Google's Cloud TPUs are a promising new hardware architecture for machine learning workloads. They have powered many of Google's milestone machine learning achievements in recent years. Google has now made TPUs available for general use on their cloud platform and as of very recently has opened them up further to allow…
On-device research index
arXiv research
A locally-built, LLM-digested index of recent arXiv papers in quant finance, geometry/topology, and statistical ML — keyword search served straight from SQLite on this machine.
168,742 papers · 148 categories
Trend · papers per month
48 results for “TPU”
Optimized neural networks for Edge TPU achieve high accuracy in real-time image classification.
problem Designing neural networks for hardware accelerators to achieve optimal performance.
method Hardware-aware neural architecture search and model customization for Edge TPU.
result Improved accuracy-latency tradeoff on Pixel 4's Edge TPU compared to existing models.
Monte Carlo methods are critical to many routines in quantitative finance such as derivatives pricing, hedging and risk metrics. Unfortunately, Monte Carlo methods are very computationally expensive when it comes to running simulations in high-dimensional state spaces where they are still a method of choice in the fina…
Training deep learning models is compute-intensive and there is an industry-wide trend towards hardware specialization to improve performance. To systematically benchmark deep learning platforms, we introduce ParaDnn, a parameterized benchmark suite for deep learning that generates end-to-end models for fully connected…
Trade uncertainty impact on stock-bond correlations: Insights from conditional correlation modelsq-fin.ST
The study shows how trade uncertainty affects stock-bond correlations over time.
problem Impact of trade policy uncertainty on stock-bond correlations.
method Daily data analysis using GARCH-based models (CCC, STCC, DCC) with TPU and political dummy variables.
result Time-varying correlation models better capture the dynamics of stock-bond correlations than constant models.
The paper models and analyzes faults in TPU-based neural networks.
problem Faults in TPU matrix multiplication units cause errors in DNN inference.
method Formal model using DTMC, probabilistic model checking, and experiments.
result Classification accuracy is sensitive to fault type, location, and neural network layers.
Framework for precise recall control in spatial conflation tasks.
problem Precise recall control in large-scale spatial conflation tasks to avoid downstream analytics failures and excessive manual review.
method End-to-end framework using equigrid bounding-box filter, CSR representation, neural ranker, and inverse-variance weighted ensemble of threshold estimators.
result Achieves exact recall with sub-percent variance over tens of millions of geometry pairs, runs on a single TPU v3 core.
We propose an sorting algorithm by Machine Learning method, which shows a huge potential sorting big data. This sorting algorithm can be applied to parallel sorting and is suitable for GPU or TPU acceleration. Furthermore, we discuss the application of this algorithm to sparse hash table.
This paper improves financial simulations using Tensor Processing Units and Tensorflow.
problem Estimating sensitivities in financial models efficiently.
method Utilizing Tensor Processing Units and Tensorflow for fast and automated differentiation.
result Single line of code for estimating sensitivities in financial models.
CWY Parametrization: a Solution for Parallelized Optimization of Orthogonal and Stiefel Matricescs.LG
Efficiently optimizes orthogonal and Stiefel matrices on parallel units.
problem Optimization over orthogonal groups on parallel units.
method CWY and T-CWY transforms for parametrization and optimization.
result CWY and T-CWY methods lead to convergence on parallel units.
GShard enables scaling of large neural networks with automatic sharding and lightweight APIs.
problem Scaling neural networks to handle vast training data and compute efficiently.
method GShard uses lightweight annotation APIs and XLA compiler extensions for parallel computation.
result GShard successfully trained a 600 billion parameter model on 2048 TPUs in 4 days.
LAD-BNet improves real-time energy forecasting on edge devices.
problem Real-time energy forecasting on edge devices for smart grid optimization and intelligent buildings.
method Hybrid neural architecture combining temporal lag exploitation and TCN with dilated convolutions.
result 14.49% MAPE at 1-hour horizon with 18ms inference time on Edge TPU, 8-12x faster than CPU.
This paper optimizes deep learning training by efficiently sharding weight updates across replicas.
problem Redundant weight update computation on all replicas in data-parallel training.
method Automatic sharding of weight updates using static analysis and transformations on the training graph.
result Substantial speedups achieved on large-scale models using Cloud TPUs.
