Spatial audio is essential for enhancing the immersiveness of audio-visual experiences, yet its production typically demands complex recording systems and specialized expertise. In this work, we address a novel problem of generating first-order ambisonics, a widely used spatial audio format, directly from silent videos. To support this task, we introduce YT-Ambigen, a dataset comprising 102K 5-second YouTube video clips paired with corresponding first-order ambisonics. We also propose new evaluation metrics to assess the spatial aspect of generated audio based on audio energy maps and saliency metrics. Furthermore, we present Video-to-Spatial Audio Generation (ViSAGe), an end-to-end framework that generates first-order ambisonics from silent video frames by leveraging CLIP visual features, autoregressive neural audio codec modeling with both directional and visual guidance. Experimental results demonstrate that ViSAGe produces plausible and coherent first-order ambisonics, outperforming two-stage approaches consisting of video-to-audio generation and audio spatialization. Qualitative examples further illustrate that ViSAGe generates temporally aligned high-quality spatial audio that adapts to viewpoint changes.

Low-rank adaptation (LoRA) has become a standard approach for fine-tuning large foundation models. However, our theoretical understanding of LoRA remains limited as prior analyses of LoRA's training dynamics either rely on linearization arguments or consider highly simplified setups. In this work, we analyze the LoRA loss landscape without such restrictive assumptions. We define two regimes: a "special regime", which includes idealized setups where linearization arguments hold, and a "generic regime" representing more realistic setups where linearization arguments do not hold. In the generic regime, we show that LoRA training converges to a global minimizer with low rank and small magnitude, or a qualitatively distinct solution with high rank and large magnitude. Finally, we argue that the zero-initialization and weight decay in LoRA training induce an implicit bias toward the low-rank, small-magnitude region of the parameter space—where global minima lie—thus shedding light on why LoRA training usually succeeds in finding global minima.

In recent years, masked diffusion models (MDMs) have emerged as a promising alternative approach for generative modeling over discrete domains. Compared to autoregressive models (ARMs), MDMs trade off complexity at training time with flexibility at inference time. At training time, they must learn to solve an exponentially large number of infilling problems, but at inference time, they can decode tokens in essentially arbitrary order. In this work we closely examine these two competing effects. On the training front, we theoretically and empirically demonstrate that MDMs indeed train on computationally intractable subproblems compared to their autoregressive counterparts. On the inference front, we show that a suitable strategy for adaptively choosing the token decoding order significantly enhances the capabilities of MDMs, allowing them to sidestep hard subproblems. On logic puzzles like Sudoku, we show that adaptive inference can boost solving accuracy in pretrained MDMs from $<7$\% to $\approx 90$\%, even outperforming ARMs that were explicitly trained via teacher forcing to learn the right order of decoding.

People continuously perceive and interact with their surroundings based on underlying intentions that drive their exploration and behaviors. While research in egocentric user and scene understanding has focused primarily on motion and contact-based interaction, forecasting human visual perception itself remains less explored despite its fundamental role in guiding human actions and its implications for AR/VR and assistive technologies. We address the challenge of egocentric 3D visual span forecasting, predicting where a person's visual perception will focus next within their three-dimensional environment. To this end, we propose EgoSpanLift, a novel method that transforms egocentric visual span forecasting from 2D image planes to 3D scenes. EgoSpanLift converts SLAM-derived keypoints into gaze-compatible geometry and extracts volumetric visual span regions. We further combine EgoSpanLift with 3D U-Net and unidirectional transformers, enabling spatio-temporal fusion to efficiently predict future visual span in the 3D grid. In addition, we curate a comprehensive benchmark from raw egocentric multisensory data, creating a testbed with 364.6K samples for 3D visual span forecasting. Our approach outperforms competitive baselines for egocentric gaze anticipation and 3D localization, while achieving comparable results even when projected back onto 2D image planes without additional 2D-specific training.

Neural radiance field (NeRF) is an emerging technique for 3D scene reconstruction and modeling. However, current NeRF-based methods are limited in the capabilities of adding or removing objects. This paper fills the aforementioned gap by proposing a new language-driven method for object manipulation in NeRFs through dataset updates. Specifically, to insert an object represented by a set of multi-view images into a background NeRF, we use a text-to-image diffusion model to blend the object into the given background across views. The generated images are then used to update the NeRF so that we can render view-consistent images of the object within the background. To ensure view consistency, we propose a dataset update strategy that prioritizes the radiance field training based on camera poses in a pose-ordered manner. We validate our method in two case studies: object insertion and object removal. Experimental results show that our method can generate photo-realistic results and achieves state-of-the-art performance in NeRF editing.

We propose a framework to learn semantics from raw audio signals using two types of representations, encoding contextual and phonetic information respectively. Specifically, we introduce a speech-to-unit processing pipeline that captures two types of representations with different time resolutions. For the language model, we adopt a dual-channel architecture to incorporate both types of representation. We also present new training objectives, masked context reconstruction and masked context prediction, that push models to learn semantics effectively. Experiments on the sSIMI metric of Zero Resource Speech Benchmark 2021 and Fluent Speech Command dataset show our framework learns semantics better than models trained with only one type of representation.

