Tehrim Yoon

Modern generative models, particularly denoising diffusion probabilistic models (DDPMs), provide high-quality synthetic images, enabling users to generate diverse images and videos that are realistic. However, in a number of situations, edge devices or individual institutions may possess locally collected data that is highly sensitive and should ensure data privacy, such as in the field of healthcare and finance. Under such federated learning (FL) settings, various methods on training generative models have been studied, but most of them assume generative adversarial networks (GANs), and the algorithms are specific to GANs and not other forms of generative models such as DDPM. This paper proposes a new algorithm for training DDPMs under federated learning settings, VQ-FedDiff, which provides a personalized algorithm for training diffusion models that can generate higher-quality images FID while still keeping risk of breaching sensitive information as low as locally-trained secure models. We demonstrate that VQ-FedDiff shows state-of-the-art performance on existing federated learning of diffusion models in both IID and non-IID settings, and in benchmark photorealistic and medical image datasets. Our results show that diffusion models can efficiently learn with decentralized, sensitive data, generating high-quality images while preserving data privacy.

Neural ordinary differential equations (NODE) present a new way of considering a deep residual network as a continuous structure by layer depth. However, it fails to overcome its representational limits, where it cannot learn all possible homeomorphisms of input data space, and therefore quickly saturates in terms of performance even as the number of layers increases. Here, we show that simply stacking Neural ODE blocks could easily improve performance by alleviating this issue. Furthermore, we suggest a more effective way of training neural ODE by using a time-evolving mixture weight on multiple ODE functions that also evolves with a separate neural ODE. We provide empirical results that are suggestive of improved performance over stacked as well as vanilla neural ODEs where we also confirm our approach can be orthogonally combined with recent advances in neural ODEs.

Decisions are made based on the subjective value that the brain assigns to options. However, subjective value is a mathematical construct that cannot be measured directly, but rather inferred from choices. Recent results have demonstrated that reaction time and velocity of movements are modulated by reward, raising the possibility that there is a link between how the brain evaluates an option, and how it controls movements toward that option. Here, we asked people to choose among risky options represented by abstract stimuli, some associated with gain, others with loss. From their choices in decision trials we estimated the subjective value that they assigned to each stimulus. In probe trials, they were presented with a single stimulus at center and made a saccade to a peripheral location. We found that the reaction time and peak velocity of that saccade varied roughly linearly from loss to gain with the subjective value of the stimulus. Naturally, participants differed in how much they valued a given stimulus. Remarkably, those who valued a stimulus more, as evidenced by their choices in decision trials, tended to move with greater vigor in response to that stimulus in probe trials. Thus, saccade vigor partly reflected the subjective value that the brain assigned the stimulus. However, the influence of subjective value on vigor was only a modest predictor of preference: vigor in probe trials allowed us to predict choice in decision trials with roughly 60% accuracy. New and Noteworthy: We found that saccade vigor tends to vary monotonically with subjective value: smallest for stimuli that predict a loss, and highest for stimuli that predict a gain. Notably, between-subject differences in valuation could be gleaned from the between-subject differences in their patterns of vigor. However, the influence of subjective value on vigor was modest, allowing partial ability to infer subjective value for the purpose of predicting choice in decision trials.