Training machine learning (ML) models on large datasets requires considerable computing power. To speed up training, it is typical to distribute training across several machines, often with specialized hardware like GPUs or TPUs. Managing a distributed training job is complex and requires dealing with resource contenti…
Deep neural networks (DNN) are increasingly being accelerated on application-specific hardware such as the Google TPU designed especially for deep learning. Timing speculation is a promising approach to further increase the energy efficiency of DNN accelerators. Architectural exploration for timing speculation requires…
NS-RGS improves orthogonal group synchronization with faster convergence.
problem Orthogonal group synchronization from pairwise measurements.
method Newton-Schulz iteration for Riemannian gradient optimization.
result NS-RGS achieves linear convergence and near-optimal accuracy.
The impact of the maximally possible batch size (for the better runtime) on performance of graphic processing units (GPU) and tensor processing units (TPU) during training and inference phases is investigated. The numerous runs of the selected deep neural network (DNN) were performed on the standard MNIST and Fashion-M…
RETINA Benchmark evaluates Bayesian deep learning on diabetic retinopathy detection.
problem Reliable uncertainty quantification for deep learning models in medical applications.
method Design and evaluation of a real-world diabetic retinopathy dataset and tasks.
result Benchmarking of Bayesian deep learning methods on diabetic retinopathy detection tasks.
not-so-BigGAN: Generating High-Fidelity Images on Small Compute with Wavelet-based Super-Resolutioneess.IV
A cost-effective method to generate high-resolution images using wavelet-based super-resolution.
problem High compute resources required for training state-of-the-art image generation models.
method Wavelet-based two-step training framework: low-frequency sampling followed by super-resolution.
result Achieves FID of 10.59 on ImageNet 512x512 with half the compute of BigGAN.
Efficiently trains BERT on academic GPUs in 12 days.
problem Training large-scale BERT models is expensive and time-consuming.
method Optimizes training on multiple GPUs and nodes, reducing costs.
result Trains BERT on academic GPUs in 12 days, not requiring expensive hardware.
AdaNet is a lightweight TensorFlow-based (Abadi et al., 2015) framework for automatically learning high-quality ensembles with minimal expert intervention. Our framework is inspired by the AdaNet algorithm (Cortes et al., 2017) which learns the structure of a neural network as an ensemble of subnetworks. We designed it…
Deep learning is extremely computationally intensive, and hardware vendors have responded by building faster accelerators in large clusters. Training deep learning models at petaFLOPS scale requires overcoming both algorithmic and systems software challenges. In this paper, we discuss three systems-related optimization…
Sparse deep neural networks(DNNs) are efficient in both memory and compute when compared to dense DNNs. But due to irregularity in computation of sparse DNNs, their efficiencies are much lower than that of dense DNNs on regular parallel hardware such as TPU. This inefficiency leads to poor/no performance benefits for s…
We describe TF-Replicator, a framework for distributed machine learning designed for DeepMind researchers and implemented as an abstraction over TensorFlow. TF-Replicator simplifies writing data-parallel and model-parallel research code. The same models can be effortlessly deployed to different cluster architectures (i…
Neural Tangents is a library designed to enable research into infinite-width neural networks. It provides a high-level API for specifying complex and hierarchical neural network architectures. These networks can then be trained and evaluated either at finite-width as usual or in their infinite-width limit. Infinite-wid…
By introducing several improvements to the AlphaZero process and architecture, we greatly accelerate self-play learning in Go, achieving a 50x reduction in computation over comparable methods. Like AlphaZero and replications such as ELF OpenGo and Leela Zero, our bot KataGo only learns from neural-net-guided Monte Carl…
BlackJAX simplifies Bayesian inference with modular, fast implementations.
problem Efficient and modular Bayesian inference methods for complex models.
method Functional approach to implementing sampling and variational inference algorithms in JAX.
result Ease of use, speed, and modularity in implementing Bayesian inference.
VeLO learns versatile optimizers from deep learning tasks.
problem Training deep learning models with hand-designed optimizers.
method Meta-training a neural network optimizer on a wide variety of optimization tasks.
result The learned optimizer automatically adapts to different optimization tasks without hyperparameter tuning.
Efficient algorithm finds fast Transformer models.
problem Slow inference time of Transformer models.
method Decompose Transformer architecture into components, use sampling-based one-shot search.
result Achieved 10% to 30% speedup on pre-trained BERT and 70% on top of a previous state-of-the-art model.
