Simon Lucey

Runze Xu, Arpit Garg, Hemanth Saratchandran et al.

Low-Rank Adaptation (LoRA) has become one of the most widely used fine-tuning mechanisms for adapting large language models to new domains, tasks, and users. Yet adaptation performance alone can obscure an important failure mode: LoRA updates may improve performance on the target distribution while degrading prior capabilities learned during pretraining and alignment. We show that this forgetting becomes especially severe when the adaptation distribution differs substantially from the models original training or alignment distributions. The challenge is amplified in practical settings, where the original training and alignment data are typically unavailable. Motivated by this constraint, we study how LoRA based adaptation balances new learning against forgetting in a replay-free setting, and introduce a simple output space regularizer that can be added directly to existing training pipelines. Our method removes the ground-truth token from both the base and adapted model distributions, renormalizes the remaining probabilities, and applies KL regularization only over the non-target vocabulary. This preserves the base models relative preferences among alternative tokens without directly opposing the cross-entropy signal required for adaptation. As the regularizer acts only at the loss level, it requires no replay data, architectural changes, adapter redesign, or inference-time overhead, and can be applied directly to existing LoRA variants. Across all LoRA variants tested and across various backbones, our method improves the frontier between new learning and forgetting when the adaptation distribution differs substantially from the base models original training or alignment distributions, suggesting a broadly applicable route toward more reliable LLM updating.

Arpit Garg, Simon Lucey, Hemanth Saratchandran

Low-Rank Adaptation (LoRA) is a popular parameter-efficient fine-tuning (PEFT) method that restricts weight updates to low-rank adapters, introducing a fixed low-rank inductive bias by optimizing in a low-dimensional subspace. In this work, we question whether a fixed-rank constraint is the most effective inductive bias for parameter-efficient fine-tuning. We introduce *Learnable Rank LoRA (LR-LoRA)*, a PEFT method in which the adapter rank is learned during the training process. Instead of prescribing a uniform rank for all adapter layers, LR-LoRA allows the optimizer to determine the appropriate rank for each layer. Using this approach, we find substantial layer-wise variation in the learned ranks, with the attention and MLP layers in the transformer models exhibiting systematically different rank preferences. Across a range of language understanding and commonsense reasoning benchmarks, LR-LoRA achieves state-of-the-art performance in most settings and consistently outperforms strong PEFT baselines, demonstrating that a learnable rank provides a more flexible and effective inductive bias than fixed-rank adaptations.

Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah et al.

Despite Neural Radiance Fields (NeRF) showing compelling results in photorealistic novel views synthesis of real-world scenes, most existing approaches require accurate prior camera poses. Although approaches for jointly recovering the radiance field and camera pose exist (BARF), they rely on a cumbersome coarse-to-fine auxiliary positional embedding to ensure good performance. We present Gaussian Activated neural Radiance Fields (GARF), a new positional embedding-free neural radiance field architecture - employing Gaussian activations - that outperforms the current state-of-the-art in terms of high fidelity reconstruction and pose estimation.

Sameera Ramasinghe, Hemanth Saratchandran, Violetta Shevchenko et al. · amazon-science

Modelling dynamical systems is an integral component for understanding the natural world. To this end, neural networks are becoming an increasingly popular candidate owing to their ability to learn complex functions from large amounts of data. Despite this recent progress, there has not been an adequate discussion on the architectural regularization that neural networks offer when learning such systems, hindering their efficient usage. In this paper, we initiate a discussion in this direction using coordinate networks as a test bed. We interpret dynamical systems and coordinate networks from a signal processing lens, and show that simple coordinate networks with few layers can be used to solve multiple problems in modelling dynamical systems, without any explicit regularizers.

Nathaniel Chodosh, Deva Ramanan, Simon Lucey

Popular benchmarks for self-supervised LiDAR scene flow (stereoKITTI, and FlyingThings3D) have unrealistic rates of dynamic motion, unrealistic correspondences, and unrealistic sampling patterns. As a result, progress on these benchmarks is misleading and may cause researchers to focus on the wrong problems. We evaluate a suite of top methods on a suite of real-world datasets (Argoverse 2.0, Waymo, and NuScenes) and report several conclusions. First, we find that performance on stereoKITTI is negatively correlated with performance on real-world data. Second, we find that one of this task's key components -- removing the dominant ego-motion -- is better solved by classic ICP than any tested method. Finally, we show that despite the emphasis placed on learning, most performance gains are caused by pre- and post-processing steps: piecewise-rigid refinement and ground removal. We demonstrate this through a baseline method that combines these processing steps with a learning-free test-time flow optimization. This baseline outperforms every evaluated method.

Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A. Jeni et al.

In this paper we present a new method for deformable NeRF that can directly use optical flow as supervision. We overcome the major challenge with respect to the computationally inefficiency of enforcing the flow constraints to the backward deformation field, used by deformable NeRFs. Specifically, we show that inverting the backward deformation function is actually not needed for computing scene flows between frames. This insight dramatically simplifies the problem, as one is no longer constrained to deformation functions that can be analytically inverted. Instead, thanks to the weak assumptions required by our derivation based on the inverse function theorem, our approach can be extended to a broad class of commonly used backward deformation field. We present results on monocular novel view synthesis with rapid object motion, and demonstrate significant improvements over baselines without flow supervision.

Xueqian Li, Jianqiao Zheng, Francesco Ferroni et al.

