Pipeline Parallelism (PP) enables large neural network training on small, interconnected devices by splitting the model into multiple stages. To maximize pipeline utilization, asynchronous optimization is appealing as it offers 100% pipeline utilization by construction. However, it is inherently challenging as the weights and gradients are no longer synchronized, leading to *stale (or delayed) gradients*. To alleviate this, we introduce a variant of Nesterov Accelerated Gradient (NAG) for asynchronous optimization in PP. Specifically, we modify the look-ahead step in NAG to effectively address the staleness in gradients. We theoretically prove that our approach converges at a sublinear rate in the presence of fixed delay in gradients. Our experiments on large-scale language modelling tasks using decoder-only architectures with up to **1B parameters**, demonstrate that our approach significantly outperforms existing asynchronous methods, even surpassing the synchronous baseline.

We consider a decentralized setup in which the participants collaboratively train and serve a large neural network, and where each participant only processes a subset of the model. In this setup, we explore the possibility of unmaterializable weights, where a full weight set is never available to any one participant. We introduce Unextractable Protocol Models (UPMs): a training and inference framework that leverages the sharded model setup to ensure model shards (i.e.,, subsets) held by participants are incompatible at different time steps. UPMs periodically inject time-varying, random, invertible transforms at participant boundaries; preserving the overall network function yet rendering cross-time assemblies incoherent. On Qwen-2.5-0.5B and Llama-3.2-1B, 10 000 transforms leave FP32 perplexity unchanged ($\Delta$PPL$< 0.01$; Jensen–Shannon drift $<4 \times 10^{-5}$), and we show how to control growth for lower precision datatypes. Applying a transform every 30s adds 3% latency, 0.1% bandwidth, and 10% GPU-memory overhead at inference, while training overhead falls to 1.6% time and < 1% memory. We consider several attacks, showing that the requirements of direct attacks are impractical and easy to defend against, and that gradient-based fine-tuning of stitched partitions consumes $\geq 60\%$ of the tokens required to train from scratch. By enabling models to be collaboratively trained yet not extracted, UPMs make it practical to embed programmatic incentive mechanisms in community-driven decentralized training.

Scaling models has led to significant advancements in deep learning, but training these models in decentralized settings remains challenging due to communication bottlenecks. While existing compression techniques are effective in data-parallel, they do not extend to model parallelism. Unlike data-parallel training, where weight gradients are exchanged, model-parallel requires compressing activations and activation gradients as they propagate through layers, accumulating compression errors. We propose a novel compression algorithm that compresses both forward and backward passes, enabling up to 99% compression with no convergence degradation with negligible memory/compute overhead. By leveraging a recursive structure in transformer networks, we predefine a low-dimensional subspace to confine the activations and gradients, allowing full reconstruction in subsequent layers. Our method achieves up to 100x improvement in communication efficiency and enables training billion-parameter-scale models over low-end GPUs connected via consumer-grade internet speeds as low as 80Mbps, matching the convergence of centralized datacenter systems with 100Gbps connections with model parallel.

Pretraining language models with extended context windows enhances their ability to leverage rich information during generation. Existing methods split input sequences into chunks, broadcast them across multiple devices, and compute attention block by block which incurs significant communication overhead. While feasible in high-speed clusters, these methods are impractical for decentralized training over low-bandwidth connections. We propose a compression method for communication-efficient context parallelism in decentralized settings, achieving a remarkable compression rate of over 95% with negligible overhead and no loss in convergence. Our key insight is to exploit the intrinsic low-rank structure of activation outputs by dynamically constraining them to learned mixtures of subspaces via efficient reparameterizations. We demonstrate scaling billion-parameter decentralized models to context lengths exceeding 100K tokens on networks as slow as 300Mbps, matching the wall-clock convergence speed of centralized models on 100Gbps interconnects.

