The paper shows how curated synthetic data can optimize human preferences in generative models.
problem Contamination of web-scale datasets by synthetic data affects future model training.
method Theoretical study of iterated retraining of generative models with curated synthetic data.
result Data curation can be seen as an implicit preference optimization mechanism, maximizing expected reward.
Model compression has gained a lot of attention due to its ability to reduce hardware resource requirements significantly while maintaining accuracy of DNNs. Model compression is especially useful for memory-intensive recurrent neural networks because smaller memory footprint is crucial not only for reducing storage re…
This paper optimizes retraining models using their own predictions and noisy labels.
problem Improving model performance through optimal retraining of noisy labels.
method Developed a principled framework based on approximate message passing (AMP) to analyze iterative retraining procedures.
result Derivation of the Bayes optimal aggregator function to minimize prediction error.
Active learning (AL) concerns itself with learning a model from as few labelled data as possible through actively and iteratively querying an oracle with selected unlabelled samples. In this paper, we focus on analyzing a popular type of AL in which the utility of a sample is measured by a specified goal achieved by th…
Classifier learns to ignore unreliable feedback from end users.
problem Improving classifier performance by filtering unreliable feedback.
method Modeling end users as autonomous agents, periodically retraining classifier with filtered feedback.
result Classifier can identify and filter out unreliable feedback, improving performance.
Paper tackles model collapse in synthetic data retraining.
problem Iterative retraining of generative models on synthetic data can lead to performance deterioration.
method Integrates an external synthetic data verifier to prevent model collapse.
result Synthetic retraining with a verifier can improve model performance initially but may lead to convergence to the verifier's knowledge center.
Dynamic pruning during training reduces deep network complexity without significant accuracy loss.
problem High memory and computational requirements of deep networks during training and inference.
method Dynamic pruning of convolutional filters during training, using L1 normalization for optimization.
result L1 normalization-based pruning yields up to 50% reduction in filters with minimal accuracy loss.
Quantum computing speeds up neural network training and retraining.
problem Inefficient classical training and retraining of neural networks.
method Adiabatic quantum computing to optimize Kolmogorov-Arnold Networks using Bezier curves.
result Quantum optimization achieves 100x faster retraining compared to classical methods.
New method protects neural networks from adversarial attacks without generating adversarial examples.
problem Vulnerability of neural networks to adversarial examples.
method Entropic retraining, inspired by information theory.
result Significant increase in NNs' security and robustness.
Method generates resource-optimized ML models for different platforms.
problem Deployment of deep learning models on edge platforms requires model reduction and retraining.
method Conditional Neural Architecture Search using Generative Adversarial Networks (GAN)
result Successfully generates resource-optimized ML models for different platforms.
Paper identifies and solves a 'scrambled translation' issue in UNMT models.
problem Scrambled translation in UNMT models using word shuffle noise.
method Retraining UNMT models without noise after a pre-defined number of iterations.
result Retraining strategy improves BLEU scores by 10-20% on various language pairs.
Proposes a method to retrain neural networks incrementally for continuous data flow.
problem Continuous data flow and the challenges of catastrophic forgetting and efficient retraining.
method Incremental retraining using multi-armed bandits to select important samples and weights, and a new regularization term for synapse and neuron importance.
result Mitigates catastrophic forgetting and boosts model performance.
Active learning methods for neural networks are usually based on greedy criteria which ultimately give a single new design point for the evaluation. Such an approach requires either some heuristics to sample a batch of design points at one active learning iteration, or retraining the neural network after adding each da…
Kernel methods can learn hierarchical polynomials efficiently.
problem Learning hierarchical structure from data.
method Iteratively reweighting kernel machines using derivatives.
result Efficient learning of hierarchical polynomials.
Retraining with predicted labels improves model accuracy in noisy settings.
problem Improving model accuracy with noisy or corrupted labels.
method Retraining with predicted hard labels in a linearly separable binary classification setting.
result Retraining with predicted labels can increase model accuracy, as proven theoretically.
DeltaGrad rapidly retrain models with minimal data changes.
problem Rapid retraining of machine learning models with minimal data changes.
method DeltaGrad algorithm based on cached training information.
result DeltaGrad compares favorably to state-of-the-art methods.