Neural Scene Flow Prior (NSFP) is of significant interest to the vision community due to its inherent robustness to out-of-distribution (OOD) effects and its ability to deal with dense lidar points. The approach utilizes a coordinate neural network to estimate scene flow at runtime, without any training. However, it is up to 100 times slower than current state-of-the-art learning methods. In other applications such as image, video, and radiance function reconstruction innovations in speeding up the runtime performance of coordinate networks have centered upon architectural changes. In this paper, we demonstrate that scene flow is different -- with the dominant computational bottleneck stemming from the loss function itself (i.e., Chamfer distance). Further, we rediscover the distance transform (DT) as an efficient, correspondence-free loss function that dramatically speeds up the runtime optimization. Our fast neural scene flow (FNSF) approach reports for the first time real-time performance comparable to learning methods, without any training or OOD bias on two of the largest open autonomous driving (AV) lidar datasets Waymo Open and Argoverse.

Cameron Gordon, Shin-Fang Chng, Lachlan MacDonald et al.

The role of quantization within implicit/coordinate neural networks is still not fully understood. We note that using a canonical fixed quantization scheme during training produces poor performance at low-rates due to the network weight distributions changing over the course of training. In this work, we show that a non-uniform quantization of neural weights can lead to significant improvements. Specifically, we demonstrate that a clustered quantization enables improved reconstruction. Finally, by characterising a trade-off between quantization and network capacity, we demonstrate that it is possible (while memory inefficient) to reconstruct signals using binary neural networks. We demonstrate our findings experimentally on 2D image reconstruction and 3D radiance fields; and show that simple quantization methods and architecture search can achieve compression of NeRF to less than 16kb with minimal loss in performance (323x smaller than the original NeRF).

Jianqiao Zheng, Sameera Ramasinghe, Xueqian Li et al.

It is well noted that coordinate-based MLPs benefit -- in terms of preserving high-frequency information -- through the encoding of coordinate positions as an array of Fourier features. Hitherto, the rationale for the effectiveness of these positional encodings has been mainly studied through a Fourier lens. In this paper, we strive to broaden this understanding by showing that alternative non-Fourier embedding functions can indeed be used for positional encoding. Moreover, we show that their performance is entirely determined by a trade-off between the stable rank of the embedded matrix and the distance preservation between embedded coordinates. We further establish that the now ubiquitous Fourier feature mapping of position is a special case that fulfills these conditions. Consequently, we present a more general theory to analyze positional encoding in terms of shifted basis functions. In addition, we argue that employing a more complex positional encoding -- that scales exponentially with the number of modes -- requires only a linear (rather than deep) coordinate function to achieve comparable performance. Counter-intuitively, we demonstrate that trading positional embedding complexity for network deepness is orders of magnitude faster than current state-of-the-art; despite the additional embedding complexity. To this end, we develop the necessary theoretical formulae and empirically verify that our theoretical claims hold in practice.

Kavisha Vidanapathirana, Shin-Fang Chng, Xueqian Li et al.

The test-time optimization of scene flow - using a coordinate network as a neural prior - has gained popularity due to its simplicity, lack of dataset bias, and state-of-the-art performance. We observe, however, that although coordinate networks capture general motions by implicitly regularizing the scene flow predictions to be spatially smooth, the neural prior by itself is unable to identify the underlying multi-body rigid motions present in real-world data. To address this, we show that multi-body rigidity can be achieved without the cumbersome and brittle strategy of constraining the $SE(3)$ parameters of each rigid body as done in previous works. This is achieved by regularizing the scene flow optimization to encourage isometry in flow predictions for rigid bodies. This strategy enables multi-body rigidity in scene flow while maintaining a continuous flow field, hence allowing dense long-term scene flow integration across a sequence of point clouds. We conduct extensive experiments on real-world datasets and demonstrate that our approach outperforms the state-of-the-art in 3D scene flow and long-term point-wise 4D trajectory prediction. The code is available at: https://github.com/kavisha725/MBNSF.

Sameera Ramasinghe, Lachlan MacDonald, Moshiur Farazi et al.

Characterizing the remarkable generalization properties of over-parameterized neural networks remains an open problem. In this paper, we promote a shift of focus towards initialization rather than neural architecture or (stochastic) gradient descent to explain this implicit regularization. Through a Fourier lens, we derive a general result for the spectral bias of neural networks and show that the generalization of neural networks is heavily tied to their initialization. Further, we empirically solidify the developed theoretical insights using practical, deep networks. Finally, we make a case against the controversial flat-minima conjecture and show that Fourier analysis grants a more reliable framework for understanding the generalization of neural networks.

Ming-Fang Chang, Yipu Zhao, Rajvi Shah et al.

We address the problem of map sparsification for long-term visual localization. For map sparsification, a commonly employed assumption is that the pre-build map and the later captured localization query are consistent. However, this assumption can be easily violated in the dynamic world. Additionally, the map size grows as new data accumulate through time, causing large data overhead in the long term. In this paper, we aim to overcome the environmental changes and reduce the map size at the same time by selecting points that are valuable to future localization. Inspired by the recent progress in Graph Neural Network(GNN), we propose the first work that models SfM maps as heterogeneous graphs and predicts 3D point importance scores with a GNN, which enables us to directly exploit the rich information in the SfM map graph. Two novel supervisions are proposed: 1) a data-fitting term for selecting valuable points to future localization based on training queries; 2) a K-Cover term for selecting sparse points with full map coverage. The experiments show that our method selected map points on stable and widely visible structures and outperformed baselines in localization performance.

Lachlan Ewen MacDonald, Jack Valmadre, Hemanth Saratchandran et al.

We introduce a general theoretical framework, designed for the study of gradient optimisation of deep neural networks, that encompasses ubiquitous architecture choices including batch normalisation, weight normalisation and skip connections. Our framework determines the curvature and regularity properties of multilayer loss landscapes in terms of their constituent layers, thereby elucidating the roles played by normalisation layers and skip connections in globalising these properties. We then demonstrate the utility of this framework in two respects. First, we give the only proof of which we are aware that a class of deep neural networks can be trained using gradient descent to global optima even when such optima only exist at infinity, as is the case for the cross-entropy cost. Second, we identify a novel causal mechanism by which skip connections accelerate training, which we verify predictively with ResNets on MNIST, CIFAR10, CIFAR100 and ImageNet.

