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.
Study shows scaling up models doesn't always improve downstream tasks.
problem Understanding why scaling up models doesn't always improve downstream performance.
method Systematic study of 4800 experiments on various models, analyzing performance on 20 downstream tasks.
result Performance on downstream tasks saturates as model size increases, revealing a nonlinear relationship.
End-to-end approach for weak supervision improves downstream model performance.
problem Data-labeling bottleneck in machine learning applications.
method Directly learning the downstream model by maximizing its agreement with probabilistic labels generated from weak supervision sources.
result Improved performance over prior work in terms of downstream model performance and robustness.
Theoretical analysis shows pretext-based self-supervised learning can be boosted by downstream data under certain conditions.
problem Theoretical analysis of pretext-based self-supervised learning and downstream data refinement.
method Theoretical analysis and experiments on synthetic and real-world datasets.
result Theoretical lower bounds and experiments show that downstream data refinement can boost or hurt performance depending on conditions.
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.
This work optimizes alignment and uniformity of features on a hypersphere for better downstream performance.
problem Improving the performance of contrastive representation learning.
method Identifying and optimizing alignment and uniformity of features on a hypersphere.
result Directly optimizing alignment and uniformity leads to comparable or better performance than contrastive learning.
Contrastive learning performance doesn't degrade with more negative samples.
problem Theoretical and empirical evidence of negative samples hurting performance in contrastive learning.
method Simple theoretical setting and empirical support on CIFAR-10 and CIFAR-100 datasets.
result Contrastive learning performance does not degrade with the number of negative samples.
This paper characterizes VAE training pathologies and their effects on tasks.
problem Characterizing VAE training pathologies and their impact on downstream tasks.
method Concretely characterizing conditions for VAE training pathologies and their connection to specific downstream tasks.
result Connects VAE training pathologies to specific downstream tasks like learning compressed and disentangled representations, adversarial robustness, and semi-supervised learning.
This work studies the impact of intra-/inter-class diversity on pre-training datasets and finds a balance for optimal performance.
problem The impact of intra-/inter-class diversity on supervised pre-training datasets and their effect on downstream tasks.
method Empirical study and theoretical analysis of the relationship between diversity types and downstream performance.
result The optimal class-to-sample ratio is invariant to the size of the pre-training dataset and can be predicted.
New framework analyzes why more negative samples improve self-supervised learning performance.
problem Inconsistency between theoretical degradation and empirical improvement of downstream supervised tasks with more negative samples.
method Coupon collector's problem framework to analyze self-supervised representation learning with more negative samples.
result Bound can implicitly incorporate supervised loss in self-supervised loss by increasing negative samples.
Optimal feature transfer identified through bias-variance analysis.
problem Optimizing feature transfer in transfer learning.
method Simple linear model with fine-grained bias-variance decomposition.
result Optimal pretrained feature transform is naturally sparse.
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.
Paper designs a lossy compression method for lossless prediction.
problem Ensuring high performance on predictive tasks with minimal data.
method Characterizes bit-rate requirements for invariant transformations, designs unsupervised objectives for neural compressors.
result Achieves substantial rate savings on ImageNet compared to JPEG without compromising classification performance.
Enhances performance on downstream tasks using multi-domain data.
problem Improving performance on tasks with limited downstream data.
method Deep transfer learning framework leveraging shared and domain-specific features.
result Significantly improves convergence rate for learning Lipschitz functions.
Identifies optimal base features for zero-shot adaptation in reinforcement learning.
problem Unclear what constitutes a good set of base features for a wide range of downstream tasks.
method Identifies optimal base features based on downstream performance, without assuming downstream tasks are linear.
result Optimal base features are the same across three task families, differing from Laplacian eigenfunctions.
New insights into contrastive learning reveal how projectors affect downstream performance.
problem Understanding how projectors in contrastive learning impact downstream linear classification accuracy.
method Identified and modeled two effects: expansion and shrinkage induced by contrastive loss.
result Linear projectors operating in the shrinkage regime hinder downstream classification accuracy.
This paper evaluates fairness in deep metric learning and proposes a method to reduce subgroup performance gaps.
problem The negative impact of deep metric learning representations on minority subgroup performance in downstream tasks.
method Definition of fairness in DML through inter-class, intra-class, and uniformity properties; finDML benchmark; Partial Attribute De-correlation (PARADE) method.
result Bias in DML representations propagates to downstream tasks, even with balanced training data.