Inferring the 3D structure of a non-rigid dynamic scene from a single moving camera is an under-constrained problem. Inspired by the remarkable progress of neural radiance fields (NeRFs) in photo-realistic novel view synthesis of static scenes, it has also been extended to dynamic settings. Such methods heavily rely on implicit neural priors to regularize the problem. In this work, we take a step back and investigate how current implementations may entail deleterious effects including limited expressiveness, entanglement of light and density fields, and sub-optimal motion localization. Further, we devise a factorisation-based framework that represents the scene as a composition of bandlimited, high-dimensional signals. We demonstrate compelling results across complex dynamic scenes that involve changes in lighting, texture and long-range dynamics.

Recent advancements in machine learning have spotlighted the potential of hyperbolic spaces as they effectively learn hierarchical feature representations. While there has been progress in leveraging hyperbolic spaces in single-modality contexts, its exploration in multimodal settings remains under explored. A recent work has sought to transpose Euclidean multimodal learning techniques to hyperbolic spaces, by adopting a geodesic distance based contrastive loss. However, we show both theoretically and empirically that such spatial proximity based contrastive loss significantly disrupts hierarchies in the latent space. To remedy this, we advocate that the cross-modal representations should accept the inherent modality gap between text and images, and introduce a novel approach to measure cross-modal similarity that does not enforce spatial proximity. Our approach shows remarkable capabilities in preserving unimodal hierarchies while aligning the two modalities. Our experiments on a series of downstream tasks demonstrate that a better latent structure emerges with our objective function while being superior in text-to-image and image-to-text retrieval tasks.

Novel-view synthesis through diffusion models has demonstrated remarkable potential for generating diverse and high-quality images. Yet, the independent process of image generation in these prevailing methods leads to challenges in maintaining multiple view consistency. To address this, we introduce ViewFusion, a novel, training-free algorithm that can be seamlessly integrated into existing pre-trained diffusion models. Our approach adopts an auto-regressive method that implicitly leverages previously generated views as context for next view generation, ensuring robust multi-view consistency during the novel-view generation process. Through a diffusion process that fuses known-view information via interpolated denoising, our framework successfully extends single-view conditioned models to work in multiple-view conditional settings without any additional fine-tuning. Extensive experimental results demonstrate the effectiveness of View Fusion in generating consistent and detailed novel views.

Synthesizing novel views for dynamic scenes from a collection of RGB inputs poses significant challenges due to the inherent under-constrained nature of the problem. To mitigate this ill-posedness, practitioners in the field of neural radiance fields (NeRF) often resort to the adoption of intricate geometric regularization techniques, including scene flow, depth estimation, or learned perceptual similarity. While these geometric cues have demonstrated their effectiveness, their incorporation leads to evaluation of computationally expensive off-the-shelf models, introducing substantial computational overhead into the pipeline. Moreover, seamlessly integrating such modules into diverse dynamic NeRF models can be a non-trivial task, hindering their utilization in an architecture-agnostic manner. In this paper, we propose a theoretically grounded, lightweight regularizer by treating the dynamics of a time-varying scene as a low-frequency change of a probability distribution of the light intensity. We constrain the dynamics of this distribution using optimal transport (OT) and provide error bounds under reasonable assumptions. Our regularization is learning-free, architecture agnostic, and can be implemented with just a few lines of code. Finally, we demonstrate the practical efficacy of our regularizer across state-of-the-art architectures.

Lipstick virtual try-on (VTO) experiences have become widespread across the e-commerce sector and assist users in eliminating the guesswork of shopping online. However, such experiences still lack in both realism and accuracy. In this work, we propose LipAT, a neural framework that blends the strengths of Physics-Based Rendering (PBR) and Neural Style Transfer (NST) approaches to directly apply lipstick onto face images given lipstick attributes (e.g., colour, finish type). LipAT consists of a physics aware neural lipstick application module (LAM) to apply lipstick on face images given its attributes and Lipstick Refiner Module (LRM) to improve the realism by refining the imperfections. Unlike the NST approaches, LipAT allows precise and controllable lipstick attribute preservation, without requiring crude approximations and inference of various intertwined environment factors (e.g., scene lighting, face structure etc) involved in image generation that is required for accurate PBR. We propose an experimental framework with quantitative metrics to evaluate different desirable aspects of the lipstick attribute driven try-on alongside user studies to further validate our findings. Our results show that LipAT considerably outperforms fully-automated PBR approaches in preserving realism and the NST approaches in preserving various lipstick attributes such as finish types.

