Thomas Walker

Thomas Walker, Ahmed Imtiaz Humayun, Randall Balestriero et al.

Identifying and understanding the features that a deep network (DN) extracts from its inputs to produce its outputs is a focal point of interpretability research. The Linear Representation Hypothesis (LRH) identifies features in terms of the linear directions formed by the inputs in a DN's latent space. However, the LRH is limited as it abstracts away from individual components (e.g., neurons and layers), is susceptible to identifying spurious features, and cannot be applied across sub-components (e.g., multiple layers). In this paper, we introduce the Linear Centroids Hypothesis (LCH) as a new framework for identifying the features of a DN. The LCH posits that features correspond to linear directions of centroids, which are vector summarizations of the functional behavior of a DN in a local region of its input space. Interpretability studies under the LCH can leverage existing LRH tools, such as sparse autoencoders, by applying them to the DN's centroids rather than to its latent activations. We demonstrate that doing so yields sparser feature dictionaries for DINO vision transformers, which also perform better on downstream tasks. The LCH also inspires novel approaches to interpretability; for example, LCH can readily identify circuits in GPT2-Large. For code to study the LCH https://github.com/ThomasWalker1/LinearCentroidsHypothesis .

Thomas Walker, Octave Mariotti, Amir Vaxman et al.

We introduce Explicit Neural Surfaces (ENS), an efficient smooth surface representation that directly encodes topology with a deformation field from a known base domain. We apply this representation to reconstruct explicit surfaces from multiple views, where we use a series of neural deformation fields to progressively transform the base domain into a target shape. By using meshes as discrete surface proxies, we train the deformation fields through efficient differentiable rasterization. Using a fixed base domain allows us to have Laplace-Beltrami eigenfunctions as an intrinsic positional encoding alongside standard extrinsic Fourier features, with which our approach can capture fine surface details. Compared to implicit surfaces, ENS trains faster and has several orders of magnitude faster inference times. The explicit nature of our approach also allows higher-quality mesh extraction whilst maintaining competitive surface reconstruction performance and real-time capabilities.

Thomas Walker, Alessio Lomuscio

Probably Approximately Correct (PAC) bounds are widely used to derive probabilistic guarantees for the generalisation of machine learning models. They highlight the components of the model which contribute to its generalisation capacity. However, current state-of-the-art results are loose in approximating the generalisation capacity of deployed machine learning models. Consequently, while PAC bounds are theoretically useful, their applicability for evaluating a model's generalisation property in a given operational design domain is limited. The underlying classical theory is supported by the idea that bounds can be tightened when the number of test points available to the user to evaluate the model increases. Yet, in the case of neural networks, the number of test points required to obtain bounds of interest is often impractical even for small problems. In this paper, we take the novel approach of using the formal verification of neural systems to inform the evaluation of PAC bounds. Rather than using pointwise information obtained from repeated tests, we use verification results on regions around test points. We show that conditioning existing bounds on verification results leads to a tightening proportional to the underlying probability mass of the verified region.

Samuel Garcin, Thomas Walker, Steven McDonagh et al.

Interactive world models continually generate video by responding to a user's actions, enabling open-ended generation capabilities. However, existing models typically lack a 3D representation of the environment, meaning 3D consistency must be implicitly learned from data, and spatial memory is restricted to limited temporal context windows. This results in an unrealistic user experience and presents significant obstacles to down-stream tasks such as training agents. To address this, we present PERSIST, a new paradigm of world model which simulates the evolution of a latent 3D scene: environment, camera, and renderer. This allows us to synthesize new frames with persistent spatial memory and consistent geometry. Both quantitative metrics and a qualitative user study show substantial improvements in spatial memory, 3D consistency, and long-horizon stability over existing methods, enabling coherent, evolving 3D worlds. We further demonstrate novel capabilities, including synthesising diverse 3D environments from a single image, as well as enabling fine-grained, geometry-aware control over generated experiences by supporting environment editing and specification directly in 3D space. Project page: https://francelico.github.io/persist.github.io

Thomas Walker, Ahmed Imtiaz Humayun, Randall Balestriero et al.

A key challenge for the machine learning community is to understand and accelerate the training dynamics of deep networks that lead to delayed generalisation and emergent robustness to input perturbations, also known as grokking. Prior work has associated phenomena like delayed generalisation with the transition of a deep network from a linear to a feature learning regime, and emergent robustness with changes to the network's functional geometry, in particular the arrangement of the so-called linear regions in deep networks employing continuous piecewise affine nonlinearities. Here, we explain how grokking is realised in the Jacobian of a deep network and demonstrate that aligning a network's Jacobians with the training data (in the sense of cosine similarity) ensures grokking under a low-rank Jacobian assumption. Our results provide a strong theoretical motivation for the use of Jacobian regularisation in optimizing deep networks -- a method we introduce as GrokAlign -- which we show empirically to induce grokking much sooner than more conventional regularizers like weight decay. Moreover, we introduce centroid alignment as a tractable and interpretable simplification of Jacobian alignment that effectively identifies and tracks the stages of deep network training dynamics. Accompanying webpage (https://thomaswalker1.github.io/blog/grokalign.html) and code (https://github.com/ThomasWalker1/grokalign).

Thomas Walker, T. Mitchell Roddenberry, Ahmed Imtiaz Humayun et al.

