The paper presents a multi-power law for predicting loss curves across different learning rate schedules.
problem Understanding and optimizing the relationship between model performance and hyperparameters, especially learning rates.
method Proposes a multi-power law that combines power laws based on the sum of learning rates and additional laws for loss reduction due to decay.
result The multi-power law accurately predicts loss curves for unseen learning rate schedules and finds a schedule that outperforms cosine learning rate.
A new method prevents forgetting during knowledge transfer.
problem Catastrophic forgetting in transfer learning.
method Transfer without Forgetting (TwF) using a fixed pretrained network.
result TwF outperforms other CL methods by 4.81% in Class-Incremental accuracy.
New framework selects high-quality pretraining data without training LLMs.
problem Slow progress in understanding pretraining data due to costly experiments.
method Statistical framework based on perplexity-benchmark correlations.
result Approach outperforms existing methods on multiple benchmarks.
Study shows how pretraining robustness transfers to downstream tasks.
problem Understanding how robustness is transferred from pretraining to downstream tasks.
method Theoretical analysis and practical validation of robustness constraints.
result Robustness of a linear predictor on downstream tasks can be constrained by the robustness of its underlying representation.
Publicly pretraining models on Web data may undermine differential privacy.
problem The use of large Web-scraped datasets in differential privacy models.
method Critical review of leveraging pretrained models on public datasets for differential privacy.
result Publicizing pretrained models as 'private' could harm trust and generalize poorly.
This study examines how the size and alignment of pretraining data affect the performance of large language models on downstream tasks.
problem Understanding how the size and alignment of pretraining data impact the performance of large language models on downstream tasks.
method Investigated the scaling behavior of large language models in a transfer learning setting, focusing on machine translation tasks.
result The size of the finetuning dataset and the distribution alignment between pretraining and downstream data significantly influence the scaling behavior of downstream performance.
New scaling laws optimize model size, training, and inference for better performance.
problem Trade-off between model size and inference cost in modern LLMs.
method Train-to-Test (T2) scaling laws that jointly optimize model size, training tokens, and inference samples. result Optimal pretraining decisions shift into overtraining regime, leading to stronger performance.
AFTER technique improves NLP models by preventing overfitting to task-specific domains.
problem Standard fine-tuning degrades pretraining domain representations.
method Complements task-specific loss with adversarial objective.
result AFTER leads to improved performance on various NLP tasks.
A new approach to protein language models combines latent space prediction with masked language modeling.
problem Improving protein language models by predicting amino acid identities at masked positions.
method A variant of masked language modeling that predicts latent targets only at masked positions, retaining the MLM cross-entropy.
result The new approach outperforms pure masked language modeling on 11 out of 16 downstream tasks.
Framework adds invariance to pretrained networks without fine-tuning.
problem Adding invariance to pretrained networks without altering original behavior.
method Post-training augmentation invariance framework with Markov-Wasserstein minimization and Wasserstein correlation maximization losses.
result Adapter networks improve classification accuracy on rotated and noisy images.
We introduce a pretraining technique called Selfie, which stands for SELFie supervised Image Embedding. Selfie generalizes the concept of masked language modeling of BERT (Devlin et al., 2019) to continuous data, such as images, by making use of the Contrastive Predictive Coding loss (Oord et al., 2018). Given masked-o…
Study shows how specialized attention circuits emerge during transformer training.
problem Understanding the mechanisms of transformer training dynamics at large scales.
method Controlled sparse modular addition task; monitoring token evolution via visual sandbox.
result Specialized attention circuits (clustering heads) naturally emerge during training.
Flaky performance found in GNN SSL on RDBs, leading to worse linear evaluation.
problem Downstream task performances of GNN SSL on RDBs are poor.
method Proposed InfoNode to maximize mutual information between initial and final node representations.
result InfoNode improves GNN SSL performance on RDBs, supporting conjecture of conflict between SSL and GNN message passing.
A method for self-supervised learning of multivariate time series data across varying channels.
problem Labeling multivariate biomedical time series data is laborious and expensive.
method Proposes a multi-view self-supervised learning approach using a message passing neural network to extract a single representation across channels.
result Our method, combined with the TS2Vec loss, outperforms all other methods in most settings.
A pretrained LLM and its finetuned version can detect OOD data effectively.
problem Detecting out-of-distribution data in language models.
method Using the likelihood ratio between a pretrained and finetuned LLM.
result The likelihood ratio is an effective criterion for OOD detection.
