Topic: diffusion models

Antonio Ocello tests the idea that heavy-tailed noise improves diffusion and flow-based generative models by matching heavy-tailed data and encouraging diverse samples. Replacing Gaussian noise also changes the estimation task. A combined theoretical and experimental study derives sampling-error bounds for representative heavy- and light-tailed diffusion models, finding that heavy-tailed noise makes statistical estimation harder and produces less favourable bounds. Experiments with synthetic and real data recover the predicted trade-off. The results question whether heavy-tailed initialization

Google DeepMind research director Brendan O'Donoghue examines discrete diffusion as an alternative to autoregressive text generation. He discusses sample diversity, reinforcement learning, latency and potential on-device or robotics applications. The episode also addresses the costs of serving diffusion models, hardware trade-offs between computation and memory bandwidth, and practical limits on progress toward more capable AI. Hosted by Ravid Shwartz Ziv and Allen Roush. Watch the full conversation on the publisher’s YouTube channel.

Developing agents capable of modeling complex environments and human behaviors within them is a key goal of artificial intelligence research. Progress towards this goal has exciting potential for applications in video games, from new tools that empower game developers to realize new creative visions, to enabling new kinds of immersive player experiences. This talk focuses on recent advances of my team at Microsoft Research towards scalable machine learning architectures that effectively capture human gameplay data. In the first part of my talk, I will focus on diffusion models as generative mo

Developing agents capable of modeling complex environments and human behaviors within them is a key goal of artificial intelligence research. Progress towards this goal has exciting potential for applications in video games, from new tools that empower game developers to realize new creative visions, to enabling new kinds of immersive player experiences. This talk focuses on recent advances of my team at Microsoft Research towards scalable machine learning architectures that effectively capture human gameplay data. In the first part of my talk, I will focus on diffusion models as generative mo

Like all neural tissue, the retina has a high metabolic demand, and requires a constant supply of oxygen. Second and third order neurons are supplied by the retinal circulation, whose characteristics are similar to brain circulation. However, the photoreceptor region, which occupies half of the retinal thickness, is avascular, and relies on diffusion of oxygen from the choroidal circulation, whose properties are very different, as well as the retinal circulation. By fitting diffusion models to oxygen measurements made with oxygen microelectrodes, it is possible to understand the relative roles

Decision-making in noisy environments with constant sensory evidence involves integrating sequentially-sampled evidence, a strategy formalized by diffusion models which is supported by decades behavioral and neural findings. By contrast, it is unknown whether this strategy is also used during decision-making when the underlying sensory evidence is expected to change. Here, we trained monkeys to identify the dominant color of a dynamically refreshed checkerboard pattern that doesn't become informative until after a variable delay. Animals' behavioral responses were briefly suppressed after an a

Modern datasets in neuroscience enable unprecedented inquiries into the relationship between complex behaviors and the activity of many simultaneously recorded neurons. While latent variable models can successfully extract low-dimensional embeddings from such recordings, using them to generate realistic spiking data, especially in a behavior-dependent manner, still poses a challenge. Here, we present Latent Diffusion for Neural Spiking data (LDNS), a diffusion-based generative model with a low-dimensional latent space: LDNS employs an autoencoder with structured state-space (S4) layers to pro