The manifold hypothesis (MH) is often used to explain how machine learning can overcome the curse of dimensionality. However, the MH is only applicable in regimes where the training data provides a sufficiently dense sample of the underlying low-dimensional data manifold, or where such a low-dimensional manifold is conceivably present. We describe the regimes where the MH is not applicable as sparse. In this paper, we demonstrate that models succeed in the sparse regime by exploiting a highly structured local geometry, a property we formalize as normal alignment. We prove that normal-aligned classifiers -- whose input-output Jacobians are rank-one and align perfectly with the training data -- minimize the training objective under norm constraints and achieve maximal local robustness under a non-zero Jacobian constraint. For continuous piecewise-affine deep networks, normal alignment manifests geometrically as centroid alignment within the network's induced power diagram partition and results from the feature-learning regime. Motivated by these theoretical insights, we introduce GrokAlign, a regularization strategy that actively induces normal alignment. We demonstrate that GrokAlign significantly accelerates the training dynamics of deep networks relevant to the grokking phenomenon. Furthermore, we apply the principle of normal alignment to Recursive Feature Machines (RFMs) to introduce Recursive Feature Alignment Machines (RFAMs). We show that RFAMs exhibit greater adversarial robustness compared to RFMs when trained on tabular data.

Thomas Walker, Varun Anand, Pavlos Andreadis

Semantic segmentation for spherical data is a challenging problem in machine learning since conventional planar approaches require projecting the spherical image to the Euclidean plane. Representing the signal on a fundamentally different topology introduces edges and distortions which impact network performance. Recently, graph-based approaches have bypassed these challenges to attain significant improvements by representing the signal on a spherical mesh. Current approaches to spherical segmentation exclusively use variants of the UNet architecture, meaning more successful planar architectures remain unexplored. Inspired by the success of feature pyramid networks (FPNs) in planar image segmentation, we leverage the pyramidal hierarchy of graph-based spherical CNNs to design spherical FPNs. Our spherical FPN models show consistent improvements over spherical UNets, whilst using fewer parameters. On the Stanford 2D-3D-S dataset, our models achieve state-of-the-art performance with an mIOU of 48.75, an improvement of 3.75 IoU points over the previous best spherical CNN.

Amr Sharafeldin, Shrisudhan Govindarajan, Thomas Walker et al.

Modern scene reconstruction methods, such as 3D Gaussian Splatting, enable photo-realistic novel view synthesis at real-time speeds. However, their adoption in interactive graphics applications remains limited due to the difficulty of interacting with these representations compared to traditional, human-authored 3D assets. While prior work has attempted to impose semantic decomposition on these models, significant challenges remain in segmentation quality and cross-view consistency.To address these limitations, we introduce Semantic Foam, which extends the recently proposed Radiant Foam representation to semantic decomposition tasks. Our approach leverages the inherent spatial structure of Radiant Foam's volumetric Voronoi mesh and augments it with an explicit semantic feature field defined at the cell level. This design enables direct spatial regularization, improving consistency across views and mitigating artifacts caused by occlusion and inconsistent supervision, which are common issues in point-based representations.Experimental results demonstrate that our method achieves superior object-level segmentation performance compared to state-of-the-art approaches such as Gaussian Grouping and SAGA.Project page: http://semanticfoam.github.io/

Thomas Walker, Octave Mariotti, Amir Vaxman et al.

Positional encodings are a common component of neural scene reconstruction methods, and provide a way to bias the learning of neural fields towards coarser or finer representations. Current neural surface reconstruction methods use a "one-size-fits-all" approach to encoding, choosing a fixed set of encoding functions, and therefore bias, across all scenes. Current state-of-the-art surface reconstruction approaches leverage grid-based multi-resolution hash encoding in order to recover high-detail geometry. We propose a learned approach which allows the network to choose its encoding basis as a function of space, by masking the contribution of features stored at separate grid resolutions. The resulting spatially adaptive approach allows the network to fit a wider range of frequencies without introducing noise. We test our approach on standard benchmark surface reconstruction datasets and achieve state-of-the-art performance on two benchmark datasets.

Thomas Walker

Machine learning models are trained with relatively simple objectives, such as next token prediction. However, on deployment, they appear to capture a more fundamental representation of their input data. It is of interest to understand the nature of these representations to help interpret the model's outputs and to identify ways to improve the salience of these representations. Concept vectors are constructions aimed at attributing concepts in the input data to directions, represented by vectors, in the model's latent space. In this work, we introduce concept boundary vectors as a concept vector construction derived from the boundary between the latent representations of concepts. Empirically we demonstrate that concept boundary vectors capture a concept's semantic meaning, and we compare their effectiveness against concept activation vectors.

Thomas Walker, Salvatore Esposito, Daniel Rebain et al.

Reconstructing complex structures from planar cross-sections is a challenging problem, with wide-reaching applications in medical imaging, manufacturing, and topography. Out-of-the-box point cloud reconstruction methods can often fail due to the data sparsity between slicing planes, while current bespoke methods struggle to reconstruct thin geometric structures and preserve topological continuity. This is important for medical applications where thin vessel structures are present in CT and MRI scans. This paper introduces CrossSDF, a novel approach for extracting a 3D signed distance field from 2D signed distances generated from planar contours. Our approach makes the training of neural SDFs contour-aware by using losses designed for the case where geometry is known within 2D slices. Our results demonstrate a significant improvement over existing methods, effectively reconstructing thin structures and producing accurate 3D models without the interpolation artifacts or over-smoothing of prior approaches.