In many applications, one works with neural network models trained by someone else. For such pretrained models, one may not have access to training data or test data. Moreover, one may not know details about the model, e.g., the specifics of the training data, the loss function, the hyperparameter values, etc. Given on…
Transformers learn low-dimensional target functions efficiently in-context.
problem Efficiently learning nonlinear target functions in-context using transformers.
method Nonlinear MLP layer in transformers optimized by gradient descent, focusing on single-index target functions.
result Transformers can learn target functions with low-dimensional structures efficiently in-context.
DreamFusion uses text-to-image diffusion models to create 3D images efficiently.
problem Lack of large-scale 3D datasets and efficient architectures for 3D synthesis.
method Adapting a 2D diffusion model to 3D synthesis using a loss based on probability density distillation.
result A 3D model can be optimized from a 2D diffusion model, allowing for text-to-3D synthesis.
WSD schedule improves model training efficiency by adapting learning rates dynamically.
problem Fixed compute budgets limit training efficiency of language models.
method Introduces a WSD schedule that uses a constant learning rate followed by a rapid decay phase.
result WSD schedule generates a non-traditional loss curve with stable and decay phases.
SFAVEL distills features from language models for fact verification without annotations.
problem Fact verification with semantically meaningful and compact features.
method Self-supervised pretraining using contrastive loss with language models.
result Achieved state-of-the-art results on FB15k-237 and FEVER.
Big models pretrain and fine-tune for semi-supervised learning on ImageNet.
problem Learning from few labeled examples with a large amount of unlabeled data.
method Unsupervised pretraining of a big ResNet model followed by supervised fine-tuning and distillation.
result 73.9% ImageNet top-1 accuracy with just 1% of labels (≤13 labeled images per class). FAST improves fast and stable task adaptation in DNNs.
problem Catastrophic forgetting in fine-tuned pretrained models.
method Introducing FAST, an easy-to-implement fine-tuning algorithm.
result FAST learns target tasks faster and retains source knowledge longer.
Advances few-shot classification by treating it as supervised learning and proposing new training techniques.
problem Formulating the ability of humans to learn from limited data in machine learning.
method Formulated few-shot classification as a supervised learning problem and introduced multi-episode and cross-way training techniques.
result Proposed training strategies accelerate the training process without accuracy loss.
Task-agnostic data augmentation shows little benefit for pretrained transformers.
problem Evaluating the effectiveness of task-agnostic data augmentation on pretrained transformers.
method Conducted a systematic examination of two data augmentation techniques (Easy Data Augmentation and Back-Translation) across 5 tasks, 6 datasets, and 3 pretrained transformer models.
result Data augmentation techniques previously effective for non-pretrained models fail to consistently improve performance for pretrained transformers, even with limited training data.
Two CSSL-based methods improve graph classification with limited labeled data.
problem Limited labeled data for graph classification leads to overfitting.
method Contrastive self-supervised learning (CSSL) for graph encoders pretraining and regularization.
result CSSL methods reduce overfitting and improve graph classification accuracy.
Pretraining models improves text classification accuracy, but diminishing returns are observed with large datasets.
problem Improving text classification accuracy with pretrained models.
method Examined the benefits of pretrained models on text classification tasks with varying amounts of training data.
result As the number of training examples grows into the millions, the accuracy gap between pretrained BERT-based models and vanilla LSTM narrows to within 1%.
This paper improves model fusion by training-time neuron alignment, reducing barriers in multi-model fusion.
problem Diverse neuron permutations across different settings hinder model fusion performances.
method Training-time neuron alignment using fixed neuron anchors to reduce training-time permutations.
result Training-time neuron alignment improves fusion of pretrained models and federated learning performances.
APLC-XLNet improves XMTC by clustering labels and reducing computational time.
problem Efficiently tagging texts with many labels from a large set.
method Fine-tunes XLNet with APLC to approximate cross entropy loss.
result Achieved state-of-the-art results on XMTC benchmarks.
Paper studies how few pretraining tasks are needed for a linear model to solve new tasks.
problem How many pretraining tasks are needed for a linear model to solve new tasks?
method Pretrained a linear attention model for linear regression with a Gaussian prior.
result Effective pretraining requires a small number of independent tasks, and the model closely matches Bayes optimal.
SiD distills pretrained diffusion models into a fast one-step generator.
problem Efficiently distilling pretrained diffusion models into a fast generator.
method Reformulates forward diffusion processes as semi-implicit distributions and uses three score-related identities to create a loss mechanism.
result Achieves high FID performance and significantly reduces generation time.
Fine-grained pretraining improves neural network's ability to learn rare features.
problem Improving generalization in deep learning models.
method Introducing a hierarchical multi-view structure to confine input data distribution.
result Fine-grained pretraining leads to better accuracy on hard downstream test samples.
