New techniques improve 16-bit training accuracy without 32-bit units.
problem Training deep learning models with only 16-bit floating-point units.
method Studied BFloat16 units and applied stochastic rounding and Kahan summation techniques.
result Up to 7% absolute validation accuracy gain in 16-bit-FPU training.
Recently, the posit numerical format has shown promise for DNN data representation and compute with ultra-low precision ([5..8]-bit). However, majority of studies focus only on DNN inference. In this work, we propose DNN training using posits and compare with the floating point training. We evaluate on both MNIST and F…
The state-of-the-art hardware platforms for training Deep Neural Networks (DNNs) are moving from traditional single precision (32-bit) computations towards 16 bits of precision -- in large part due to the high energy efficiency and smaller bit storage associated with using reduced-precision representations. However, un…
Low-precision DNNs have been extensively explored in order to reduce the size of DNN models for edge devices. Recently, the posit numerical format has shown promise for DNN data representation and compute with ultra-low precision in [5..8]-bits. However, previous studies were limited to studying posit for DNN inference…
Reduced precision computation for deep neural networks is one of the key areas addressing the widening compute gap driven by an exponential growth in model size. In recent years, deep learning training has largely migrated to 16-bit precision, with significant gains in performance and energy efficiency. However, attemp…
Training of large-scale deep neural networks is often constrained by the available computational resources. We study the effect of limited precision data representation and computation on neural network training. Within the context of low-precision fixed-point computations, we observe the rounding scheme to play a cruc…
The high computational complexity associated with training deep neural networks limits online and real-time training on edge devices. This paper proposed an end-to-end training and inference scheme that eliminates multiplications by approximate operations in the log-domain which has the potential to significantly reduc…
Paper applies FloatSD8 to LSTM networks, reducing complexity and power.
problem Training and inference complexity of LSTM networks.
method Applied FloatSD8 for weights, 8-bit quantization for gradients/activations, reduced arithmetic precision.
result Successfully trained LSTM models with reduced complexity and preserved accuracy.
Secure Aggregation protocols allow a collection of mutually distrust parties, each holding a private value, to collaboratively compute the sum of those values without revealing the values themselves. We consider training a deep neural network in the Federated Learning model, using distributed stochastic gradient descen…
Paper explores reducing precision in SVM for faster text classification.
problem Efficiency in multi-class text classification training.
method Comparison of SVM trained with reduced precision (16-bit, half) vs original.
result Reduced precision training maintains text classification accuracy.
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 neura…
We propose a new approach to combine Restricted Boltzmann Machines (RBMs) that can be used to solve combinatorial optimization problems. This allows synthesis of larger models from smaller RBMs that have been pretrained, thus effectively bypassing the problem of learning in large RBMs, and creating a system able to mod…
This paper proposes a learning framework for n-bit quantized neural networks that improves accuracy and speed on FPGAs.
problem Efficiently implementing quantized neural networks on FPGAs to maintain accuracy and speed.
method A novel learning framework for n-bit QNNs, constrained weights, reconstructed gradient function, n-BQ-NN structure, and SVPE array.
result Quantized models achieve almost the same accuracy as full-precision models and outperform typical low-precision QNNs.
A system for federated learning with private data, adding discrete Gaussian noise and secure aggregation.
problem Training models on private data distributed across devices while ensuring privacy.
method Discretizes data, adds discrete Gaussian noise, and uses secure aggregation to protect privacy.
result Matches the accuracy of central differential privacy with less than 16 bits of precision per value.
Hardware-accelerated RBM solves large combinatorial problems and integer factorization.
problem Solving large combinatorial optimization and integer factorization problems.
method Logically synthesized RBM architecture, hardware acceleration, and efficient training methods.
result Hardware-accelerated RBM factorizes 16-bit numbers with 10000x speed and 32x power improvements.