Mosam Dabhi, Chaoyang Wang, Tim Clifford et al.

Labeling articulated objects in unconstrained settings have a wide variety of applications including entertainment, neuroscience, psychology, ethology, and many fields of medicine. Large offline labeled datasets do not exist for all but the most common articulated object categories (e.g., humans). Hand labeling these landmarks within a video sequence is a laborious task. Learned landmark detectors can help, but can be error-prone when trained from only a few examples. Multi-camera systems that train fine-grained detectors have shown significant promise in detecting such errors, allowing for self-supervised solutions that only need a small percentage of the video sequence to be hand-labeled. The approach, however, is based on calibrated cameras and rigid geometry, making it expensive, difficult to manage, and impractical in real-world scenarios. In this paper, we address these bottlenecks by combining a non-rigid 3D neural prior with deep flow to obtain high-fidelity landmark estimates from videos with only two or three uncalibrated, handheld cameras. With just a few annotations (representing 1-2% of the frames), we are able to produce 2D results comparable to state-of-the-art fully supervised methods, along with 3D reconstructions that are impossible with other existing approaches. Our Multi-view Bootstrapping in the Wild (MBW) approach demonstrates impressive results on standard human datasets, as well as tigers, cheetahs, fish, colobus monkeys, chimpanzees, and flamingos from videos captured casually in a zoo. We release the codebase for MBW as well as this challenging zoo dataset consisting image frames of tail-end distribution categories with their corresponding 2D, 3D labels generated from minimal human intervention.

Yiping Ji, Mahalakshmi Sabanayagam, Peyman Moghadam et al.

Large-scale transformer training and deployment are increasingly constrained by the transfer of activations, gradients, and optimizer states across accelerators. Low-bit quantization offers a natural remedy, but transformer activations are often heavy-tailed and outlier-dominated, making simple quantization highly lossy. We show that this difficulty is not only a property of the quantizer, but also of the architecture. Specifically, residual connections can drive transformer activations away from Gaussianity during training. Using controlled comparisons between residual and residual-free transformers, we demonstrate that this effect leads to substantially higher quantization error and accuracy degradation at low precision in residual models. We explain the phenomenon through an excess kurtosis analysis, showing that residual mixing can amplify non-Gaussianity, whereas dense mixing in residual-free contracts non-Gaussianity. We then show that residual-free transformers can be made trainable using orthogonal initialization, spectral or second-order optimization, and depth-aware scaling of attention temperature. In language tasks, while there is a small drop in full precision performance, these models retain near-Gaussian activations and exhibit significantly improved robustness to low-bit quantization. Our results identify an accuracy--compressibility trade-off in transformer design and motivate architecture-level approaches to quantization-friendly foundation models.

Shin-Fang Chng, Ravi Garg, Hemanth Saratchandran et al.

This paper tackles the simultaneous optimization of pose and Neural Radiance Fields (NeRF). Departing from the conventional practice of using explicit global representations for camera pose, we propose a novel overparameterized representation that models camera poses as learnable rigid warp functions. We establish that modeling the rigid warps must be tightly coupled with constraints and regularization imposed. Specifically, we highlight the critical importance of enforcing invertibility when learning rigid warp functions via neural network and propose the use of an Invertible Neural Network (INN) coupled with a geometry-informed constraint for this purpose. We present results on synthetic and real-world datasets, and demonstrate that our approach outperforms existing baselines in terms of pose estimation and high-fidelity reconstruction due to enhanced optimization convergence.

Hemanth Saratchandran, Thomas X. Wang, Simon Lucey

In this article, we introduce a novel normalization technique for neural network weight matrices, which we term weight conditioning. This approach aims to narrow the gap between the smallest and largest singular values of the weight matrices, resulting in better-conditioned matrices. The inspiration for this technique partially derives from numerical linear algebra, where well-conditioned matrices are known to facilitate stronger convergence results for iterative solvers. We provide a theoretical foundation demonstrating that our normalization technique smoothens the loss landscape, thereby enhancing convergence of stochastic gradient descent algorithms. Empirically, we validate our normalization across various neural network architectures, including Convolutional Neural Networks (CNNs), Vision Transformers (ViT), Neural Radiance Fields (NeRF), and 3D shape modeling. Our findings indicate that our normalization method is not only competitive but also outperforms existing weight normalization techniques from the literature.

Vishagar Arunan, Saeedha Nazar, Hashiru Pramuditha et al.

Splatting-based 3D reconstruction methods have gained popularity with the advent of 3D Gaussian Splatting, efficiently synthesizing high-quality novel views. These methods commonly resort to using exponential family functions, such as the Gaussian function, as reconstruction kernels due to their anisotropic nature, ease of projection, and differentiability in rasterization. However, the field remains restricted to variations within the exponential family, leaving generalized reconstruction kernels largely underexplored, partly due to the lack of easy integrability in 3D to 2D projections. In this light, we show that a class of decaying anisotropic radial basis functions (DARBFs), which are non-negative functions of the Mahalanobis distance, supports splatting by approximating the Gaussian function's closed-form integration advantage. With this fresh perspective, we demonstrate up to 34% faster convergence during training and a 45% reduction in memory consumption across various DARB reconstruction kernels, while maintaining comparable PSNR, SSIM, and LPIPS results. We will make the code available.

