Paper presents efficient algorithms for convolutional neural networks using Winograd minimal filtering.
problem Resource-efficient implementation of convolutional neural networks.
method Winograd minimal filtering trick applied to M-tap filters (M=3,5,7,9,11) for parallel hardware implementation.
result Approximately 30% reduction in multipliers for fully parallel hardware implementation.
We present a design and implementation of the Thomas algorithm optimized for hardware acceleration on an FPGA, the Thomas Core. The hardware-based algorithm combined with the custom data flow and low level parallelism available in an FPGA reduces the overall complexity from 8N down to 5N serial arithmetic operations, a…
Bio-inspired neuromorphic hardware is a research direction to approach brain's computational power and energy efficiency. Spiking neural networks (SNN) encode information as sparsely distributed spike trains and employ spike-timing-dependent plasticity (STDP) mechanism for learning. Existing hardware implementations of…
PipeMare enables efficient DNN training with minimal memory and pipeline sacrifices.
problem Sacrificing hardware efficiency to maintain statistical efficiency in pipeline parallel DNN training.
method PipeMare is a simple yet robust training method that tolerates asynchronous updates during pipeline parallelism without sacrificing pipeline utilization or memory.
result PipeMare achieves up to 2.7x less memory usage or 4.3x higher pipeline utilization compared to state-of-the-art synchronous PP training techniques.
Recurrent neural networks (RNNs) are omnipresent in sequence modeling tasks. Practical models usually consist of several layers of hundreds or thousands of neurons which are fully connected. This places a heavy computational and memory burden on hardware, restricting adoption in practical low-cost and low-power devices…
PipeDream-2BW accelerates large model training by 20x with minimal memory usage.
problem Training large models requires memory beyond single accelerator capacity.
method Pipeline parallelism, weight gradient coalescing, double buffering.
result Accelerates large model training by up to 20x.
We derive scaling laws for optimizing neural networks in hardware.
problem Optimizing the large parameter space of neural networks in hardware.
method Analytical derivation of scaling laws for Coordinate Descent optimization.
result Convergence is exponential and scales linearly with the number of neurons.
Pipelined Backpropagation trains large models without batches efficiently.
problem Training large models efficiently on hardware with limited batch sizes.
method Fine-grained Pipelined Backpropagation with Spike Compensation and Linear Weight Prediction.
result Fine-grained Pipelined Backpropagation with a batch size of one matches the accuracy of SGD for multiple networks.
As a result of the growing size of Deep Neural Networks (DNNs), the gap to hardware capabilities in terms of memory and compute increases. To effectively compress DNNs, quantization and connection pruning are usually considered. However, unconstrained pruning usually leads to unstructured parallelism, which maps poorly…
A deep-learning inference accelerator is synthesized from a C-language software program parallelized with Pthreads. The software implementation uses the well-known producer/consumer model with parallel threads interconnected by FIFO queues. The LegUp high-level synthesis (HLS) tool synthesizes threads into parallel FPG…
Serial problems can't be efficiently parallelized, affecting machine learning models.
problem Inefficiency of parallelization in inherently serial problems.
method Formalized distinction in complexity theory, demonstrated with diffusion models.
result Diffusion models cannot solve inherently serial problems.
We describe ASAGA, an asynchronous parallel version of the incremental gradient algorithm SAGA that enjoys fast linear convergence rates. Through a novel perspective, we revisit and clarify a subtle but important technical issue present in a large fraction of the recent convergence rate proofs for asynchronous parallel…
New principle reduces load imbalance in LLM serving systems, saving up to 52% energy.
problem Wasted computational power due to load imbalance in LLM serving systems.
method Developed a universal load-balancing principle for barrier-synchronized systems with non-migratable state.
result Proves worst-case theoretical guarantees for imbalance reduction and energy savings.
This paper improves parallel belief propagation for scalable machine learning.
problem Efficient parallelization of belief propagation for large-scale machine learning tasks.
method Use of scalable relaxed schedulers to parallelize belief propagation.
result Our approach outperforms previous methods in scalability and convergence time.
Efficiently quantifies uncertainty in DeepONets for function spaces.
problem Uncertainty quantification in deep operator networks.
method Randomized prior ensembles for frequentist inference.
result Improved robustness and accuracy, reliable uncertainty estimates, out-of-distribution detection, and model bias quantification.