We present an overview of techniques for quantizing convolutional neural networks for inference with integer weights and activations. Per-channel quantization of weights and per-layer quantization of activations to 8-bits of precision post-training produces classification accuracies within 2% of floating point networks…
Quantized-TinyLLaVA reduces communication costs in split learning for multimodal models.
problem High communication costs in split learning for multimodal models.
method Integrates a compression module that quantizes intermediate features into discrete representations before transmission.
result Achieves an approximate 87.5% reduction in communication overhead with 2-bit quantization.
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.
TinyCNN accelerates CNN models on embedded FPGA with 15x speedup.
problem Limited memory on embedded FPGAs restricts CNN performance.
method Software and hardware design tool for FPGA resource-aware CNN accelerator.
result 3% accuracy loss with 15.75x speedup on image classification.
State of the art deep learning models have made steady progress in the fields of computer vision and natural language processing, at the expense of growing model sizes and computational complexity. Deploying these models on low power and mobile devices poses a challenge due to their limited compute capabilities and str…
Taylorized training improves neural network training at finite width.
problem Understanding and improving neural network training at finite width.
method Training the k-th order Taylor expansion of the neural network at initialization.
result Taylorized training agrees with full neural network training better as k increases and can significantly close the performance gap.
Self-training outperforms pre-training on COCO object detection and segmentation datasets.
problem The effectiveness of pre-training in improving object detection and segmentation models is limited.
method Investigated self-training as an alternative method to utilize additional data.
result Self-training consistently improves model performance across various dataset sizes and data augmentation levels.
Un-trained neural networks outperform trained methods in MRI reconstruction.
problem Accelerated MRI reconstruction with minimal training data.
method Variation of Deep Decoder without training data.
result Un-trained approach significantly outperforms other methods in reconstruction accuracy.
Meta-learning performance depends on train-validation split type.
problem Understanding the importance of train-validation split in meta-learning.
method Theoretical and experimental study comparing train-val and train-train methods.
result Train-train method can achieve strictly better excess loss in realizable cases.
New method reveals how training data influence diffusion model outputs.
problem Difficulty in assessing training data impact on diffusion model outputs.
method Use of ensembles trained on carefully engineered splits of training data to identify influential training examples.
result Demonstrated the viability of ensembles as generative models and validity of assessing influence.
Paper shows adversarial training can be fooled by new type of noise.
problem Adversarial training can be fooled by new types of noise.
method Designing ADVIN, a new type of inducing noise.
result ADVIN can degrade adversarial training robustness by 99.9%.
New MIP methods improve training of integer-valued neural networks.
problem Training integer-valued neural networks with limited data and resources.
method Formulated new MIP models to optimize training efficiency and handle more data.
result Significantly outperforms previous state-of-the-art methods in accuracy, training time, and data usage.
Free adversarial training reduces the generalization gap compared to vanilla method.
problem Improving generalization in adversarial training.
method Analysis of algorithmic stability in free adversarial training.
result Free adversarial training shows a lower generalization gap.
Making neural networks robust against adversarial inputs has resulted in an arms race between new defenses and attacks. The most promising defenses, adversarially robust training and verifiably robust training, have limitations that restrict their practical applications. The adversarially robust training only makes the…
learn2mix trains neural nets faster by adjusting class proportions dynamically.
problem Training neural nets efficiently with limited resources and imbalanced classes.
method Adaptive class proportion adjustment during training.
result Neural nets trained with learn2mix converge faster than static methods.
Loss-guided training accelerates node embedding methods on graphs.
problem Training efficiency in graph learning methods with implicit positive examples.
method Dynamic adjustment of training distribution based on loss values.
result Significant acceleration in training and computation over static methods.
Paper explores fast adversarial training to improve robustness with less computation.
problem Efficiently defending against adversarial examples.
method Integrates simple self-attacks for faster training, focusing on overfitting recovery.
result Shows superior robust accuracy with reduced training time compared to strong adversarial training.
Detects backdoors in trained models without poisoned training data.
problem Detecting backdoors in DNNs trained without access to the poisoned training set.
method Proposes a novel detector using the maximum achievable misclassification fraction (MAMF) statistic.
result Detects backdoors and infers source and target classes.