DKPS provides guarantees for synthetic data from Transformer models, improving downstream tasks.
problem Lack of labeled data for building performant AI models.
method Data Kernel Perspective Space (DKPS) for mathematical analysis of synthetic data quality.
result Concrete statistical guarantees for the quality of transformer model outputs.
Analysis of pretrained models' effectiveness in downstream tasks.
problem Understanding why pretrained models perform well in NLP tasks.
method Analyzed head and prompt tuning approaches using latent variable models.
result Prompt tuning provides stronger guarantees than head tuning.
PEARL combines multiple representation learning methods to enhance model performance.
problem Different representation learning methods extract distinct data aspects, potentially missing important insights.
method Combines multiple representation learning approaches using surrogate loss functions for efficient weight estimation.
result Asymptotically achieves optimal performance in downstream tasks, assigning nonzero weights to correctly specified models.
New method optimizes treatment policies to avoid winner's curse.
problem Winner's curse in treatment policy optimization.
method Inference-aware policy optimization.
result Optimizes for both estimated performance and downstream evaluation.
A comprehensive benchmark of 15 scRNA-seq imputation methods across various datasets and analyses.
problem Imputation of single-cell RNA sequencing data to recover latent transcriptional signals.
method Evaluation of 15 imputation methods across 30 datasets and 6 downstream analyses.
result Traditional methods generally outperform DL-based methods in scRNA-seq data analysis.
New framework explains how larger pre-trained models reduce downstream learning sample complexity.
problem Understanding why larger pre-trained models reduce sample complexity in downstream tasks.
method Introducing a novel framework called Caulking inspired by PEFT methods.
result Improved pre-trained models provably decrease downstream task sample complexity.
Study shows uncertainty of deep learning models can be measured from their embeddings.
problem Uncertainty in contrastive learning models for critical applications.
method Estimating the distribution of training data in embedding space and accounting for local consistency.
result Uncertainty of an embedding vector correlates strongly with downstream accuracy.
Object-centric learning improves generalization and robustness in multi-object scenes.
problem Improving generalization and robustness in neural networks for scenes with multiple objects.
method Training state-of-the-art unsupervised models on multi-object datasets and evaluating segmentation metrics and downstream tasks.
result Object-centric representations are useful for downstream tasks and generally robust to most distribution shifts affecting objects, but less so for less structured shifts.
Compressing word embeddings is important for deploying NLP models in memory-constrained settings. However, understanding what makes compressed embeddings perform well on downstream tasks is challenging---existing measures of compression quality often fail to distinguish between embeddings that perform well and those th…
Optimal Word2Vec hyper-parameters improve NLP tasks.
problem Finding the best Word2Vec hyper-parameters for NLP tasks.
method Empirical evaluation of various hyper-parameter combinations on NLP tasks.
result The best hyper-parameters vary by task, and high analogy scores don't always correlate with performance.
Pre-training improves model coverage, crucial for downstream performance.
problem Understanding why pre-training enhances model performance.
method Coverage principle, focusing on next-token prediction and model quality.
result Coverage generalizes faster than cross-entropy, improving downstream performance.
Contrastive learning outperforms autoencoders and GANs in feature recovery and downstream tasks.
problem Theoretical understanding of contrastive learning's superiority in feature learning.
method Theoretical analysis of contrastive learning in linear representation settings.
result Contrastive learning outperforms autoencoders and GANs for feature recovery and in-domain downstream tasks.
This work investigates how neural collapse improves transfer learning for large-scale models.
problem Improving transfer learning for large-scale models with limited labeled data.
method Investigates neural collapse and develops a fine-tuning method using skip-connections.
result Feature collapse on downstream data correlates with higher transfer accuracy.
Proposes a new metric to quantify the difference between neural network representations based on downstream task performance.
problem The lack of a consistent metric to measure the difference between neural network representations.
method Introduced the Transferred Discrepancy (TD) metric, which evaluates the difference between representations based on their performance on downstream tasks.
result TD provides fine-grained information for various downstream tasks and can evaluate the effectiveness of different training strategies.
New theory explains contrastive learning via overlapping augmented views.
problem Lack of theoretical understanding of contrastive learning.
method Augmentation overlap perspective to improve downstream performance.
result Asymptotically closed bounds for downstream performance under weaker assumptions.
Improved tabular models learn better from real-world data.
problem Tabular models perform poorly on real-world datasets when trained only on synthetic data.
method Continued pre-training on a curated set of real-world datasets.
result Real-TabPFN achieves superior predictive accuracy on 29 datasets.
