DAM with MRL improves relational reasoning in MANNs.
problem Limited performance of associative memory networks on complex relational reasoning tasks.
method Distributed Associative Memory architecture with Memory Refreshing Loss.
result Enhanced relation reasoning performance of MANNs on long temporal sequence data.
LiDER refreshes past experiences in RL by dreaming about them.
problem Improving data efficiency in off-policy RL algorithms.
method Refreshing past experiences in a replay buffer using the current policy.
result LiDER consistently improves performance in Atari games.
Memory-efficient optimizers fail to track a subspace, leading to unpredictable model performance.
problem Memory-efficient optimizers fail to track a subspace, leading to unpredictable model performance.
method Analyzing the behavior of memory-efficient optimizers like GaLore, which project gradients onto a rank-r subspace recomputed every T steps.
result Memory-efficient optimizers fail to track a subspace, leading to unpredictable model performance.
Improved efficiency in HMC samplers reduces dissipative behavior.
problem Reducing dissipative behavior in HMC samplers.
method Variable integration time and partial velocity refreshment.
result Efficiency improved by a √κ factor in Wasserstein-2 distance.
A new method for efficiently updating large-scale matrices in real-time.
problem Updating large-scale matrices with evolving data in real-time.
method Incremental SVD approach that handles row/column appends, rank-1 updates, and refresh strategies.
result Incremental SVD achieves accuracy close to full SVD with a fraction of the computational cost.
Extends FC-RAG to anytime-valid sequential coverage for language model swarms.
problem Maintain distribution-free coverage for a swarm of weak language models over time.
method Introduces Anytime-FC-RAG, a sequential extension with a summable calibration-deviation budget.
result Achieves time-uniform alarm validity and safety under predictable adaptive control.
New method reduces Bayesian inference time by optimizing Hamiltonian flows.
problem Efficiently constructing a Bayesian coreset with low inferential error.
method Sparse Hamiltonian flows, involving uniform subsampling and momentum quasi-refreshment steps.
result Exponential compression of the dataset and reduced KL divergence to the target.
We propose a max-pooling based loss function for training Long Short-Term Memory (LSTM) networks for small-footprint keyword spotting (KWS), with low CPU, memory, and latency requirements. The max-pooling loss training can be further guided by initializing with a cross-entropy loss trained network. A posterior smoothin…
TentacleNet improves binarized CNNs, reducing accuracy loss and memory usage.
problem Excessive accuracy loss in binarized CNNs.
method Parallelization inspired by ensemble learning theory, end-to-end trainable compact topology.
result Significant memory savings compared to state-of-the-art binary ensemble methods.
Enhanced Hopfield model boosts memory retrieval capacity.
problem Memory retrieval in modern Hopfield models with limited capacity.
method Introduces a learnable feature map transforming energy function into kernel space, minimizing separation loss for uniform memory distribution.
result Significant reduction in metastable states, enhancing memory capacity and retrieval accuracy.
New insights link memory loss to system stability in dynamical systems.
problem Understanding the relationship between dynamics and computation, especially stability and memory loss.
method Analyzing driven dynamical systems responding to temporal inputs.
result Memory loss in driven systems leads to consistent responses to similar inputs and affects stability.
Extended LSTMs improve volatility prediction by 20%.
problem Predicting asset price volatility with long memory.
method Extended LSTMs with multiple flexible timescales.
result Extended LSTMs outperform rough volatility predictions by 20%.
A technique identifies memoryless algorithms approximating memory-dependent optimization methods.
problem Understanding how memory in optimization algorithms affects loss and generalization.
method Introducing a general technique to replace past iterates with the current one and adding a correction term.
result Lion does not have the same implicit anti-regularization as AdamW, explaining its better generalization performance.
