arXiv subjects

Urs Köster

Publications and source records attributed to Urs Köster.

7 recordsLinked to original sources

arXiv2026

Towards a future space-based, highly scalable AI infrastructure system design

If AI is a foundational general-purpose technology, we should anticipate that demand for AI compute -- and energy -- will continue to grow. The Sun is by far the largest energy source in our solar system, and thus it warrants consideration how future AI infrastructure could most efficiently tap into that power. This work explores a scalable compute system for machine learning in space, using fleets of satellites equipped with solar arrays, inter-satellite links using free-space optics, and Google tensor processing unit (TPU) accelerator chips. To facilitate high-bandwidth, low-latency inter-satellite communication, the satellites would be flown in close proximity. We illustrate the basic approach to formation flight via an 81-satellite cluster of 1 km radius, and describe an approach for using high-precision ML-based models to control large-scale constellations. Trillium TPUs are radiation tested. They survive a total ionizing dose equivalent to a 5 year mission life without permanent failures, and are characterized for bit-flip errors. Launch costs are a critical part of overall system cost; a learning curve analysis suggests launch to low-Earth orbit (LEO) may reach $\lesssim$\$200/kg by the mid-2030s.

cs.DC↗

arXiv2024

Scalable Spatiotemporal Prediction with Bayesian Neural Fields

Spatiotemporal datasets, which consist of spatially-referenced time series, are ubiquitous in diverse applications, such as air pollution monitoring, disease tracking, and cloud-demand forecasting. As the scale of modern datasets increases, there is a growing need for statistical methods that are flexible enough to capture complex spatiotemporal dynamics and scalable enough to handle many observations. This article introduces the Bayesian Neural Field (BayesNF), a domain-general statistical model that infers rich spatiotemporal probability distributions for data-analysis tasks including forecasting, interpolation, and variography. BayesNF integrates a deep neural network architecture for high-capacity function estimation with hierarchical Bayesian inference for robust predictive uncertainty quantification. Evaluations against prominent baselines show that BayesNF delivers improvements on prediction problems from climate and public health data containing tens to hundreds of thousands of measurements. Accompanying the paper is an open-source software package (https://github.com/google/bayesnf) that runs on GPU and TPU accelerators through the JAX machine learning platform.

cs.LG↗

arXiv2021

Pipelined Backpropagation at Scale: Training Large Models without Batches

New hardware can substantially increase the speed and efficiency of deep neural network training. To guide the development of future hardware architectures, it is pertinent to explore the hardware and machine learning properties of alternative training algorithms. In this work we evaluate the use of small batch, fine-grained Pipelined Backpropagation, an asynchronous pipeline parallel training algorithm that has significant hardware advantages. We introduce two methods, Spike Compensation and Linear Weight Prediction, that effectively mitigate the downsides caused by the asynchronicity of Pipelined Backpropagation and outperform existing techniques in our setting. We show that appropriate normalization and small batch sizes can also aid training. With our methods, fine-grained Pipelined Backpropagation using a batch size of one can match the accuracy of SGD for multiple networks trained on CIFAR-10 and ImageNet. Simple scaling rules allow the use of existing hyperparameters for traditional training without additional tuning.

cs.LG↗

arXiv2020

Adaptive Braking for Mitigating Gradient Delay

Neural network training is commonly accelerated by using multiple synchronized workers to compute gradient updates in parallel. Asynchronous methods remove synchronization overheads and improve hardware utilization at the cost of introducing gradient delay, which impedes optimization and can lead to lower final model performance. We introduce Adaptive Braking (AB), a modification for momentum-based optimizers that mitigates the effects of gradient delay. AB dynamically scales the gradient based on the alignment of the gradient and the velocity. This can dampen oscillations along high curvature directions of the loss surface, stabilizing and accelerating asynchronous training. We show that applying AB on top of SGD with momentum enables training ResNets on CIFAR-10 and ImageNet-1k with delays $D \geq$ 32 update steps with minimal drop in final test accuracy.

cs.LG↗

arXiv2017

Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks

Deep neural networks are commonly developed and trained in 32-bit floating point format. Significant gains in performance and energy efficiency could be realized by training and inference in numerical formats optimized for deep learning. Despite advances in limited precision inference in recent years, training of neural networks in low bit-width remains a challenging problem. Here we present the Flexpoint data format, aiming at a complete replacement of 32-bit floating point format training and inference, designed to support modern deep network topologies without modifications. Flexpoint tensors have a shared exponent that is dynamically adjusted to minimize overflows and maximize available dynamic range. We validate Flexpoint by training AlexNet, a deep residual network and a generative adversarial network, using a simulator implemented with the neon deep learning framework. We demonstrate that 16-bit Flexpoint closely matches 32-bit floating point in training all three models, without any need for tuning of model hyperparameters. Our results suggest Flexpoint as a promising numerical format for future hardware for training and inference.

cs.LG↗

arXiv2013

Testing our conceptual understanding of V1 function

Here we test our conceptual understanding of V1 function by asking two experimental questions: 1) How do neurons respond to the spatiotemporal structure contained in dynamic, natural scenes? and 2) What is the true range of visual responsiveness and predictability of neural responses obtained in an unbiased sample of neurons across all layers of cortex? We address these questions by recording responses to natural movie stimuli with 32 channel silicon probes. By simultaneously recording from cells in all layers, and taking all recorded cells, we reduce recording bias that results from "hunting" for neural responses evoked from drifting bars and gratings. A nonparametric model reveals that many cells that are visually responsive do not appear to be captured by standard receptive field models. Using nonlinear Radial Basis Function kernels in a support vector machine, we can explain the responses of some of these cells better than standard linear and phase-invariant complex cell models. This suggests that V1 neurons exhibit more complex and diverse responses than standard models can capture, ranging from simple and complex cells strongly driven by their classical receptive fields, to cells with more nonlinear receptive fields inferred from the nonparametric and RFB model, and cells that are not visually responsive despite robust firing.

q-bio.NC↗

arXiv2013

Higher Order Correlations within Cortical Layers Dominate Functional Connectivity in Microcolumns

We report on simultaneous recordings from cells in all layers of visual cortex and models developed to capture the higher order structure of population spiking activity. Specifically, we use Ising, Restricted Boltzmann Machine (RBM) and semi-Restricted Boltzmann Machine (sRBM) models to reveal laminar patterns of activity. While the Ising model describes only pairwise couplings, the RBM and sRBM capture higher-order dependencies using hidden units. Applied to 32- channel polytrode data recorded from cat visual cortex, the higher-order mod- els discover functional connectivity preferentially linking groups of cells within a cortical layer. Both RBM and sRBM models outperform Ising models in log- likelihood. Additionally, we train all three models on spatiotemporal sequences of states, exposing temporal structure and allowing us to predict spiking from network history. This demonstrates the importance of modeling higher order in- teractions across space and time when characterizing activity in cortical networks.

q-bio.NC↗