Method constructs finance LLMs without instruction data using pretraining and model merging.
problem Developing domain-specific LLMs for finance is resource-intensive.
method Continual pretraining on financial data + model merging of instruction-tuned and domain-specific pretrained vectors.
result Successfully constructs instruction-tuned LLMs for finance without additional instruction data.
FAIRIF improves fairness in deep learning models without changing the model architecture.
problem Improving fairness in deep learning models trained on sensitive data.
method Two-stage training algorithm that minimizes loss over a reweighted data set, balancing model performance across demographic groups.
result FAIRIF reduces disparity among different groups in classification settings, improving fairness-utility trade-offs.
Mask-reconstruction pretraining helps in downstream tasks by capturing more semantic features.
problem How mask-reconstruction pretraining helps in downstream tasks and why it surpasses supervised learning.
method Theoretical analysis and experimental validation of mask-reconstruction pretraining (MRP) on auto-encoders.
result MRP provably captures more semantic features than supervised learning, leading to better performance in downstream tasks.
The paper proposes a method to estimate heterogeneous treatment effects using pretraining strategies.
problem Estimating conditional average treatment effects (CATE) in the presence of many covariates.
method The approach leverages prognostic factors that also predict treatment effect heterogeneity, using the R-learner framework.
result The proposed method improves estimation accuracy and power for detecting treatment effect heterogeneity.
Improved motion prediction for self-driving cars using trajectory sets and auxiliary losses.
problem Accurately predicting future vehicle motion for self-driving cars.
method Classification over trajectory sets with an auxiliary loss for off-road predictions and spatial-temporal relationships.
result Significant improvement in motion prediction performance on small datasets using map information.
Paper introduces a multi-stage influence function to track model predictions.
problem Improving natural language processing and computer vision performance.
method Develops a multi-stage influence function to track predictions from finetuned models back to pretraining data.
result Identifies pretraining examples contributing most to finetuning task predictions.
Pretraining method enhances dialogue representation learning across various tasks.
problem Scarce labeled data for specific dialogue tasks.
method Multi-task unsupervised pretraining with natural training objectives.
result Significant improvement in downstream tasks without encoder discrimination.
Synthetic continued pretraining enhances model performance with synthetic data.
problem Data inefficiency in pretrained models when adapting to domain-specific documents.
method Synthetic data augmentation using EntiGraph to create a large synthetic corpus.
result Language models can answer questions and follow instructions without access to domain-specific documents.
The study introduces anytime learning schedules for large language models without fixed horizons.
problem Training large language models without knowing the total training horizon.
method Theoretical analysis and weight averaging to create anytime learning schedules.
result Theoretical and empirical evidence shows that weight averaging with simple step sizes can achieve comparable final loss to well-tuned cosine schedules.
Deep neural networks trained on a wide range of datasets demonstrate impressive transferability. Deep features appear general in that they are applicable to many datasets and tasks. Such property is in prevalent use in real-world applications. A neural network pretrained on large datasets, such as ImageNet, can signifi…
WeatherFormer learns robust weather features from small datasets.
problem Modeling complex weather dynamics from limited data.
method Pretrained transformer encoder on large satellite dataset, with spatiotemporal encoding.
result State-of-the-art performance in county-level soybean yield prediction and influenza forecasting.
Given two networks with the same training loss on a dataset, when would they have drastically different test losses and errors? Better understanding of this question of generalization may improve practical applications of deep networks. In this paper we show that with cross-entropy loss it is surprisingly simple to ind…
Subset pretraining speeds up neural network training.
problem Efficiently training neural networks with partial gradients.
method Use small subsets to approximate the loss minimum of the full training set.
result Subset minima can approximate the full training set's loss minimum efficiently.
Theoretical analysis of data quality and synergies in LLMs.
problem Understanding why different training methods require different amounts of data.
method Theoretical analysis of transformers trained on a weight prediction task for linear regression.
result SFT excels on smaller datasets challenging for the pretrained model, while RL benefits from large, not overly difficult data.
Study compares metric learning loss functions for speaker verification.
problem Comparing metric learning loss functions for end-to-end speaker verification.
method Cross entropy loss, cosine loss, angular margin loss, center loss, contrastive loss, triplet loss.
result Additive angular margin loss outperforms other loss functions.
Denoised smoothing defends pretrained classifiers against adversarial attacks.
problem Adversarial attacks on pretrained classifiers.
method Prepending a denoiser to any off-the-shelf classifier using randomized smoothing.
result Guaranteed ℓp-robustness to adversarial examples without modifying the pretrained classifier. In this paper, we explore the ways to improve POS-tagging using various types of auxiliary losses and different word representations. As a baseline, we utilized a BiLSTM tagger, which is able to achieve state-of-the-art results on the sequence labelling tasks. We developed a new method for character-level word represen…