Study non-oblivious adversarial bandits with delayed feedback and propose algorithms with improved regret bounds.
problem Adversarial bandit problem with delayed, composite anonymous feedback.
method Propose wrapper algorithm for non-oblivious delay setting, achieving o ( T ) o(T) o ( T ) policy regret. result Achieve o ( T ) o(T) o ( T ) policy regret for many adversarial bandit problems with bounded memory loss sequences. New framework captures long-term decision dependence in online learning.
problem Long-term dependence on past decisions in online learning.
method Introduces Online Convex Optimization with Unbounded Memory (OCO-UMB) and p p p -effective memory capacity. result Proves O ( H p T ) O(\sqrt{H_p T}) O ( H p T ) upper bound on policy regret and matching lower bound. Portfolio allocation with gross-exposure constraint is an effective method to increase the efficiency and stability of selected portfolios among a vast pool of assets, as demonstrated in Fan et al (2008). The required high-dimensional volatility matrix can be estimated by using high frequency financial data. This enabl…
This paper proposes an alternative to E2E training for deep networks, reducing memory footprint.
problem High GPUs memory footprint in end-to-end training of deep networks.
method Locally supervised learning with information propagation loss to avoid information collapse.
result The proposed method achieves competitive performance with less than 40% memory footprint compared to E2E training.
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.
Second-order optimizers retain residual information after data deletion, affecting machine unlearning.
problem Residual information in second-order optimizers after data deletion.
method Comparison of first-order and second-order learners, eigendecomposition analysis.
result Second-order optimizers retain residual information, not detectable by first-order analysis.
The paper explores how gradient descent trains associative memories, revealing oscillations and convergence issues.
problem Training dynamics of associative memories in overparameterized and underparameterized settings.
method Reduction to particle system dynamics, theory, and experiments.
result Oscillatory transitory regimes and benign loss spikes in overparameterized settings, suboptimal memorization in underparameterized settings.
RNN-HAR model improves VaR forecasting with long-memory and non-linear dynamics.
problem Efficiently forecasting Value at Risk (VaR) with long-memory and non-linear realized volatility.
method Loss-based generalized Bayesian inference with Sequential Monte Carlo for model estimation and prediction.
result RNN-HAR model consistently outperforms other VaR forecasting models.
New algorithms for constrained online optimization with memory and predictions.
problem Control of constrained dynamical systems and scheduling with reconfiguration budgets.
method Proposed algorithms achieving sublinear regret and constraint violation under time-varying constraints, both with and without predictions.
result First algorithms achieving sublinear regret and constraint violation in constrained online optimization with memory.
Novel method learns memory kernels in Langevin equations.
problem Estimating memory kernels in Langevin equations.
method Regularized Prony method for correlation functions, followed by regression over Sobolev norm-based loss function with RKHS regularization.
result Method outperforms other regression estimators in exponentially weighted L^2 space.
AMSAs adaptively manage crypto-currency trading by selecting multiple strategies based on market conditions.
problem Maximizing gains in volatile crypto-currency markets with high uncertainty.
method AMSAs use multiple sub-agents with different strategies, dynamically selecting them based on market conditions.
result AMSAs can achieve high positive alpha in long-term crypto-currency trading.
Learning long-term dependencies is a key long-standing challenge of recurrent neural networks (RNNs). Hierarchical recurrent neural networks (HRNNs) have been considered a promising approach as long-term dependencies are resolved through shortcuts up and down the hierarchy. Yet, the memory requirements of Truncated Bac…
AsylADMM improves gossip-based learning for non-smooth objectives.
problem Efficient and robust decentralized learning on edge devices.
method Asynchronous gossip algorithm for non-smooth optimization.
result AsylADMM converges faster on non-smooth problems.
Many engineers wish to deploy modern neural networks in memory-limited settings; but the development of flexible methods for reducing memory use is in its infancy, and there is little knowledge of the resulting cost-benefit. We propose structural model distillation for memory reduction using a strategy that produces a …
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…
New methods improve neural network extrapolation to long sequences.
problem Neural networks struggle with extrapolation to very long or adversarial sequences.
method Activation binning and localized differentiable memory architecture.
result No extrapolation errors detected within memory constraints.
OEUVRE estimates online loss with constant time and memory, outperforming other methods.
problem Accurately estimating expected loss in online learning.
method Recursive evaluation of each sample on current and previous models, using algorithmic stability for updates.
result Consistency, convergence rates, and concentration bounds proved for OEUVRE.
