PARMESAN learns from memory without parameters for fast, efficient continual learning.
problem Inflexibility in deep learning methods for continual learning.
method Transductive reasoning and memory search for parameter-free learning.
result 3-4 orders of magnitude faster than baselines, comparable performance.
Memory-Augmented Meta-Optimization improves cold-start recommendation.
problem Cold-start problem in recommender systems for new users or items.
method Memory-Augmented Meta-Optimization approach with personalized and task-specific memories.
result Significant improvement in cold-start recommendation performance on multiple datasets.
Deep neural networks estimate long memory parameters efficiently.
problem Estimating long memory parameters in stochastic processes.
method Scale-invariant 1D Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) models trained with synthetic data.
result Neural models outperform conventional methods in precision, speed, consistency, and robustness.
NeuroMemFPP uses LSTM to estimate FPP parameters with high accuracy.
problem Estimating parameters of fractional Poisson process with memory and long-range dependence.
method Recurrent Neural Network (RNN), specifically Long Short-Term Memory (LSTM), for parameter estimation.
result The LSTM-based approach reduces MSE by about 55.3% compared to traditional MOM method.
New model captures long-term memory effects in epidemic dynamics.
problem Identifying memory effects in disease progression and recovery.
method Physics-informed neural networks (PINN) with fractional SEIRD model.
result Fractional memory order α improves predictive performance over classical models. 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.…
Paper proposes a GPU-based system for training massive deep learning models in ads systems.
problem Training massive deep learning models with terabyte-scale parameters in ads systems.
method Hierarchical GPU parameter server with 3-layer storage (GPU High-Bandwidth Memory, CPU main memory, SSD).
result 4-node hierarchical GPU parameter server trains a model 2X faster than a 150-node in-memory system.
Memory split advantage: thinner networks outperform a single wide network.
problem Optimizing deep learning models with limited memory.
method Investigated training a single wide network vs. an ensemble of thinner networks with the same total number of parameters.
result An ensemble of several thinner networks outperforms a single wide network for large memory budgets.
Transformers can store facts efficiently using associative memories.
problem Understanding how transformers store and recall factual information.
method Proved linear scaling of storage capacities for linear and MLP associative memories, introduced a synthetic task, and analyzed gradient flow.
result Shallow transformers can achieve near optimal storage capacity for factual recall tasks using associative memories.
Augmenting a neural network with memory that can grow without growing the number of trained parameters is a recent powerful concept with many exciting applications. We propose a design of memory augmented neural networks (MANNs) called Labeled Memory Networks (LMNs) suited for tasks requiring online adaptation in class…
Compact DNNs increase memory footprint and reduce energy efficiency.
problem Designing compact deep neural networks (DNNs) for improved energy efficiency.
method Evaluation of recently proposed compact DNNs on a Tesla P100 GPU.
result Higher number of activations and memory footprint lead to reduced energy efficiency.
Large deep learning models offer significant accuracy gains, but training billions to trillions of parameters is challenging. Existing solutions such as data and model parallelisms exhibit fundamental limitations to fit these models into limited device memory, while obtaining computation, communication and development …
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.
Combinatorial dimensions play an important role in the theory of machine learning. For example, VC dimension characterizes PAC learning, SQ dimension characterizes weak learning with statistical queries, and Littlestone dimension characterizes online learning. In this paper we aim to develop combinatorial dimensions th…
We augment recurrent neural networks with an external memory mechanism that builds upon recent progress in metalearning. We conceptualize this memory as a rapidly adaptable function that we parameterize as a deep neural network. Reading from the neural memory function amounts to pushing an input (the key vector) throug…
NPAS trains neural networks with a fixed parameter budget, improving performance and compactness.
problem Training neural networks requires memory, and existing methods struggle with arbitrary parameter budgets.
method NPAS learns to share parameters automatically, covering low and high budgets.
result NPAS and SSNs improve network performance and compactness across various tasks.
Survey on statistical inference under memory constraints.
problem Effect of memory limitations on statistical inference performance.
method Review of state-of-the-art in several canonical problems.
result Identification of fundamental building blocks and useful techniques.
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.
A new method learns from multi-modal sequences with external memory.
problem Learning new modes in a dynamic environment without prior knowledge.
method Maintains a neural episodic memory with a Dirichlet Process prior to store mode descriptors and transfers knowledge through retrieval.
result Performs continual learning favorably compared to mainstream approaches.
The fully connected layers of a deep convolutional neural network typically contain over 90% of the network parameters, and consume the majority of the memory required to store the network parameters. Reducing the number of parameters while preserving essentially the same predictive performance is critically important …
The properties of statistical tests for hypotheses concerning the parameters of the multifractal model of asset returns (MMAR) are investigated, using Monte Carlo techniques. We show that, in the presence of multifractality, conventional tests of long memory tend to over-reject the null hypothesis of no long memory. Ou…
Vanishing long-term gradients are a major issue in training standard recurrent neural networks (RNNs), which can be alleviated by long short-term memory (LSTM) models with memory cells. However, the extra parameters associated with the memory cells mean an LSTM layer has four times as many parameters as an RNN with the…
New neural networks model complex phenomena with fewer parameters.
problem Challenges in studying higher-order interactions in neural networks.
method Introducing curved neural networks using the maximum entropy principle.
result Curved neural networks accelerate memory retrieval and exhibit explosive phase transitions.
CoNNTrA trains DNNs with low-power, low-memory constraints.
problem Training deep neural networks on edge computing systems with low power and memory usage.
method Coordinate gradient descent-based approach for training DNNs with constrained learning parameters.
result CoNNTrA models use 32x less memory and have comparable errors to Backpropagation models.