New types of machine learning hardware in development and entering the market hold the promise of revolutionizing deep learning in a manner as profound as GPUs. However, existing software frameworks and training algorithms for deep learning have yet to evolve to fully leverage the capability of the new wave of silicon.…
Paper proposes a new ML approach using only additions and thresholding.
problem Energy efficiency and reduced complexity for IoT ML devices.
method Margin-Propagation (MP) network for inference and learning without MVMs.
result MP-based classifiers achieve comparable results to traditional ML methods with energy savings.
The omnipresence of deep learning architectures such as deep convolutional neural networks (CNN)s is fueled by the synergistic combination of ever-increasing labeled datasets and specialized hardware. Despite the indisputable success, the reliance on huge amounts of labeled data and specialized hardware can be a limiti…
Efficient FPGA dropout algorithm reduces memory usage.
problem Overfitting in deep neural networks.
method Hardware-oriented dropout algorithm for FPGA implementation.
result Significant resource reduction in FPGA implementation.
Packed-Ensembles improve uncertainty estimation in constrained hardware.
problem Hardware limitations restrict the size of ensembles and network capacity, degrading performance.
method Packed-Ensembles (PE) design and train lightweight structured ensembles by modulating encoding space and parallelizing into a single backbone.
result PE accurately preserves diversity and maintains performance on key metrics like accuracy, calibration, and out-of-distribution detection.
Photonic co-processor speeds up training of large neural networks.
problem Training large neural networks with backpropagation is inefficient and communication is a bottleneck.
method Direct Feedback Alignment (DFA) with a photonic accelerator.
result Photonic accelerator can compute random projections with trillions of parameters.
Recent hardware developments have dramatically increased the scale of data parallelism available for neural network training. Among the simplest ways to harness next-generation hardware is to increase the batch size in standard mini-batch neural network training algorithms. In this work, we aim to experimentally charac…
TriTPP models enable faster and more flexible event data modeling.
problem Inflexibility and slow sampling in traditional TPP models.
method Triangular Maps and Normalizing Flows for parallel sampling and likelihood computation.
result TriTPP models achieve orders of magnitude faster sampling while maintaining flexibility.
SELD-TCN improves sound event localization and detection efficiency.
problem Efficient sound event localization and detection on embedded hardware.
method Developed a novel temporal convolutional network (TCN) architecture.
result SELD-TCN outperforms state-of-the-art SELDnet on four datasets.
dSMC improves parallel processing of state-space models.
problem Processing multiple observations efficiently in state-space models.
method A parallel-in-time particle smoother that reduces complexity to log(T).
result dSMC achieves O(log(T)) time complexity on parallel architectures.
Deep neural network has recently shown very promising applications in different research directions and attracted the industry attention as well. Although the idea was introduced in the past but just recently the main limitation of using this class of algorithms is solved by enabling parallel computing on GPU hardware.…
This dissertation advances scalable Gaussian processes using iterative methods and pathwise conditioning.
problem The classical Gaussian process formulation is not scalable for large datasets and modern hardware.
method Combining iterative methods and pathwise conditioning to improve scalability.
result Significantly reduced memory requirements and facilitated application to larger datasets.
MLModelScope streamlines ML/DL model evaluation and benchmarking.
problem Challenges in evaluating and benchmarking ML/DL models.
method Open-source, framework/hardware agnostic platform with distributed design.
result Demonstrates the impact of model evaluation pipelines and HW/SW choices.
We present POLO --- a C++ library for large-scale parallel optimization research that emphasizes ease-of-use, flexibility and efficiency in algorithm design. It uses multiple inheritance and template programming to decompose algorithms into essential policies and facilitate code reuse. With its clear separation between…
Predictability enables efficient parallelization of nonlinear models.
problem Understanding which nonlinear state space models can be efficiently parallelized.
method Established a relationship between system dynamics and optimization problem conditioning, quantified by the largest Lyapunov exponent.
result Predictable systems can be evaluated in O((logT)2) time, improving over conventional sequential approaches. A high-parallelism SNN improves feature learning efficiency and robustness.
problem Slow learning speed and limited learning capability in existing SNNs.
method Inspired by Inception modules, high-parallelism architecture, Vote-for-All decoding, adaptive repolarization mechanism.
result Superior performance and competitive accuracy compared to state-of-the-art unsupervised SNNs.