Less frequent retraining improves forecast accuracy in retail demand forecasting.
problem Balancing forecast accuracy and computational efficiency in global models.
method Analysis of ten machine learning and deep learning models across two large retail datasets with various retraining scenarios.
result Less frequent retraining strategies maintain forecast accuracy while reducing computational costs.
K-priors enable quick adaptation with minimal retraining.
problem Machine learning models struggle to adapt to changes efficiently.
method Combines weight and function-space priors to reconstruct past gradients.
result Adaptation with K-priors achieves similar performance to full retraining with less data.
Two retraining techniques outperform fine-tuning in neural network pruning.
problem Improving accuracy and compression in neural network pruning.
method Weight rewinding and learning rate rewinding compared to fine-tuning.
result Rewinding techniques outperform fine-tuning in accuracy and compression.
Traditional classification algorithms assume that training and test data come from similar distributions. This assumption is violated in adversarial settings, where malicious actors modify instances to evade detection. A number of custom methods have been developed for both adversarial evasion attacks and robust learni…
The paper studies stability of generative models trained on mixed data.
problem Training generative models on mixed datasets (real and synthetic data).
method Developed a framework to rigorously study stability under specific conditions.
result Proved the stability of iterative training under certain conditions.
New method prevents RLHF alignment collapse by accounting for policy's influence on reward model updates.
problem Iterative RLHF leads to alignment collapse where policies exploit RM's blind spots.
method Foresighted policy optimization (FPO) restores missing steering term via regularization.
result FPO prevents alignment collapse on LLM alignment pipelines using Llama-3.2-1B.
DNN pruning reduces memory footprint and computational work of DNN-based solutions to improve performance and energy-efficiency. An effective pruning scheme should be able to systematically remove connections and/or neurons that are unnecessary or redundant, reducing the DNN size without any loss in accuracy. In this p…
Retraining stabilizes model influence on data.
problem Performativity in predictive models leads to feedback loops.
method Developed the stable signal principle to address retraining dynamics.
result Repeated risk minimization converges geometrically to stable signal direction.
This research shows loss weighting remains effective in last layer retraining despite model overparameterization.
problem Overcoming biases in machine learning models at scale.
method Theoretical and practical exploration of last layer retraining in an overparameterized setting.
result Loss weighting is still effective in last layer retraining, but weights must account for model overparameterization.
New method detects data distribution changes and retraining is advised.
problem Detecting when data distribution changes for retraining prediction algorithms.
method Exchangeable martingales and conformal prediction.
result Guaranteed validity of the method, with efficiency explored.
Layer fusion reduces deep neural network layers with minimal loss in accuracy.
problem Model compression to reduce neural network size and computation.
method Fusion of similar layers to reduce model size with minimal performance loss.
result Deep networks can be compressed up to 3.33x with minimal accuracy loss.
Data augmentation methods improve worst-case model performance.
problem Ensuring fair predictions across subpopulations in large models.
method Linear last layer retraining with data augmentation techniques.
result Optimal worst-group accuracy achieved for Gaussian latent representation distribution.
The most common method for DNN pruning is hard thresholding of network weights, followed by retraining to recover any lost accuracy. Recently developed smart pruning algorithms use the DNN response over the training set for a variety of cost functions to determine redundant network weights, leading to less accuracy deg…
Last layer retraining improves robustness to spurious correlations without high computational costs.
problem Neural networks can rely on spurious features like backgrounds for predictions.
method Simple last layer retraining on large models.
result Last layer retraining matches or outperforms state-of-the-art approaches on spurious correlation benchmarks.
In this paper, we study a simple and generic framework to tackle the problem of learning model parameters when a fraction of the training samples are corrupted. We first make a simple observation: in a variety of such settings, the evolution of training accuracy (as a function of training epochs) is different for clean…
New framework for predicting decisions that influence their own outcomes.
problem Predictions that affect the outcomes they predict, leading to undesirable distribution shift.
method Risk minimization framework combining statistics, game theory, and causality.
result Necessary and sufficient conditions for retraining to converge to a performatively stable point of minimal loss.