Jianqiao Zheng, Hemanth Saratchandran, Simon Lucey

Implicit Neural Representations (INRs) have revolutionized continuous signal modeling, yet they struggle to recover fine-grained details within finite training budgets. While empirical techniques, such as positional encoding (PE), sinusoidal activations (SIREN), and batch normalization (BN), effectively mitigate this, their theoretical justifications are predominantly post hoc, focusing on the global NTK spectrum only after modifications are applied. In this work, we reverse this paradigm by introducing a structural diagnostic framework. By performing a layer-wise decomposition of the NTK, we mathematically identify the ``Inlet Rank Collapse'': a phenomenon where the low-dimensional input coordinates fail to span the high-dimensional embedding space, creating a fundamental rank deficiency at the first layer that acts as an expressive bottleneck for the entire network. This framework provides a unified perspective to re-interpret PE, SIREN, and BN as different forms of rank restoration. Guided by this diagnosis, we derive a Rank-Expanding Initialization, a minimalist remedy that ensures the representation rank scales with the layer width without architectural modifications or computational overhead. Our results demonstrate that this principled remedy enables standard MLPs to achieve high-fidelity reconstructions, proving that the key to empowering INRs lies in the structural optimization of the initial rank propagation to effectively populate the latent space.

Jianqiao Zheng, Sameera Ramasinghe, Simon Lucey

It is well noted that coordinate based MLPs benefit -- in terms of preserving high-frequency information -- through the encoding of coordinate positions as an array of Fourier features. Hitherto, the rationale for the effectiveness of these positional encodings has been solely studied through a Fourier lens. In this paper, we strive to broaden this understanding by showing that alternative non-Fourier embedding functions can indeed be used for positional encoding. Moreover, we show that their performance is entirely determined by a trade-off between the stable rank of the embedded matrix and the distance preservation between embedded coordinates. We further establish that the now ubiquitous Fourier feature mapping of position is a special case that fulfills these conditions. Consequently, we present a more general theory to analyze positional encoding in terms of shifted basis functions. To this end, we develop the necessary theoretical formulae and empirically verify that our theoretical claims hold in practice. Codes available at https://github.com/osiriszjq/Rethinking-positional-encoding.

Vinit Sarode, Animesh Dhagat, Rangaprasad Arun Srivatsan et al.

Point clouds have grown in importance in the way computers perceive the world. From LIDAR sensors in autonomous cars and drones to the time of flight and stereo vision systems in our phones, point clouds are everywhere. Despite their ubiquity, point clouds in the real world are often missing points because of sensor limitations or occlusions, or contain extraneous points from sensor noise or artifacts. These problems challenge algorithms that require computing correspondences between a pair of point clouds. Therefore, this paper presents a fully-convolutional neural network that identifies which points in one point cloud are most similar (inliers) to the points in another. We show improvements in learning-based and classical point cloud registration approaches when retrofitted with our network. We demonstrate these improvements on synthetic and real-world datasets. Finally, our network produces impressive results on test datasets that were unseen during training, thus exhibiting generalizability. Code and videos are available at https://github.com/vinits5/masknet

Vinit Sarode, Xueqian Li, Hunter Goforth et al.

PointNet has recently emerged as a popular representation for unstructured point cloud data, allowing application of deep learning to tasks such as object detection, segmentation and shape completion. However, recent works in literature have shown the sensitivity of the PointNet representation to pose misalignment. This paper presents a novel framework that uses PointNet encoding to align point clouds and perform registration for applications such as 3D reconstruction, tracking and pose estimation. We develop a framework that compares PointNet features of template and source point clouds to find the transformation that aligns them accurately. In doing so, we avoid computationally expensive correspondence finding steps, that are central to popular registration methods such as ICP and its variants. Depending on the prior information about the shape of the object formed by the point clouds, our framework can produce approaches that are shape specific or general to unseen shapes. Our framework produces approaches that are robust to noise and initial misalignment in data and work robustly with sparse as well as partial point clouds. We perform extensive simulation and real-world experiments to validate the efficacy of our approach and compare the performance with state-of-art approaches. Code is available at https://github.com/vinits5/pointnet-registrationframework.

Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan et al.

PointNet has revolutionized how we think about representing point clouds. For classification and segmentation tasks, the approach and its subsequent extensions are state-of-the-art. To date, the successful application of PointNet to point cloud registration has remained elusive. In this paper we argue that PointNet itself can be thought of as a learnable "imaging" function. As a consequence, classical vision algorithms for image alignment can be applied on the problem - namely the Lucas & Kanade (LK) algorithm. Our central innovations stem from: (i) how to modify the LK algorithm to accommodate the PointNet imaging function, and (ii) unrolling PointNet and the LK algorithm into a single trainable recurrent deep neural network. We describe the architecture, and compare its performance against state-of-the-art in common registration scenarios. The architecture offers some remarkable properties including: generalization across shape categories and computational efficiency - opening up new paths of exploration for the application of deep learning to point cloud registration. Code and videos are available at https://github.com/hmgoforth/PointNetLK.

Mehdi Hosseinzadeh, Shin-Fang Chng, Yi Xu et al.

3D Gaussians have recently emerged as an effective scene representation for real-time splatting and accurate novel-view synthesis, motivating several works to adapt multi-view structure prediction networks to regress per-pixel 3D Gaussians from images. However, most prior work extends these networks to predict additional Gaussian parameters -- orientation, scale, opacity, and appearance -- while relying almost exclusively on view-synthesis supervision. We show that a view-synthesis loss alone is insufficient to recover geometrically meaningful splats in this setting. We analyze and address the ambiguities of learning 3D Gaussian splats under self-supervision for pose-free generalizable splatting, and introduce G3Splat, which enforces geometric priors to obtain geometrically consistent 3D scene representations. Trained on RE10K, our approach achieves state-of-the-art performance in (i) geometrically consistent reconstruction, (ii) relative pose estimation, and (iii) novel-view synthesis. We further demonstrate strong zero-shot generalization on ScanNet, substantially outperforming prior work in both geometry recovery and relative pose estimation. Code and pretrained models are released on our project page (https://m80hz.github.io/g3splat/).