A new method reduces memory requirements for sorting high-dimensional data.
problem Efficiently sorting and organizing high-dimensional data with low memory usage.
method Iteratively shuffles N indices and applies SoftSort optimization steps.
result Significantly improves sorting quality for multidimensional data.
The paper studies scaling laws for associative memory mechanisms.
problem Understanding and optimizing learning and memorization processes.
method High-dimensional matrices of outer products of embeddings, relating to transformer models. Derived scaling laws with sample and parameter sizes. Extensive numerical experiments.
result Precise scaling laws and statistical efficiency of estimators.
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…
Embedding representations power machine intelligence in many applications, including recommendation systems, but they are space intensive -- potentially occupying hundreds of gigabytes in large-scale settings. To help manage this outsized memory consumption, we explore mixed dimension embeddings, an embedding layer arc…
In this work we propose a new class of long-memory models with time-varying fractional parameter. In particular, the dynamics of the long-memory coefficient, d, is specified through a stochastic recurrence equation driven by the score of the predictive likelihood, as suggested by Creal et al. (2013) and Harvey (2013)…
Deep neural networks have excelled on a wide range of problems, from vision to language and game playing. Neural networks very gradually incorporate information into weights as they process data, requiring very low learning rates. If the training distribution shifts, the network is slow to adapt, and when it does adapt…
NeuZip compresses neural network weights to save memory during training and inference.
problem Memory constraints in neural network training and inference.
method Entropy-based dynamic weight compression.
result Significant reduction in memory usage without performance loss.
Paper compresses ViTs models by 60% with minimal accuracy loss.
problem Memory constraints in ViTs models.
method Activation-aware low-rank tensor approximations.
result 60% reduction in model size with <1% accuracy loss.
HOPE improves SSMs for long-memory tasks with robust initialization and training.
problem Improving state-space models for long-memory tasks with robust initialization and training.
method Developed a new parameterization scheme called HOPE using Hankel operators and Markov parameters.
result HOPE improves SSMs' performance on Long-Range Arena tasks and demonstrates non-decaying memory.
This work investigates power laws in deep neural network ensembles and predicts their performance.
problem Understanding the performance of deep neural network ensembles and their optimal structure.
method Investigated the behavior of negative log-likelihood (CNLL) of a deep ensemble as a function of ensemble size and member network size, identifying power law dependencies.
result One large network may perform worse than an ensemble of several medium-size networks, known as a memory split.
We propose two optimization techniques to minimize memory usage and computation while meeting system timing constraints for real-time classification in wearable systems. Our method derives a hierarchical classifier structure for Support Vector Machine (SVM) in order to reduce the amount of computations, based on the pr…
Widely popular transformer-based NLP models such as BERT and Turing-NLG have enormous capacity trending to billions of parameters. Current execution methods demand brute-force resources such as HBM devices and high speed interconnectivity for data parallelism. In this paper, we introduce a new relay-style execution tec…
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.
Earlier we proposed the stochastic point process model, which reproduces a variety of self-affine time series exhibiting power spectral density S(f) scaling as power of the frequency f and derived a stochastic differential equation with the same long range memory properties. Here we present a stochastic differential eq…
NeuroMem-FHP framework estimates FHP parameters efficiently.
problem Estimating parameters of fractional Hawkes process (FHP) with long-range dependence.
method Developed LSTM and Transformer neural architectures to estimate FHP parameters directly from inter-arrival times.
result Transformer achieves highest estimation accuracy (MSE = 0.1634) compared to classical MLE (MSE = 2.8032).
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.
Memory-based models can learn to approximate Bayes-optimal predictors for non-stationary data.
problem Learning from non-stationary data with unobserved switching points.
method Memory-based neural models, including Transformers, LSTMs, and RNNs, trained to minimize log loss.
result Memory-based models can accurately approximate known Bayes-optimal algorithms and perform Bayesian inference over latent switching points.
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.
Extended LSTM improves language modeling performance.
problem Improving LSTM for large-scale language modeling.
method Exponential gating, modified memory structure, and residual stacking.
result xLSTM outperforms state-of-the-art models in performance and scaling.
KM method reduces ConvNet parameters to 9% higher accuracy with minimal additional memory.
problem Expensive memory usage for training ConvNets on embedded devices.
method Kernel Modulation (KM) method that adapts all network parameters for each task.
result KM delivers up to 9% higher accuracy than other parameter-efficient methods.
Trojans can be inserted into deep neural networks without altering training data.
problem Inserting malicious code into deep learning models without changing training data.
method Patches model parameters in memory to achieve desired behavior on specific inputs.
result Efficient patches can induce trojan behavior with minimal risk of detection.
In several recently proposed stochastic optimization methods (e.g. RMSProp, Adam, Adadelta), parameter updates are scaled by the inverse square roots of exponential moving averages of squared past gradients. Maintaining these per-parameter second-moment estimators requires memory equal to the number of parameters. For …
GraphGP: Scalable Gaussian Processes with Vecchia's Approximation
problem Naive Gaussian Process computation limits practical use
method GPU algorithm for Vecchia's approximation
result Linear time and memory requirements for nearly a billion parameters
Paper studies statistical-computational trade-offs in tensor PCA and related problems.
problem Statistical-computational gap in tensor PCA estimation.
method Derives computational lower bounds using communication complexity.
result Lower bounds specify trade-off among passes, sample size, and memory.