New algorithm extends Greville's method for partitioned matrices efficiently and stably.
problem Efficiently compute pseudoinverse of partitioned matrices without retraining.
method Incorporates inverse Cholesky factorization to reduce computational complexity and improve stability.
result 1 iteration to compute pseudoinverse of whole matrix from first part, addressing all cases.
New method uses neural tangent kernel for efficient active learning.
problem Efficiently approximating deep learning's look-ahead selection criteria.
method Approximates retraining with neural tangent kernel for active learning.
result Approximation works asymptotically and enables sequential active learning.
Real-time pruning during training reduces network size and training time.
problem Efficiently reducing neural network size without sacrificing accuracy.
method Activation density-based pruning during training.
result Up to 200x reduction in parameters and 60x reduction in inference compute operations.
We present a weakly-supervised data augmentation approach to improve Named Entity Recognition (NER) in a challenging domain: extracting biomedical entities (e.g., proteins) from the scientific literature. First, we train a neural NER (NNER) model over a small seed of fully-labeled examples. Second, we use a reference s…
Improved method for efficient black-box optimization in latent space.
problem Efficiently optimize expensive black-box functions over complex input spaces.
method Optimize in latent space of deep generative models, retrain model periodically and weight data points.
result Significantly improved efficiency and performance on synthetic and real-world problems.
Efficiently removes specific data subsets without retraining for GDPR compliance.
problem Efficiently removing specific data subsets to comply with GDPR regulations.
method Statistical framework for machine unlearning with minimax optimality for squared loss.
result Developed Unlearning Least Squares (ULS) achieving minimax optimality for estimating model parameters.
Many active learning methods belong to the retraining-based approaches, which select one unlabeled instance, add it to the training set with its possible labels, retrain the classification model, and evaluate the criteria that we base our selection on. However, since the true label of the selected instance is unknown, …
The paper proposes a method to monitor deep learning predictions for retraining, reducing costs.
problem Reducing computational costs in deep learning by detecting when predictions are no longer valid.
method Sequential monitoring of network predictions based on projected second moments monitoring.
result The proposed method can drastically reduce computational costs in deep learning.
Study on self-consuming generative models with diverse human curation, focusing on convergence and stability.
problem Analyzing self-consuming generative models with heterogeneous human curation.
method Investigates the asymptotic behavior of retraining dynamics using nonlinear Perron--Frobenius theory and Banach contraction mapping.
result Improves convergence results and provides stability and non-stability analyses for the model.
SUMER updates deployed models with new data, improving performance.
problem Updating deployed machine learning models with new data is difficult.
method SUMER uses semi-supervised learning and noise remediation to iteratively retrain models.
result SUMER improves model performance, especially with limited initial training data.
Wide and Deep GNN learns from distributed graphs and retrain online.
problem Decentralized graph support changes over time, causing mismatch between training and testing graphs.
method Wide and Deep GNN architecture with distributed online learning.
result Convergence guarantees for online retraining of the wide part of the GNN.
Automatically detects and down-weights noisy samples in machine learning training.
problem Numerical noise in reference data hampers the accuracy of machine learning models.
method On-the-fly outlier detection using exponential moving average to identify and down-weight noisy samples.
result The method prevents overfitting and matches the performance of iterative refinement with reduced overhead.
This paper analyzes adversarial attacks methods and their effectiveness.
problem Understanding the effectiveness and theoretical properties of adversarial attacks.
method Comparative and formal analysis of loss functions of three adversarial attack methods.
result The Iterative Fast Gradient Sign attack is the slowest in creating adversarial examples.
IMP finds sparse subnetworks that match full networks, revealing geometric insights.
problem Finding sparse subnetworks that match full, overparameterized networks.
method Iterative magnitude pruning (IMP) algorithm, analyzing error landscape geometry.
result IMP masks found at end of training convey useful information for rewound networks.
Unified framework for arbitrary conditional inference using AI and Bayesian methods.
problem Limited flexibility in existing conditional inference methods.
method Bayesian generative modeling with stochastic iterative algorithm.
result Single learned model for universal conditional prediction with uncertainty quantification.
Decomposable-Net compresses neural networks without retraining for various sizes.
problem Performance degradation when changing model size after training.
method Decomposes weight matrices via SVD and adjusts ranks for different sizes.
result Maintains and improves performance across multiple model sizes.