Improved pre-trained embeddings through effective entropy maximization.
problem Developing high-quality pre-trained embeddings for future tasks.
method E2MC criterion defined in terms of low-dimensional constraints.
result Significant improvement in downstream performance.
SidAE combines autoencoders and Siamese networks for self-supervised feature extraction.
problem Efficiently extract meaningful features from unlabeled data.
method Proposes SidAE, a combination of autoencoders and Siamese networks for self-supervised image classification.
result SidAE outperforms self-supervised baselines across multiple datasets and scenarios, especially with limited labeled data.
The paper finds a surprising positive correlation between upstreamness and downstreamness in global value chains.
problem The puzzling positive correlation between upstreamness and downstreamness in industries and countries.
method Analysis of a simple model of random Input/Output tables and experiments on empirical data.
result Upstreamness and downstreamness of the same industrial sector/country are positively correlated with a slope close to +1.
This paper improves cross-domain learning using random forests for manifold alignment.
problem Improving cross-domain learning and feature integration.
method Semi-supervised manifold alignment using random forest proximities.
result Random forest proximities enhance downstream classification accuracy.
A new framework designs experiments for better decision-making.
problem Suboptimal experimental designs for downstream decision-making.
method Amortized decision-aware Bayesian Experimental Design (BED) with Transformer Neural Decision Process (TNDP).
result TNDP effectively designs experiments and facilitates accurate decision-making.
New bounds for contrastive learning handle domain shifts and generalization.
problem Domain shifts and generalization challenges in downstream tasks.
method Novel generalization bounds accounting for both domain shift and generalization.
result Performance of contrastively learned representations depends on statistical discrepancy between pretraining and downstream distributions.
Study improves healthcare time series imputation by considering structured missingness.
problem Structured missingness in clinical data impacts time series imputation models.
method Analysis of different masking strategies on imputation methods using PhysioNet Challenge 2012 dataset.
result Masking choices significantly affect imputation accuracy and clinical prediction.
A new robust scaling approach improves downstream metabolomics analysis.
problem Challenges in choosing scaling techniques for metabolomics data.
method Introduces a weighted scaling approach robust to outliers.
result The proposed method outperforms traditional scaling techniques in both outlier-free and outlier-present datasets.
The paper explores how word embeddings affect the stability of downstream NLP models.
problem Small changes in training data can cause significant changes in model predictions.
method Empirical and theoretical analysis of embedding instability, including the introduction of eigenspace instability measure.
result Increasing embedding memory can reduce the disagreement in predictions by 5% to 37%.
PRAGMA models financial event sequences for various banking tasks.
problem Handling diverse financial data for multiple applications.
method Pre-training a Transformer model on a large banking event corpus with a self-supervised objective.
result PRAGMA achieves superior performance across multiple financial domains from raw event sequences.
TaskMet learns a metric to improve model performance on unseen tasks.
problem Deep models trained on one task may struggle on another task due to conflicting objectives.
method TaskMet learns a metric in the prediction space to balance task and prediction losses.
result TaskMet achieves better performance on downstream tasks without altering the prediction model.
Adversarially robust models transfer better than standard models in image classification.
problem Improving transfer learning performance in image classification.
method Focused on adversarially robust ImageNet classifiers, compared to standard models.
result Adversarially robust models yield improved accuracy on downstream classification tasks.
This paper examines how labeling error affects contrastive learning and proposes data dimensionality reduction methods to mitigate its impact.
problem The impact of labeling error on the performance of contrastive learning.
method Data dimensionality reduction methods (e.g., SVD) are applied to reduce false positive samples and improve downstream classification accuracy.
result Data dimensionality reduction methods can mitigate the negative impacts of labeling error on downstream classification performance.
New measure quantifies contrastive self-supervised learning's generalization ability.
problem Limited theoretical understanding of contrastive self-supervised learning's generalization.
method Defined (σ,δ)-measure to mathematically quantify data augmentation and provide an upper bound for downstream classification error. result Generalization ability is related to alignment of positive samples, divergence of class centers, and concentration of augmented data.
Geospatial ML models need special evaluation methods due to their unique challenges.
problem Evaluating geospatial machine learning models is challenging due to their specific characteristics.
method Delineated unique challenges and proposed concrete takeaways for improving geospatial model evaluations.
result Concrete takeaways for improving evaluations of geospatial model performance.