The implementation of artificial neural networks in hardware substrates is a major interdisciplinary enterprise. Well suited candidates for physical implementations must combine nonlinear neurons with dedicated and efficient hardware solutions for both connectivity and training. Reservoir computing addresses the proble…
As datasets continue to increase in size and multi-core computer architectures are developed, asynchronous parallel optimization algorithms become more and more essential to the field of Machine Learning. Unfortunately, conducting the theoretical analysis asynchronous methods is difficult, notably due to the introducti…
New method learns SDEs without integrators, speeding up computation.
problem Computational expense in learning SDEs using neural networks.
method Importance-sampling estimator for SDEs, leveraging parallelism.
result Lower-variance gradient estimates and massive computation time reductions.
New methods accelerate NCGP inference by trading computation for uncertainty.
problem Prohibitively expensive exact inference in NCGPs for large datasets.
method Iterative methods explicitly modeling approximation error, leveraging parallel computing.
result Significant acceleration of posterior inference compared to baselines.
2D-PT improves sampling in constrained optimization problems.
problem Sampling Boltzmann distributions with soft constraints.
method Two-dimensional extension of parallel tempering.
result 2D-PT achieves near-ideal mixing in constrained problems.
NEST optimizes deep learning training by placing devices efficiently across networks and memory.
problem Inefficient device placement in distributed deep learning leads to high communication and memory overhead.
method NEST uses network-, compute-, and memory-aware dynamic programming to optimize device placement.
result NEST achieves up to 2.43 times higher throughput and better memory efficiency.
A new parallel BO method with exact gradients for multi-objective optimization.
problem Efficiently optimizing multiple objectives in a sample-efficient manner.
method Derive q-Expected Hypervolume Improvement (qEHVI) for parallel, constrained evaluation.
result qEHVI is computationally tractable and outperforms state-of-the-art methods.
The flourish of deep learning frameworks and hardware platforms has been demanding an efficient compiler that can shield the diversity in both software and hardware in order to provide application portability. Among the existing deep learning compilers, TVM is well known for its efficiency in code generation and optimi…
FANN-on-MCU enables efficient neural network inference on IoT devices.
problem Energy-efficient neural network inference on resource-constrained IoT devices.
method Open-source toolkit for ARM Cortex-M and RISC-V microcontrollers.
result Efficient neural network execution with low latency and power consumption.
Modern machine learning models are typically trained using Stochastic Gradient Descent (SGD) on massively parallel computing resources such as GPUs. Increasing mini-batch size is a simple and direct way to utilize the parallel computing capacity. For small batch an increase in batch size results in the proportional red…
Improved time complexity for parallel stochastic optimization in heterogeneous systems.
problem Time complexity in parallel stochastic optimization for large-scale machine learning models.
method Proposes Rennala MVR, a variance-reduced extension of Rennala SGD based on momentum-based variance reduction.
result Variance reduction improves time complexity in relevant parameter regimes for parallel stochastic optimization in heterogeneous systems.
New method trains sparse Gaussian processes without matrix inversion.
problem Costly training of Gaussian processes at scale.
method Inverse-free approach using matmul-only natural-gradient updates.
result Significantly improved stability and convergence in training.
Efficient kernel methods for large datasets using GPU acceleration.
problem Handling large-scale nonparametric learning problems efficiently.
method Preconditioned gradient solver, GPU acceleration, parallelization, out-of-core linear algebra, numerical precision optimization.
result Dramatic speedups on datasets with billions of points, maintaining state-of-the-art performance.
Machine Learning (ML) is making a strong resurgence in tune with the massive generation of unstructured data which in turn requires massive computational resources. Due to the inherently compute- and power-intensive structure of Neural Networks (NNs), hardware accelerators emerge as a promising solution. However, with …
This paper uses supervised learning to predict optimal chunk-size for parallel linear algebra operations.
problem Finding the optimal chunk-size for parallel linear algebra operations.
method The paper uses supervised learning models (logistic regression, neural networks, decision trees) to predict the optimal chunk-size for multiple linear algebra operations.
result The custom decision tree model outperforms classical decision trees and other models in predicting optimal chunk-size for linear algebra operations.
Training deep networks is expensive and time-consuming with the training period increasing with data size and growth in model parameters. In this paper, we provide a framework for distributed training of deep networks over a cluster of CPUs in Apache Spark. The framework implements both Data Parallelism and Model Paral…
Deep network solves maze path planning without training.
problem Efficient path planning in large mazes with obstacles.
method Max pooling layers without training.
result Solves mazes with over half a billion nodes in short time.