Whilst Generative Adversarial Networks (GANs) have gained a reputation as powerful generative models, they are notoriously difficult to train and suffer from instability in optimisation. Recent methods for tackling this drawback have typically approached it by inducing better behaviour on the discriminator component of the GAN; these include loss function modification, gradient regularisation and weight normalisation to create a discriminator that is well-behaved from a Lipschitz perspective. In this paper, we propose a novel and orthogonal contribution which modifies the architecture of a GAN. Our method embeds the powerful discriminating capabilities inherent in decision forests within the discriminator of a GAN. Empirically, we test the effectiveness of our approach on the CIFAR-10, Oxford Flowers and CUB Birds datasets. We show that our technique is easy to incorporate into existing GAN baselines and offers improvements on Fréchet-Inception Distance (FID) scores by as high as 56.1% over several GAN baselines.

Domain adaptation and generative modelling have collectively mitigated the expensive nature of data collection and labelling by leveraging the rich abundance of accurate, labelled data in simulation environments. In this work, we study the performance gap that exists between representations optimised for localisation on simulation environments and the application of such representations in a real-world setting. Our method exploits the shared geometric similarities between simulation and real-world environments whilst maintaining invariance towards visual discrepancies. This is achieved by optimising a representation extractor to project both simulated and real representations into a shared representation space. Our method uses a symmetrical adversarial approach which encourages the representation extractor to conceal the domain that features are extracted from and simultaneously preserves robust attributes between source and target domains that are beneficial for localisation. We evaluate our method by adapting representations optimised for indoor Habitat simulated environments (Matterport3D and Replica) to a real-world indoor environment (Active Vision Dataset), showing that it compares favourably against fully-supervised approaches.

Although Generative Adversarial Networks (GANs) are known for their sharp realism in image generation, they often fail to estimate areas of the data density. This leads to low modal diversity and at times distorted generated samples. These problems essentially arise from poor estimation of the distance metric responsible for training these networks. To address these issues, we introduce an additional regularisation term which performs optimal transport in parallel within a low dimensional representation space. We demonstrate that operating in a low dimension representation of the data distribution benefits from convergence rate gains in estimating the Wasserstein distance, resulting in more stable GAN training. We empirically show that our regulariser achieves a stabilising effect which leads to higher quality of generated samples and increased mode coverage of the given data distribution. Our method achieves significant improvements on the CIFAR-10, Oxford Flowers and CUB Birds datasets over several GAN baselines both qualitatively and quantitatively.

Continuously estimating an agent's state space and a representation of its surroundings has proven vital towards full autonomy. A shared common ground among systems which successfully achieve this feat is the integration of previously encountered observations into the current state being estimated. This necessitates the use of a memory module for incorporating previously visited states whilst simultaneously offering an internal representation of the observed environment. In this work we develop a memory module which contains rigidly aligned point-embeddings that represent a coherent scene structure acquired from an RGB-D sequence of observations. The point-embeddings are extracted using modern convolutional neural network architectures, and alignment is performed by computing a dense correspondence matrix between a new observation and the current embeddings residing in the memory module. The whole framework is end-to-end trainable, resulting in a recurrent joint optimisation of the point-embeddings contained in the memory. This process amplifies the shared information across states, providing increased robustness and accuracy. We show significant improvement of our method across a set of experiments performed on the synthetic VIZDoom environment and a real world Active Vision Dataset.