Hemanth Saratchandran, Sameera Ramasinghe, Violetta Shevchenko et al. · amazon-science

Implicit Neural Representations (INRs) have gained popularity for encoding signals as compact, differentiable entities. While commonly using techniques like Fourier positional encodings or non-traditional activation functions (e.g., Gaussian, sinusoid, or wavelets) to capture high-frequency content, their properties lack exploration within a unified theoretical framework. Addressing this gap, we conduct a comprehensive analysis of these activations from a sampling theory perspective. Our investigation reveals that sinc activations, previously unused in conjunction with INRs, are theoretically optimal for signal encoding. Additionally, we establish a connection between dynamical systems and INRs, leveraging sampling theory to bridge these two paradigms.

Yiping Ji, Hemanth Saratchandran, Cameron Gordon et al.

Low-rank decomposition has emerged as a vital tool for enhancing parameter efficiency in neural network architectures, gaining traction across diverse applications in machine learning. These techniques significantly lower the number of parameters, striking a balance between compactness and performance. However, a common challenge has been the compromise between parameter efficiency and the accuracy of the model, where reduced parameters often lead to diminished accuracy compared to their full-rank counterparts. In this work, we propose a novel theoretical framework that integrates a sinusoidal function within the low-rank decomposition process. This approach not only preserves the benefits of the parameter efficiency characteristic of low-rank methods but also increases the decomposition's rank, thereby enhancing model performance. Our method proves to be a plug in enhancement for existing low-rank models, as evidenced by its successful application in Vision Transformers (ViT), Large Language Models (LLMs), Neural Radiance Fields (NeRF) and 3D shape modelling.

Jenny Seidenschwarz, Aljoša Ošep, Francesco Ferroni et al.

We tackle semi-supervised object detection based on motion cues. Recent results suggest that heuristic-based clustering methods in conjunction with object trackers can be used to pseudo-label instances of moving objects and use these as supervisory signals to train 3D object detectors in Lidar data without manual supervision. We re-think this approach and suggest that both, object detection, as well as motion-inspired pseudo-labeling, can be tackled in a data-driven manner. We leverage recent advances in scene flow estimation to obtain point trajectories from which we extract long-term, class-agnostic motion patterns. Revisiting correlation clustering in the context of message passing networks, we learn to group those motion patterns to cluster points to object instances. By estimating the full extent of the objects, we obtain per-scan 3D bounding boxes that we use to supervise a Lidar object detection network. Our method not only outperforms prior heuristic-based approaches (57.5 AP, +14 improvement over prior work), more importantly, we show we can pseudo-label and train object detectors across datasets.

Mosam Dabhi, Laszlo A. Jeni, Simon Lucey

The lifting of 3D structure and camera from 2D landmarks is at the cornerstone of the entire discipline of computer vision. Traditional methods have been confined to specific rigid objects, such as those in Perspective-n-Point (PnP) problems, but deep learning has expanded our capability to reconstruct a wide range of object classes (e.g. C3DPO and PAUL) with resilience to noise, occlusions, and perspective distortions. All these techniques, however, have been limited by the fundamental need to establish correspondences across the 3D training data -- significantly limiting their utility to applications where one has an abundance of "in-correspondence" 3D data. Our approach harnesses the inherent permutation equivariance of transformers to manage varying number of points per 3D data instance, withstands occlusions, and generalizes to unseen categories. We demonstrate state of the art performance across 2D-3D lifting task benchmarks. Since our approach can be trained across such a broad class of structures we refer to it simply as a 3D Lifting Foundation Model (3D-LFM) -- the first of its kind.

Hemanth Saratchandran, Sameera Ramasinghe, Simon Lucey

In the realm of computer vision, Neural Fields have gained prominence as a contemporary tool harnessing neural networks for signal representation. Despite the remarkable progress in adapting these networks to solve a variety of problems, the field still lacks a comprehensive theoretical framework. This article aims to address this gap by delving into the intricate interplay between initialization and activation, providing a foundational basis for the robust optimization of Neural Fields. Our theoretical insights reveal a deep-seated connection among network initialization, architectural choices, and the optimization process, emphasizing the need for a holistic approach when designing cutting-edge Neural Fields.

Yiping Ji, Hemanth Saratchandran, Peyman Moghadam et al.

We highlight a curious empirical result within modern Vision Transformers (ViTs). Specifically, self-attention catastrophically fails to train unless it is used in conjunction with a skip connection. This is in contrast to other elements of a ViT that continue to exhibit good performance (albeit suboptimal) when skip connections are removed. Further, we show that this critical dependence on skip connections is a relatively new phenomenon, with previous deep architectures (\eg, CNNs) exhibiting good performance in their absence. In this paper, we theoretically characterize that the self-attention mechanism is fundamentally ill-conditioned and is, therefore, uniquely dependent on skip connections for regularization. Additionally, we propose Token Graying -- a simple yet effective complement (to skip connections) that further improves the condition of input tokens. We validate our approach in both supervised and self-supervised training methods.

George Cazenavette, Joel Julin, Simon Lucey

Until quite recently, the backbone of nearly every state-of-the-art computer vision model has been the 2D convolution. At its core, a 2D convolution simultaneously mixes information across both the spatial and channel dimensions of a representation. Many recent computer vision architectures consist of sequences of isotropic blocks that disentangle the spatial and channel-mixing components. This separation of the operations allows us to more closely juxtapose the effects of spatial and channel mixing in deep learning. In this paper, we take an initial step towards garnering a deeper understanding of the roles of these mixing operations. Through our experiments and analysis, we discover that on both classical (ResNet) and cutting-edge (ConvMixer) models, we can reach nearly the same level of classification performance by and leaving the spatial mixers at their random initializations. Furthermore, we show that models with random, fixed spatial mixing are naturally more robust to adversarial perturbations. Lastly, we show that this phenomenon extends past the classification regime, as such models can also decode pixel-shuffled images.

