g., for rotated MNIST, but adjust it whenever it becomes harmful, e.g., for classification of 6 / 9 digits or natural images. We empirically show that partial G-CNNs pair G-CNNs when full equivariance is advantageous, and outperform them otherwise.
Training a Convolutional Neural Network (CNN) to be robust against rotation has mostly been done with data augmentation. In this paper, another progressive vision of research direction is highlighted to encourage less dependence on data augmentation by achieving structural rotational invariance of a network. The deep equivariance-bridged SO(2) invariant network is proposed to echo such vision. First, Self-Weighted Nearest Neighbors Graph Convolutional Network (SWN-GCN) is proposed to implement Graph Convolutional Network (GCN) on the graph representation of an image to acquire rotationally equivariant representation, as GCN is more suitable for constructing deeper network than spectral graph convolution-based approaches. Then, invariant representation is eventually obtained with Global Average Pooling (GAP), a permutation-invariant operation suitable for aggregating high-dimensional representations, over the equivariant set of vertices retrieved from SWN-GCN. Our method achieves the state-of-the-art image classification performance on rotated MNIST and CIFAR-10 images, where the models are trained with a non-augmented dataset only. Quantitative validations over invariance of the representations also demonstrate strong invariance of deep representations of SWN-GCN over rotations.
Cryo-Electron Microscopy (Cryo-EM) is a Nobel prize-winning technology for determining the 3D structure of particles at near-atomic resolution. A fundamental step in the recovering of the 3D single-particle structure is to align its 2D projections; thus, the construction of a canonical representation with a fixed rotation angle is required. Most approaches use discrete clustering which fails to capture the continuous nature of image rotation, others suffer from low-quality image reconstruction. We propose a novel method that leverages the recent development in the generative adversarial networks. We introduce an encoder-decoder with a rotation angle classifier. In addition, we utilize a discriminator on the decoder output to minimize the reconstruction error. We demonstrate our approach with the Cryo-EM 5HDB and the rotated MNIST datasets showing substantial improvement over recent methods.