We propose EnCLAP, a novel framework for automated audio captioning. EnCLAP employs two acoustic representation models, EnCodec and CLAP, along with a pretrained language model, BART. We also introduce a new training objective called masked codec modeling that improves acoustic awareness of the pretrained language model. Experimental results on AudioCaps and Clotho demonstrate that our model surpasses the performance of baseline models. Source code will be available at https://github.com/jaeyeonkim99/EnCLAP.1

Recently, accelerated algorithms using the anchoring mechanism for minimax optimization and fixed-point problems have been proposed, and matching complexity lower bounds establish their optimality. In this work, we present the surprising observation that the optimal acceleration mechanism in minimax optimization and fixed-point problems is not unique. Our new algorithms achieve exactly the same worst-case convergence rates as existing anchor-based methods while using materially different acceleration mechanisms. Specifically, these new algorithms are dual to the prior anchor-based accelerated methods in the sense of H-duality. This finding opens a new avenue of research on accelerated algorithms since we now have a family of methods that empirically exhibit varied characteristics while having the same optimal worst-case guarantee.

In convex optimization, first-order optimization methods efficiently minimizing function values have been a central subject study since Nesterov's seminal work of 1983. Recently, however, Kim and Fessler's OGM-G and Lee et al.'s FISTA-G have been presented as alternatives that efficiently minimize the gradient magnitude instead. In this paper, we present H-duality, which represents a surprising one-to-one correspondence between methods efficiently minimizing function values and methods efficiently minimizing gradient magnitude. In continuous-time formulations, H-duality corresponds to reversing the time dependence of the dissipation/friction term. To the best of our knowledge, H-duality is different from Lagrange/Fenchel duality and is distinct from any previously known duality or symmetry relations. Using H-duality, we obtain a clearer understanding of the symmetry between Nesterov's method and OGM-G, derive a new class of methods efficiently reducing gradient magnitudes of smooth convex functions, and find a new composite minimization method that is simpler and faster than FISTA-G.

In this paper, we propose a new method for mapping a 3D point cloud to the latent space of a 3D generative adversarial network. Our generative model for 3D point clouds is based on SP-GAN, a state-of-the-art sphere-guided 3D point cloud generator. We derive an efficient way to encode an input 3D point cloud to the latent space of the SP-GAN. Our point cloud encoder can resolve the point ordering issue during inversion, and thus can determine the correspondences between points in the generated 3D point cloud and those in the canonical sphere used by the generator. We show that our method outperforms previous GAN inversion methods for 3D point clouds, achieving state-of-the-art results both quantitatively and qualitatively. Our code is available at https://github.com/hkust-vgd/point_inverter.

High-resolution daytime satellite imagery has become a promising source to study economic activities. These images display detailed terrain over large areas and allow zooming into smaller neighborhoods. Existing methods, however, have utilized images only in a single-level geographical unit. This research presents a deep learning model to predict economic indicators via aggregating traits observed from multiple levels of geographical units. The model first measures hyperlocal economy over small communities via ordinal regression. The next step extracts district-level features by summarizing interconnection among hyperlocal economies. In the final step, the model estimates economic indicators of districts via aggregating the hyperlocal and district information. Our new multi-level learning model substantially outperforms strong baselines in predicting key indicators such as population, purchasing power, and energy consumption. The model is also robust against data shortage; the trained features from one country can generalize to other countries when evaluated with data gathered from Malaysia, the Philippines, Thailand, and Vietnam. We discuss the multi-level model's implications for measuring inequality, which is the essential first step in policy and social science research on inequality and poverty.

As more and more mobile/embedded applications employ Deep Neural Networks (DNNs) involving sensitive user data, mobile/embedded devices must provide a highly secure DNN execution environment to prevent privacy leaks. Aimed at securing DNN data, recent studies execute part of a DNN in a trusted execution environment (e.g., TrustZone) to isolate DNN execution from the other processes; however, as the trusted execution environments for mobile/embedded devices provide limited memory protection, DNN data remain unencrypted in DRAM and become vulnerable to physical attacks. The devices can prevent the physical attacks by keeping DNN data encrypted in DRAM; when DNN data get referenced during DNN execution, they get loaded to the SRAM and get decrypted by a CPU core. Unfortunately, using the SRAM with demand paging greatly increases DNN execution time due to the inefficient use of the SRAM and the high CPU consumption of data encryption/decryption.

We present dataflow mirroring, architectural support for low-overhead fine-grained systolic array allocation which overcomes the limitations of prior coarse-grained spatial-multitasking Neural Processing Unit (NPU) architectures. The key idea of dataflow mirroring is to reverse the dataflows of co-located Neural Networks (NNs) in horizontal and/or vertical directions, allowing allocation boundaries to be set between any adjacent rows and columns of a systolic array and supporting up to four-way spatial multitasking. Our detailed experiments using MLPerf NNs and a dataflow-mirroring-augmented NPU prototype which extends Google’s TPU with dataflow mirroring shows that dataflow mirroring can significantly improve the multitasking performance by up to 46.4%.

With recent developments of convolutional neural net-works, deep learning for 3D point clouds has shown significant progress in various 3D scene understanding tasks, e.g., object recognition, semantic segmentation. In a safety-critical environment, it is however not well understood how such deep learning models are vulnerable to adversarial examples. In this work, we explore adversarial attacks for point cloud-based neural networks. We propose a unified formulation for adversarial point cloud generation that can generalise two different attack strategies. Our method generates adversarial examples by attacking the classification ability of point cloud-based networks while considering the perceptibility of the examples and ensuring the minimal level of point manipulations. Experimental results show that our method achieves the state-of-the-art performance with higher than 89% and 90% of attack success rate on synthetic and real-world data respectively, while manipulating only about 4% of the total points.