Jim James, Ben Wilson, Simon Lucey et al.

In this work, we introduce the 3D Gaussian Point Encoder, an explicit per-point embedding built on mixtures of learned 3D Gaussians. This explicit geometric representation for 3D recognition tasks is a departure from widely used implicit representations such as PointNet. However, it is difficult to learn 3D Gaussian encoders in end-to-end fashion with standard optimizers. We develop optimization techniques based on natural gradients and distillation from PointNets to find a Gaussian Basis that can reconstruct PointNet activations. The resulting 3D Gaussian Point Encoders are faster and more parameter efficient than traditional PointNets. As in the 3D reconstruction literature where there has been considerable interest in the move from implicit (e.g., NeRF) to explicit (e.g., Gaussian Splatting) representations, we can take advantage of computational geometry heuristics to accelerate 3D Gaussian Point Encoders further. We extend filtering techniques from 3D Gaussian Splatting to construct encoders that run 2.7 times faster as a comparable accuracy PointNet while using 46% less memory and 88% fewer FLOPs. Furthermore, we demonstrate the effectiveness of 3D Gaussian Point Encoders as a component in Mamba3D, running 1.27 times faster and achieving a reduction in memory and FLOPs by 42% and 54% respectively. 3D Gaussian Point Encoders are lightweight enough to achieve high framerates on CPU-only devices.

Hemanth Saratchandran, Damien Teney, Simon Lucey

Transformers have reshaped machine learning by utilizing attention mechanisms to capture complex patterns in large datasets, leading to significant improvements in performance. This success has contributed to the belief that "bigger means better", leading to ever-increasing model sizes. This paper challenge this ideology by showing that many existing transformers might be unnecessarily oversized. We discover a theoretical principle that redefines the role of multi-head attention. An important benefit of the multiple heads is in improving the conditioning of the attention block. We exploit this theoretical insight and redesign popular architectures with an increased number of heads. The improvement in the conditioning proves so significant in practice that model depth can be decreased, reducing the parameter count by up to 30-50% while maintaining accuracy. We obtain consistent benefits across a variety of transformer-based architectures of various scales, on tasks in computer vision (ImageNet-1k) as well as language and sequence modeling (GLUE benchmark, TinyStories, and the Long-Range Arena benchmark).

Hemanth Saratchandran, Simon Lucey

We present a theoretical analysis of the Jacobian of an attention block within a transformer, showing that it is governed by the query, key, and value projections that define the attention mechanism. Leveraging this insight, we introduce a method that systematically alters the spectral properties of each attention layer to reduce the Jacobian's condition number, thereby improving the overall conditioning of the attention layers within a transformer network. We empirically show that this improved Jacobian conditioning translates to enhanced performance in practice. Our approach is simple, broadly applicable, and can be easily integrated as a drop-in replacement for a wide range of existing attention mechanisms. We validate its effectiveness across diverse transformer architectures and tasks, demonstrating consistent improvements in performance.

Hemanth Saratchandran, Simon Lucey

Transformers have transformed modern machine learning, driving breakthroughs in computer vision, natural language processing, and robotics. At the core of their success lies the attention mechanism, which enables the modeling of global dependencies among input tokens. However, we reveal that the attention block in transformers suffers from inherent ill-conditioning, which hampers gradient-based optimization and leads to inefficient training. To address this, we develop a theoretical framework that establishes a direct relationship between the conditioning of the attention block and that of the embedded tokenized data. Building on this insight, we introduce conditioned embedded tokens, a method that systematically modifies the embedded tokens to improve the conditioning of the attention mechanism. Our analysis demonstrates that this approach significantly mitigates ill-conditioning, leading to more stable and efficient training. We validate our methodology across various transformer architectures, achieving consistent improvements in image classification, object detection, instance segmentation, and natural language processing, highlighting its broad applicability and effectiveness.

Jianqiao Zheng, Xueqian Li, Simon Lucey

The training of vision transformer (ViT) networks on small-scale datasets poses a significant challenge. By contrast, convolutional neural networks (CNNs) have an architectural inductive bias enabling them to perform well on such problems. In this paper, we argue that the architectural bias inherent to CNNs can be reinterpreted as an initialization bias within ViT. This insight is significant as it empowers ViTs to perform equally well on small-scale problems while maintaining their flexibility for large-scale applications. Our inspiration for this ``structured'' initialization stems from our empirical observation that random impulse filters can achieve comparable performance to learned filters within CNNs. Our approach achieves state-of-the-art performance for data-efficient ViT learning across numerous benchmarks including CIFAR-10, CIFAR-100, and SVHN.

Shin-Fang Chng, Hemanth Saratchandran, Simon Lucey

Neural fields encode continuous multidimensional signals as neural networks, enabling diverse applications in computer vision, robotics, and geometry. While Adam is effective for stochastic optimization, it often requires long training times. To address this, we explore alternative optimization techniques to accelerate training without sacrificing accuracy. Traditional second-order methods like L-BFGS are unsuitable for stochastic settings. We propose a theoretical framework for training neural fields with curvature-aware diagonal preconditioners, demonstrating their effectiveness across tasks such as image reconstruction, shape modeling, and Neural Radiance Fields (NeRF).