Current deep neural networks can achieve remarkable performance on a single task. However, when the deep neural network is continually trained on a sequence of tasks, it seems to gradually forget the previous learned knowledge. This phenomenon is referred to as \textit{catastrophic forgetting} and motivates the field c…
SliceOut speeds up deep learning training without sacrificing accuracy.
problem Frequent model re-training and large model training workloads in deep learning.
method SliceOut uses dropout-inspired scheme to drop contiguous sets of units at random, leveraging GPU memory layout.
result 10-40% speedups and memory reduction with minimal accuracy loss.
Develops a new model to predict training dynamics of large language models.
problem Lack of mechanistic understanding of training dynamics in large language models.
method A first-principles reduced-order model of training dynamics, predicting group-size invariance and stability thresholds.
result Closed-form model predicts training dynamics with high accuracy and provides new diagnostics.
Memory is increasingly often the bottleneck when training neural network models. Despite this, techniques to lower the overall memory requirements of training have been less widely studied compared to the extensive literature on reducing the memory requirements of inference. In this paper we study a fundamental questio…
In this paper, we study the generalization properties of online learning based stochastic methods for supervised learning problems where the loss function is dependent on more than one training sample (e.g., metric learning, ranking). We present a generic decoupling technique that enables us to provide Rademacher compl…
The study examines label smoothing to improve confidence calibration in fine-tuned LLMs.
problem Improving confidence calibration in fine-tuned large language models (LLMs) after instruction tuning.
method Examine various open-sourced LLMs, label smoothing, and custom kernel design.
result Label smoothing is effective in maintaining confidence calibration but faces challenges in large vocabulary LLMs.
SGQuant reduces GNN memory usage without significant accuracy loss.
problem High memory consumption in GNNs limits their applicability on memory-constrained devices.
method Proposes a specialized GNN quantization scheme (SGQuant) with a quantization algorithm, fine-tuning scheme, and multi-granularity strategy.
result SGQuant reduces GNN memory footprint from 4.25x to 31.9x with minimal accuracy loss.
New algorithm controls systems with unknown, changing losses.
problem Control systems with adversarial perturbations and unknown loss function.
method Efficient sublinear regret algorithm for bandit convex optimization with memory.
result Achieves efficient control with sublinear regret in the presence of unknown, changing losses.
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.
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.
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 α α improves predictive performance over classical models. We study the power of different types of adaptive (nonoblivious) adversaries in the setting of prediction with expert advice, under both full-information and bandit feedback. We measure the player's performance using a new notion of regret, also known as policy regret, which better captures the adversary's adaptiveness…
A model for deciding which facts to remember in a lifelong learning scenario.
problem Deciding which facts to retain in limited memory from an endless stream of information.
method Mathematical model based on online learning framework, using multiplicative weights update algorithm with modifications.
result Design of an alternative scheme with close to optimal regret guarantees for memory-constrained lifelong learning.
ActNN reduces neural network training memory by 2-bit quantization.
problem Limited memory in training neural networks.
method Randomly quantizes activations to 2 bits, proving convergence and proposing mixed-precision strategies.
result Reduces activation memory footprint by 12x with negligible accuracy loss.
A novel approach stores encoded images as centroids and covariance matrices to improve classification accuracy with less memory.
problem Catastrophic forgetting and memory limitations in continual learning.
method Trains autoencoders with Neural Style Transfer to encode images, replay encoded episodes to avoid forgetting, and use centroids and covariance matrices for pseudo-images when memory is full.
result Increases classification accuracy by 13-17% over state-of-the-art methods on benchmark datasets, while requiring 78% less storage space.
Using more than 6.7 billions of trades, we explore how the tick-by-tick dynamics of limit order books depends on the aggregate actions of large investment funds on a much larger (quarterly) timescale. In particular, we find that the well-established long memory of market order signs is markedly weaker when large invest…
New score matching method estimates local intrinsic dimension efficiently.
problem Quantifying the local intrinsic dimension of complex data.
method Denoising score matching loss and equivalent implicit score matching loss.
result Denoising score matching loss is a highly competitive and scalable LID estimator.
A method models nonlinear dynamics from data using barycentric coordinates and memory.
problem Modeling complex dynamical systems from data.
method SPA for data projection, barycentric coordinates, delay-embedding theorem for memory.
result Stable models of chaotic dynamics and attractors are reproduced.