The computation of convolution layers in deep neural networks typically rely on high performance routines that trade space for time by using additional memory (either for packing purposes or required as part of the algorithm) to improve performance. The problems with such an approach are two-fold. First, these routines…
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.
168,657 papers · 148 categories
Trend · papers per month
48 results for “Memory Overhead”
SRG improves optimization efficiency with reduced memory and computation overhead.
problem Limited applicability of variance-reduced optimization algorithms.
method SRG (stochastic reweighted gradient) using importance sampling.
result SRG outperforms SGD and provably improves convergence in strongly-convex cases.
In recent years, memory-augmented neural networks(MANNs) have shown promising power to enhance the memory ability of neural networks for sequential processing tasks. However, previous MANNs suffer from complex memory addressing mechanism, making them relatively hard to train and causing computational overheads. Moreove…
Adaptive gradient-based optimizers such as Adagrad and Adam are crucial for achieving state-of-the-art performance in machine translation and language modeling. However, these methods maintain second-order statistics for each parameter, thus introducing significant memory overheads that restrict the size of the model b…
SPARC improves continual learning with minimal memory and computational overhead.
problem Efficient continual learning for deep neural networks.
method Combines task-specific working memories and task-agnostic semantic memory.
result Significantly reduces parameter usage (6% of full-model surrogates) while maintaining performance.
Clapping reduces memory usage in distributed optimization by reusing data samples.
problem Significant communication overhead and impractical memory overhead in pipeline-parallel distributed optimization.
method Lazy sampling strategy to reuse data samples across steps, supporting convergence without unbiased gradient assumptions.
result Clapping achieves convergence in few-epoch or online training regimes without sample-size memory overhead.
Online method learns sparse models efficiently in large scale settings.
problem Sparse model learning in large scale settings with high computational and memory costs.
method Online learning approach, mini-batch methods, hard thresholding based stochastic gradient algorithm.
result Sparsity promoted by batch methods is not preserved in online fashion.
State-of-the-art models are now trained with billions of parameters, reaching hardware limits in terms of memory consumption. This has created a recent demand for memory-efficient optimizers. To this end, we investigate the limits and performance tradeoffs of memory-efficient adaptively preconditioned gradient methods.…
Quantum systems with scrambling improve temporal information processing, but scaling requires exponential overhead.
problem Scalability and memory retention of quantum reservoirs in temporal information processing.
method Examined a quantum reservoir processing framework with scrambling reservoirs modeled by high-order unitary designs, analyzed in noiseless and noisy settings.
result Memory retention improves exponentially with reservoir size but worsens with reservoir iterations, requiring exponential shot overhead for scaling.
Convolutional Neural Networks (CNN) are being actively explored for safety-critical applications such as autonomous vehicles and aerospace, where it is essential to ensure the reliability of inference results in the presence of possible memory faults. Traditional methods such as error correction codes (ECC) and Triple …
Recomputation algorithms collectively refer to a family of methods that aims to reduce the memory consumption of the backpropagation by selectively discarding the intermediate results of the forward propagation and recomputing the discarded results as needed. In this paper, we will propose a novel and efficient recompu…
HyperINF improves influence function estimation for large models with better accuracy and efficiency.
problem Inaccurate and computationally expensive influence function estimation for large-scale models.
method HyperINF leverages Schulz's iterative algorithm and GFIM for low-rank approximation of Hessian matrix.
result HyperINF achieves superior accuracy and performance compared to existing methods on LoRA-tuned models.
A new method speeds up spectral normalization for neural nets.
problem Efficiently controlling the spectral norm of convolutional layers.
method Depthwise separable convolutions with spectral normalization.
result Significant reduction in computational and memory costs.
GPA improves LLM training speed by 8.71% for Llama-160M models.
problem Training Large Language Models (LLMs) with high memory overhead and slow convergence.
method Generalized Primal Averaging (GPA) extends Nesterov's method to eliminate memory-intensive two-loop structure.
result GPA achieves up to 10.13% speedup over AdamW in training Llama-1B model.
Through the River: Understanding the Benefit of Schedule-Free Methods for Language Model Trainingcs.LG
New method SF-AdamW trains large models without decay phases or memory overhead.
problem Inadequate fixed compute budgets for large-scale training.
method Schedule-Free (SF) method revisited and refined.
result SF-AdamW effectively navigates loss landscape without decay phases or memory overhead.
Global Memory Augmentation (GMAT) improves Transformer performance on long documents.
problem Large memory requirements of Transformer pairwise dot-product attention for long sequences.
method Integrates a dense global memory of length M into sparse Transformer blocks.
result Significant improvement on various tasks, including synthetic tasks, masked language modeling, and reading comprehension.