Hemanth Saratchandran, Shin-Fang Chng, Simon Lucey

Recently, neural networks utilizing periodic activation functions have been proven to demonstrate superior performance in vision tasks compared to traditional ReLU-activated networks. However, there is still a limited understanding of the underlying reasons for this improved performance. In this paper, we aim to address this gap by providing a theoretical understanding of periodically activated networks through an analysis of their Neural Tangent Kernel (NTK). We derive bounds on the minimum eigenvalue of their NTK in the finite width setting, using a fairly general network architecture which requires only one wide layer that grows at least linearly with the number of data samples. Our findings indicate that periodically activated networks are \textit{notably more well-behaved}, from the NTK perspective, than ReLU activated networks. Additionally, we give an application to the memorization capacity of such networks and verify our theoretical predictions empirically. Our study offers a deeper understanding of the properties of periodically activated neural networks and their potential in the field of deep learning.

Jianqiao Zheng, Xueqian Li, Simon Lucey

Training vision transformer networks on small datasets poses challenges. In contrast, convolutional neural networks (CNNs) can achieve state-of-the-art performance by leveraging their architectural inductive bias. In this paper, we investigate whether this inductive bias can be reinterpreted as an initialization bias within a vision transformer network. Our approach is motivated by the finding that random impulse filters can achieve almost comparable performance to learned filters in CNNs. We introduce a novel initialization strategy for transformer networks that can achieve comparable performance to CNNs on small datasets while preserving its architectural flexibility.

Yiping Ji, James Martens, Jianqiao Zheng et al.

Transformers have achieved remarkable success across a wide range of applications, a feat often attributed to their scalability. Yet training them without skip (residual) connections remains notoriously difficult. While skips stabilize optimization, they also disrupt the hierarchical structure of representations, raising the long-standing question of whether transformers can be trained efficiently without them. In this work, we address this problem by analyzing the Jacobian of a skipless transformer block, showing why skips improve conditioning and revealing that their stabilization benefits can be recovered through a principled initialization strategy. Building on this insight, we introduce the first method that enables stable and efficient training of skipless transformers without altering the standard architecture. We validate our approach on Vision Transformers (ViTs) in both supervised and self-supervised settings, demonstrating that skipless ViTs trained with our initialization overcome the usual optimization barriers, learn richer hierarchical representations, and outperform strong baselines, that incorporate skip connections, on dense prediction benchmarks. These results show that skip connections are not a fundamental requirement for training ViTs and open new avenues for hierarchical representation learning in vision models.

Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji et al.

This paper questions whether the strong performance of softmax attention in transformers stems from producing a probability distribution over inputs. Instead, we argue that softmax's effectiveness lies in its implicit regularization of the Frobenius norm of the attention matrix, which stabilizes training. Motivated by this, we explore alternative activations, specifically polynomials, that achieve a similar regularization effect. Our theoretical analysis shows that certain polynomials can serve as effective substitutes for softmax, achieving strong performance across transformer applications despite violating softmax's typical properties of positivity, normalization, and sparsity. Extensive experiments support these findings, offering a new perspective on attention mechanisms.

Xueqian Li, Simon Lucey

In contrast to current state-of-the-art methods, such as NSFP [25], which employ deep implicit neural functions for modeling scene flow, we present a novel approach that utilizes classical kernel representations. This representation enables our approach to effectively handle dense lidar points while demonstrating exceptional computational efficiency -- compared to recent deep approaches -- achieved through the solution of a linear system. As a runtime optimization-based method, our model exhibits impressive generalizability across various out-of-distribution scenarios, achieving competitive performance on large-scale lidar datasets. We propose a new positional encoding-based kernel that demonstrates state-of-the-art performance in efficient lidar scene flow estimation on large-scale point clouds. An important highlight of our method is its near real-time performance (~150-170 ms) with dense lidar data (~8k-144k points), enabling a variety of practical applications in robotics and autonomous driving scenarios.

Hemanth Saratchandran, Shin-Fang Chng, Simon Lucey

Physics-informed neural networks (PINNs) offer a promising avenue for tackling both forward and inverse problems in partial differential equations (PDEs) by incorporating deep learning with fundamental physics principles. Despite their remarkable empirical success, PINNs have garnered a reputation for their notorious training challenges across a spectrum of PDEs. In this work, we delve into the intricacies of PINN optimization from a neural architecture perspective. Leveraging the Neural Tangent Kernel (NTK), our study reveals that Gaussian activations surpass several alternate activations when it comes to effectively training PINNs. Building on insights from numerical linear algebra, we introduce a preconditioned neural architecture, showcasing how such tailored architectures enhance the optimization process. Our theoretical findings are substantiated through rigorous validation against established PDEs within the scientific literature.

Harrison Bagley, Will Meakin, Simon Lucey et al.

Physical adversarial attacks on deep learning systems is concerning due to the ease of deploying such attacks, usually by placing an adversarial patch in a scene to manipulate the outcomes of a deep learning model. Training such patches typically requires regularization that improves physical realizability (e.g., printability, smoothness) and/or robustness to real-world variability (e.g. deformations, viewing angle, noise). One type of variability that has received little attention is scale variability. When a patch is rescaled, either digitally through downsampling/upsampling or physically through changing imaging distances, interpolation-induced color mixing occurs. This smooths out pixel values, resulting in a loss of high-frequency patterns and degrading the adversarial signal. To address this, we present a novel superpixel-based regularization method that guides patch optimization to scale-resilient structures. Our ap proach employs the Simple Linear Iterative Clustering (SLIC) algorithm to dynamically cluster pixels in an adversarial patch during optimization. The Implicit Function Theorem is used to backpropagate gradients through SLIC to update the superpixel boundaries and color. This produces patches that maintain their structure over scale and are less susceptible to interpolation losses. Our method achieves greater performance in the digital domain, and when realized physically, these performance gains are preserved, leading to improved physical performance. Real-world performance was objectively assessed using a novel physical evaluation protocol that utilizes screens and cardboard cut-outs to systematically vary real-world conditions.