Stochastic gradient decent~(SGD) and its variants, including some accelerated variants, have become popular for training in machine learning. However, in all existing SGD and its variants, the sample size in each iteration~(epoch) of training is the same as the size of the full training set. In this paper, we propose a…
Crowdsourced training of large neural networks with decentralized Mixture-of-Experts.
problem Expensive training of large neural networks limits research contributions.
method Learning@home: decentralized Mixture-of-Experts for large, poorly connected participants.
result Performance and reliability of Learning@home surpass conventional distributed training.
Pre-training is crucial for learning deep neural networks. Most of existing pre-training methods train simple models (e.g., restricted Boltzmann machines) and then stack them layer by layer to form the deep structure. This layer-wise pre-training has found strong theoretical foundation and broad empirical support. Howe…
New method reconstructs significant parts of training data from neural networks.
problem Understanding and reconstructing training data from neural networks.
method Proposes a novel reconstruction scheme based on recent theoretical results about neural network training.
result Shows that a significant fraction of training data can be reconstructed from neural network parameters.
The paper studies the asymptotic behavior of adversarial training under ℓ∞-perturbation.
problem Theoretical guarantees for sparsity-recovery in adversarial training.
method Investigation of the asymptotic distribution of the adversarial training estimator in generalized linear models.
result The asymptotic distribution of the adversarial training estimator under ℓ∞-perturbation could have a positive probability mass at 0 when the true parameter is 0. This paper introduces 'General Cyclical Training' for neural networks.
problem Improving training efficiency and performance of neural networks.
method Cyclical training phases with varying hyperparameters, batch sizes, loss functions, and data augmentation.
result Cyclical weight decay, softmax temperature, and gradient clipping enhance model accuracy.
Speeds up training and inference by pruning entire channels before training.
problem Training and inference speed in deep neural networks.
method Structured pruning applied before training, focusing on removing entire channels and hidden units.
result 2x speedup in training and 3x speedup in inference.
The paper predicts loss scaling across different datasets and compute scales.
problem Predicting loss scaling across different datasets and compute scales.
method Derive shifted power law relationships between train and test losses.
result Shifted power law relationships hold for various datasets and tasks, improving prediction accuracy.
Theoretical study shows adversarial training improves robustness in deep learning models.
problem Ensuring robustness in pre-trained deep learning models.
method Theoretical analysis of adversarial training and feature purification in two-layer neural networks.
result Adversarial training leads to feature purification, making models more robust to attacks.
Training dynamics reveal neural network complexity.
problem Understanding the complexity of neural networks during training.
method Relating the Lipschitz constant and training behavior of neural networks.
result Steady training and bounded complexity in far regions of input space.
LIIT uses large deviations to train neural networks faster.
problem Training neural networks requires large amounts of data, leading to inefficiencies.
method LIIT uses large deviations principle to generate and update training samples.
result LIIT achieves similar classification performance with smaller, more representative samples.
Empirical study shows GANs overfit and drop modes when training is deterministic.
problem Understanding overfitting and mode drop in GAN training.
method Empirical analysis of GAN training with and without stochasticity.
result GANs overfit and drop modes when training is deterministic.
Joint training improves model accuracy by selectively using privileged information.
problem Two-stage training can lead to model failure with noisy privileged information.
method Joint training of two models to use privileged information selectively.
result Joint training outperforms two-stage baselines on synthetic and real-world tasks.
ShadowSync separates background synchronization for scalable distributed training.
problem Reducing synchronization overhead in distributed training for high scalability.
method Separates synchronization from training and runs it in the background.
result Achieves both high throughput and excellent model quality at scale.
MACER trains robust models without adversarial training, faster and more effective.
problem Learning robust models without relying on attack-dependent adversarial training.
method MACER trains provably robust smoothed classifiers by maximizing certified radius.
result MACER achieves larger average certified radius and faster training time compared to state-of-the-art methods.