This work analyzes and optimizes memory and compute costs of learned optimizers.
problem High memory and compute costs of learned optimizers.
method Identified and quantified design features of learned and hand-designed optimizers, constructed a more efficient learned optimizer.
result A learned optimizer that is faster and more memory efficient than previous work.
FPG uses fractional calculus for efficient reinforcement learning with long-term memory.
problem High variance and inefficient sampling in standard policy gradient methods for long-term temporal modeling.
method Fractional Policy Gradients (FPG) incorporating Caputo fractional derivatives for power-law temporal correlations.
result Achieves asymptotic variance reduction of order O(t^(-alpha)) and sample efficiency gains.
Express improves causal attention guarantees for language models.
problem Improving causal attention guarantees for language models.
method Introducing Express, a tool for converting non-causal attention into causal with matching guarantees.
result Express improves causal attention guarantees to log^(3/2)(n)/s with minimal memory and compression overhead.
Paper proposes efficient BNN inference flow to reduce computation and memory costs.
problem High computation complexity in Bayesian Neural Networks (BNNs) limits deployment in power-constrained systems.
method Feature decomposition and memorization strategy to reduce computations and a memory-friendly computing framework to reduce memory overhead.
result Reduces computation by about half and energy consumption by 73% with 14% area overhead.
A new method combines regularization and generative rehearsal for continual learning.
problem Catastrophic forgetting in neural networks over past tasks.
method Uses a normalizing flow to conditionally store past task data and regularize network embeddings.
result Performs favorably compared to state-of-the-art approaches with constant memory overhead.
A new method combines generative models and regularization to prevent forgetting in continual learning.
problem Catastrophic forgetting in neural networks when learning new tasks.
method Uses a normalizing flow as a generative model to keep past data embeddings and regularize them.
result Performs favorably compared to existing methods with constant memory overhead.
Deep neural networks (DNNs) depend on the storage of a large number of parameters, which consumes an important portion of the energy used during inference. This paper considers the case where the energy usage of memory elements can be reduced at the cost of reduced reliability. A training algorithm is proposed to optim…
Paper proposes efficient optimizers for large language models with fast convergence and low memory usage.
problem Designing efficient optimizers for large language models with low-memory requirements and fast convergence.
method Structured Fisher information matrix approximation and low-rank extension framework.
result New optimizers (RACS and Alice) achieve better convergence and lower memory usage than existing methods.
Efficiently discovers Bayesian network structure with reduced memory usage.
problem Bayesian network structure discovery is NP-hard and memory-intensive.
method Progressively leveled scoring approach with hierarchical computation.
result Achieved processing of a 28-variable Bayesian network using only memory.
DarkneTZ protects edge devices from DNN model leaks using TEE and model partitioning.
problem Privacy risks of pre-trained DNNs on edge devices through membership inference attacks.
method Model partitioning into sensitive and untrusted parts, leveraging TEE.
result DarkneTZ provides reliable model privacy with minimal performance overhead.
GPU (graphics processing unit) has been used for many data-intensive applications. Among them, deep learning systems are one of the most important consumer systems for GPU nowadays. As deep learning applications impose deeper and larger models in order to achieve higher accuracy, memory management becomes an important …
Serenity optimizes neural network execution for edge devices by scheduling with optimal memory footprint.
problem Order of nodes in irregular neural networks affects memory footprint, complicating execution under resource constraints.
method Memory-aware compiler using dynamic programming and graph rewriting to find optimal schedules.
result Achieves optimal peak memory and further improves it with graph rewriting, reducing memory usage by 1.68x-1.86x compared to TensorFlow Lite.
AlphaGrad optimizes memory usage in RL algorithms by normalizing gradients.
problem Memory overhead and hyperparameter complexity in adaptive optimizers.
method Tensor-wise L2 normalization followed by a smooth hyperbolic tangent transformation controlled by a single parameter.
result AlphaGrad provides enhanced training stability and competitive performance in various RL algorithms.
A new model learns preferences incrementally without personal data.
problem Incremental session-based recommendation without personal data.
method Memory Augmented Neural model (MAN) that combines a neural recommender with a nonparametric memory.
result MAN consistently outperforms existing methods in incremental session-based recommendation.