Arpit Garg, Hemanth Saratchandran, Simon Lucey

Multimodal Large Language Models (MLLMs) increasingly need to forget specific knowledge such as unsafe or private information without requiring full retraining. However, existing unlearning methods often disrupt vision language alignment, causing models to reject both harmful and benign queries. We trace this failure to the projector network during unlearning, its Jacobian becomes severely illconditioned, leading to unstable optimization and drift in cross modal embeddings. We introduce SineProject, a simple method that augments the frozen projector with sinusoidally modulated trainable parameters, improving the Jacobian's spectral conditioning and stabilizing alignment throughout unlearning. Across standard safety and privacy unlearning benchmarks using LLaVA v1.5 7B and 13B, SineProject reduces benign query refusals while achieving complete forgetting of targeted information, yielding state of the art forget retain trade offs with negligible computational overhead.

Jianqiao Zheng, Cameron Gordon, Yiping Ji et al.

Task-agnostic tabular foundation models such as TabPFN have achieved impressive performance on tabular learning tasks, yet the origins of their inductive biases remain poorly understood. In this work, we study TabPFN through the lens of signal reconstruction and provide the first frequency-based analysis of its in-context learning behavior. We show that TabPFN possesses a broader effective frequency capacity than standard ReLU-MLPs, even without hyperparameter tuning. Moreover, unlike MLPs whose spectra evolve primarily over training epochs, we find that TabPFN's spectral capacity adapts directly to the number of samples provided in-context, a phenomenon we term Spectral Adaptivity. We further demonstrate that positional encoding modulates TabPFN's frequency response, mirroring classical results in implicit neural representations. Finally, we show that these properties enable TabPFN to perform training-free and hyperparameter-free image denoising, illustrating its potential as a task-agnostic implicit model. Our analysis provides new insight into the structure and inductive biases of tabular foundation models and highlights their promise for broader signal reconstruction tasks.

Arpit Garg, Hemanth Saratchandran, Ravi Garg et al.

Machine unlearning in large language models (LLMs) is essential for privacy and safety; however, existing approaches remain unstable and unreliable. A widely used strategy, the gradient difference method, applies gradient descent on retained data while performing gradient ascent on forget data, the data whose influence should be removed. However, when combined with cross-entropy loss, this procedure causes unbounded growth of weights and gradients, leading to training instability and degrading both forgetting and retention. We provide a theoretical framework that explains this failure, explicitly showing how ascent on the forget set destabilizes optimization in the feedforward MLP layers of LLMs. Guided by this insight, we propose Bounded Parameter-Efficient Unlearning, a parameter-efficient approach that stabilizes LoRA-based fine-tuning by applying bounded functions to MLP adapters. This simple modification controls the weight dynamics during ascent, enabling the gradient difference method to converge reliably. Across the TOFU, TDEC, and MUSE benchmarks, and across architectures and scales from 125M to 8B parameters, our method achieves substantial improvements in forgetting while preserving retention, establishing a novel theoretically grounded and practically scalable framework for unlearning in LLMs.

Cameron Gordon, Yiping Ji, Hemanth Saratchandran et al.

Resource-constrained weight deployment is a task of immense practical importance. Recently, there has been interest in the specific task of \textit{Delta Compression}, where parties each hold a common base model and only communicate compressed weight updates. However, popular parameter efficient updates such as Low Rank Adaptation (LoRA) face inherent representation limitations - which are especially pronounced when combined with aggressive quantization. To overcome this, we build on recent work that improves LoRA representation capacity by using fixed-frequency sinusoidal functions to increase stable rank without adding additional parameters. We extend this to the quantized setting and present the first theoretical analysis showing how stable rank evolves under quantization. From this, we introduce SineLoRA$Δ$, a principled and effective method for delta compression that improves the expressivity of quantized low-rank adapters by applying a sinusoidal activation. We validate SineLoRA$Δ$ across a diverse variety of domains - including language modeling, vision-language tasks, and text-to-image generation - achieving up to 66% memory reduction with similar performance. We additionally provide a novel application of the canonical Bjøntegaard Delta metric to consistently compare adapter compression changes across the rate-distortion curve.

Simon Lucey

We generalize the connection between activation function and spline regression/smoothing and characterize how this choice may influence spectral bias within a 1D shallow network. We then demonstrate how gradient descent (GD) can be reinterpreted as a shrinkage operator that masks the singular values of a neural network's Jacobian. Viewed this way, GD implicitly selects the number of frequency components to retain, thereby controlling the spectral bias. An explicit relationship is proposed between the choice of GD hyperparameters (learning rate & number of iterations) and bandwidth (the number of active components). GD regularization is shown to be effective only with monotonic activation functions. Finally, we highlight the utility of non-monotonic activation functions (sinc, Gaussian) as iteration-efficient surrogates for spectral bias.

Christopher Fusco, Shin-Fang Ch'ng, Mosam Dabhi et al.

We present a spatio-temporal perspective on category-agnostic 3D lifting of 2D keypoints over a temporal sequence. Our approach differs from existing state-of-the-art methods that are either: (i) object-agnostic, but can only operate on individual frames, or (ii) can model space-time dependencies, but are only designed to work with a single object category. Our approach is grounded in two core principles. First, general information about similar objects can be leveraged to achieve better performance when there is little object-specific training data. Second, a temporally-proximate context window is advantageous for achieving consistency throughout a sequence. These two principles allow us to outperform current state-of-the-art methods on per-frame and per-sequence metrics for a variety of animal categories. Lastly, we release a new synthetic dataset containing 3D skeletons and motion sequences for a variety of animal categories.