Can we automatically design a Convolutional Network (ConvNet) with the highest image classification accuracy under the latency constraint of a mobile device? Neural Architecture Search (NAS) for ConvNet design is a challenging problem due to the combinatorially large design space and search time (at least 200 GPU-hours…
Neural network quantization is becoming an industry standard to efficiently deploy deep learning models on hardware platforms, such as CPU, GPU, TPU, and FPGAs. However, we observe that the conventional quantization approaches are vulnerable to adversarial attacks. This paper aims to raise people's awareness about the …
Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performan…
To make music composition more approachable, we designed the first AI-powered Google Doodle, the Bach Doodle, where users can create their own melody and have it harmonized by a machine learning model Coconet (Huang et al., 2017) in the style of Bach. For users to input melodies, we designed a simplified sheet-music ba…
Sparsity helps reduce the computational complexity of deep neural networks by skipping zeros. Taking advantage of sparsity is listed as a high priority in next generation DNN accelerators such as TPU. The structure of sparsity, i.e., the granularity of pruning, affects the efficiency of hardware accelerator design as w…
New attacks exploit neural network energy and latency, increasing costs by 10-200x.
problem Energy and latency vulnerabilities in neural networks.
method Crafted sponge examples to maximize energy and latency.
result Adversaries can increase energy consumption by 10-200x and delay decisions.
Large, pre-trained generative models have been increasingly popular and useful to both the research and wider communities. Specifically, BigGANs a class-conditional Generative Adversarial Networks trained on ImageNet---achieved excellent, state-of-the-art capability in generating realistic photos. However, fine-tuning …
Proposes differentiable and sparse top-k operators for neural networks.
problem Discontinuity of top-k operator makes it unsuitable for end-to-end training with backpropagation.
method Formulates top-k as a linear program over permutahedron, introduces p-norm regularization, and uses isotonic optimization.
result Successfully applied to neural network pruning, fine-tuning, and routing.
Batch-splitting (data-parallelism) is the dominant distributed Deep Neural Network (DNN) training strategy, due to its universal applicability and its amenability to Single-Program-Multiple-Data (SPMD) programming. However, batch-splitting suffers from problems including the inability to train very large models (due to…
We present Noisy Student Training, a semi-supervised learning approach that works well even when labeled data is abundant. Noisy Student Training achieves 88.4% top-1 accuracy on ImageNet, which is 2.0% better than the state-of-the-art model that requires 3.5B weakly labeled Instagram images. On robustness test sets, i…
Can we automatically design a Convolutional Network (ConvNet) with the highest image classification accuracy under the runtime constraint of a mobile device? Neural architecture search (NAS) has revolutionized the design of hardware-efficient ConvNets by automating this process. However, the NAS problem remains challen…
New method uses MCTS only at test time for faster game learning.
problem High computational demands in game learning.
method Combines MCTS with TD learning agents for faster, reproducible agents.
result First learning-from-scratch agent to beat Edax up to level 7.
This study shows how trade policy uncertainty affects stock-T bill correlations.
problem The impact of trade policy uncertainty on stock-T bill relationships.
method Extended Dynamic Conditional Correlation (DCC) framework incorporating exogenous variables.
result Trade policy uncertainty significantly alters stock-T bill correlations, especially under specific political conditions.
GOBO compresses 99.9% of BERT model parameters to 3 bits, improving inference efficiency.
problem Efficient execution of attention-based NLP models, especially in terms of latency and energy consumption.
method GOBO quantizes 32-bit floating-point parameters to 3 bits without fine-tuning, using hardware compression and co-designed architectures.
result GOBO maintains model accuracy while significantly reducing inference latency and energy consumption.
ProtTrans models predict protein features without evolutionary info.
problem Predicting protein features from amino acid sequences.
method Self-supervised deep learning on large protein datasets.
result ProtT5 embeddings outperform state-of-the-art for per-residue predictions.
New asynchronous SGD algorithms achieve optimal performance in distributed learning.
problem Asynchronous training introduces staleness, complicating optimization analysis.
method Developed rigorous framework for asynchronous first order stochastic optimization.
result Asynchronous SGD can achieve optimal time complexity, matching synchronous methods.
Can we reduce the search cost of Neural Architecture Search (NAS) from days down to only few hours? NAS methods automate the design of Convolutional Networks (ConvNets) under hardware constraints and they have emerged as key components of AutoML frameworks. However, the NAS problem remains challenging due to the combin…
Deep neural networks have yielded superior performance in many applications; however, the gradient computation in a deep model with millions of instances lead to a lengthy training process even with modern GPU/TPU hardware acceleration. In this paper, we propose AutoAssist, a simple framework to accelerate training of …
Improves matrix multiplication throughput for asymmetric bit-width operands.
problem Matrix multiplications between asymmetric bit-width operands, especially 8- and 4-bit, are not efficiently handled by existing SIMD instructions.
method Proposes a new SIMD matrix multiplication instruction that uses mixed precision on inputs (8- and 4-bit) and accumulates into 16-bit output, improving throughput.
result Offers 2x improvement in throughput compared to existing symmetric-operand-size instructions, with negligible overflow.