Incremental class learning, a scenario in continual learning context where classes and their training data are sequentially and disjointedly observed, challenges a problem widely known as catastrophic forgetting. In this work, we propose a novel incremental class learning method that can significantly reduce memory ove…
Deep Neural Networks(DNNs) require huge GPU memory when training on modern image/video databases. Unfortunately, the GPU memory is physically finite, which limits the image resolutions and batch sizes that could be used in training for better DNN performance. Unlike solutions that require physically upgrade GPUs, the G…
Memory Organization for Energy-Efficient Learning and Inference in Digital Neuromorphic Acceleratorscs.NE
Improves energy efficiency of neuromorphic hardware by optimizing memory organization and encoding schemes.
problem Energy inefficiency in neuromorphic hardware, especially in digital accelerators.
method Synthesized controller and memory for different encoding schemes, introduced functional encoding for structured connectivity.
result Functional encoding offers a 58% reduction in energy for weight updates in convolutional layers.
Index structures are important for efficient data access, which have been widely used to improve the performance in many in-memory systems. Due to high in-memory overheads, traditional index structures become difficult to process the explosive growth of data, let alone providing low latency and high throughput performa…
Generative Adapter adapts LMs with a single forward pass, reducing inference overhead.
problem Efficient adaptation of large language models for new contexts.
method Generative Adapter directly maps new contexts to low-rank LM adapters via self-supervised learning.
result Significant reduction in inference overhead with no need for fine-tuning.
Compact Recurrent Transformer (CRT) improves Transformer efficiency for long sequences.
problem Efficiently scaling Transformer architecture to long sequences with limited compute resources.
method Combines shallow Transformer models with recurrent neural networks and persistent memory.
result CRT achieves comparable or superior performance to full-length Transformers with shorter segments and reduced FLOPs.
The Long-Short-Term-Memory Recurrent Neural Networks (LSTM RNNs) are a popular class of machine learning models for analyzing sequential data. Their training on modern GPUs, however, is limited by the GPU memory capacity. Our profiling results of the LSTM RNN-based Neural Machine Translation (NMT) model reveal that fea…
Modern information technology services largely depend on cloud infrastructures to provide their services. These cloud infrastructures are built on top of datacenter networks (DCNs) constructed with high-speed links, fast switching gear, and redundancy to offer better flexibility and resiliency. In this environment, net…
VPFD uses vocoder features for adversarial training in VC.
problem Adversarial training on waveform data is time-consuming and memory-intensive.
method VPFD employs vocoder features for adversarial training.
result VPFD achieves VC performance comparable to waveform discriminators with reduced training time and memory.
Learning to learn has emerged as an important direction for achieving artificial intelligence. Two of the primary barriers to its adoption are an inability to scale to larger problems and a limited ability to generalize to new tasks. We introduce a learned gradient descent optimizer that generalizes well to new tasks, …
MELO predicts electricity loads by adapting to shifts without external indicators.
problem Adapting to non-stationary prediction challenges in online settings.
method MELO combines multiple forgetting factors and aggregation rules to adaptively predict.
result MELO reduces RMSE by 34.7% compared to base predictors and external covariates.
SAG is a scalable method for adversarial attacks on GNNs.
problem Scalability and robustness of GNNs to adversarial attacks.
method Decomposing large graphs into smaller partitions, using ADMM for optimization.
result SAG reduces computation and memory overhead for large graphs.
While the backpropagation of error algorithm enables deep neural network training, it implies (i) bidirectional synaptic weight transport and (ii) update locking until the forward and backward passes are completed. Not only do these constraints preclude biological plausibility, but they also hinder the development of l…
Recurrent Neural Networks (RNNs) are used in state-of-the-art models in domains such as speech recognition, machine translation, and language modelling. Sparsity is a technique to reduce compute and memory requirements of deep learning models. Sparse RNNs are easier to deploy on devices and high-end server processors. …
Proposes ContSup to boost local learning by supplying context between isolated modules.
problem Local learning's performance degrades with more isolated modules.
method Theoretical analysis and ContSup scheme to supply context between modules.
result Significant performance improvement with minimal overhead.
KineticSim accelerates financial market simulations 3406x over CPU.
problem Simulating financial markets at scale with multi-agent models is bottlenecked by sequential processing and GPU kernel overhead.
method Formalized and implemented a reusable parallel design pattern for iterative multi-agent reductions in thread-block shared memory.
result Achieved a peak throughput of over 54.7 billion agent-events per second, delivering 3406x speedup over CPU.
KineticSim: A lightweight, high-performance execution engine for real-time market simulators
problem Simulating financial markets at scale with multi-agent models
method Reusable parallel design pattern: persistent, state-carrying clearing for iterative multi-agent reductions
result Reduces per-step critical-path depth from Theta(L+A) to Theta(log L + ceil(A/L))
Huge scale machine learning problems are nowadays tackled by distributed optimization algorithms, i.e. algorithms that leverage the compute power of many devices for training. The communication overhead is a key bottleneck that hinders perfect scalability. Various recent works proposed to use quantization or sparsifica…