Investigating the Interaction between Data and Algorithms

Daniel Pototzky, Azhar Sultan, Lars Schmidt-ThiemeProceedings of the World Congress on Electrical Engineering and Computer Systems and Science 2022 · DOI

Abstract

Research in computer vision is centered on algorithmic improvements, for example, by developing better models. Thereby, the data is considered fixed. This is in contrast to many real-world applications of computer vision systems in which algorithms and data co-evolve. To address this shortcoming of previous research, we study the properties of the data and their interaction with deep learning algorithms. Thereby, we investigate the size of the data, the share of mislabels, class imbalance and the presence of unlabeled data which can be leveraged using semi-supervised learning. In experiments on 100 classes from ImageNet, we show that a tiny network architecture outperforms a much more powerful one it if has access to only a little bit more data. Only if vast amounts of data are available so that adding even more images has little effect on performance, large architectures dominate smaller ones. If little data is provided, adding a few labeled images has a huge effect on accuracy. Once accuracy saturates, massive amounts of additional data are needed to achieve even small improvements. Furthermore, we find that mislabels severely reduce performance. To fix that, we propose a cost-efficient way of identifying mislabels which is especially beneficial if many images are already available. Conversely, if little data is available, labeling more images is more advantageous than cleaning existing annotations. In the case of imbalanced data, we illustrate that labeling more instances from rare classes has a much greater effect on performance than only increasing dataset size. Moreover, we show that leveraging unlabeled images by semi-supervised learning offers a consistent benefit even if the labeled subset contains significant label noise.

Introduction

Most computer vision research treats datasets as static benchmarks—fixed entities used to crown the "best" new architecture. However, in industrial and real-world applications, data and algorithms co-evolve. Engineers often face a "data-algorithm" trade-off: Should budget be allocated to a more complex model or to higher-quality data? This paper addresses the lack of systematic study regarding how dataset properties—specifically size, label noise (mislabels), and class imbalance—interact with different model architectures.

Methodology

The study compares two ends of the architectural spectrum: ResNet-50 (~25 million parameters) and SqueezeNet-1.0 (~1 million parameters). Experiments were conducted on a 100-class subset of ImageNet (approx. 130,000 images), chosen for its fine-grained categories.

The researchers explored several data dynamics:

  • Supervised Learning: Standard training using cross-entropy loss.
  • Semi-Supervised Learning (SSL): Utilizing FixMatch to leverage unlabeled data via consistency regularization.
  • Cost-Efficient Mislabel Correction: A k-fold approach where a model trained on a subset flags potential mislabels in the remaining data for manual review.

Results and Discussion

Data Quantity vs. Model Complexity

A primary finding is that data quantity often outweighs model complexity. SqueezeNet, despite being much smaller, can outperform ResNet-50 if it has access to slightly more data. For instance, SqueezeNet trained on 2% of the data outperforms ResNet-50 trained on 1%. ResNet-50 only begins to consistently dominate once the dataset size exceeds 40% of the total subset.

Top-1 accuracy of ResNet-50 and SqueezeNet-1.0 across varying dataset sizes.

The "Clean vs. Collect" Tipping Point

The impact of label noise is significant; at 10% dataset size, 20% label noise causes accuracy to drop from 46.0% to 33.8%. However, the optimal strategy to mitigate this depends on the current dataset scale. For small, noisy datasets, labeling new images is more effective than cleaning existing ones. Conversely, for large, noisy datasets, cleaning existing labels becomes more cost-efficient than further data collection.

Comparison of strategies (adding data vs. cleaning labels) for small (10%) and large (60%) initial datasets with 20% mislabels.

Strategic Labeling and SSL Robustness

In imbalanced scenarios (e.g., 100:1), targeted labeling of rare classes is far more impactful than blind dataset expansion. Furthermore, SSL (FixMatch) proved remarkably robust, providing consistent performance boosts even when the labeled portion was corrupted by 20% mislabels. Using FixMatch with only 25% of the labels achieved 77.3% accuracy, nearly matching the 79.8% achieved with 100% labels under standard supervised training.

Conclusion

This research highlights that the "best" model depends heavily on the data regime. While larger models dominate in data-rich environments, smaller models are highly competitive when data is scarce. Future work should extend these findings to Vision Transformers (ViTs) and include broader cost-benefit analyses, such as compute costs and environmental impact.

Summary updated

Abstract

Research in computer vision is centered on algorithmic improvements, for example, by developing better models. Thereby, the data is considered fixed. This is in contrast to many real-world applications of computer vision systems in which algorithms and data co-evolve. To address this shortcoming of previous research, we study the properties of the data and their interaction with deep learning algorithms. Thereby, we investigate the size of the data, the share of mislabels, class imbalance and the presence of unlabeled data which can be leveraged using semi-supervised learning. In experiments on 100 classes from ImageNet, we show that a tiny network architecture outperforms a much more powerful one it if has access to only a little bit more data. Only if vast amounts of data are available so that adding even more images has little effect on performance, large architectures dominate smaller ones. If little data is provided, adding a few labeled images has a huge effect on accuracy. Once accuracy saturates, massive amounts of additional data are needed to achieve even small improvements. Furthermore, we find that mislabels severely reduce performance. To fix that, we propose a cost-efficient way of identifying mislabels which is especially beneficial if many images are already available. Conversely, if little data is available, labeling more images is more advantageous than cleaning existing annotations. In the case of imbalanced data, we illustrate that labeling more instances from rare classes has a much greater effect on performance than only increasing dataset size. Moreover, we show that leveraging unlabeled images by semi-supervised learning offers a consistent benefit even if the labeled subset contains significant label noise.
BibTeX
@inproceedings{pototzky2022investigating,
  author = {Daniel Pototzky and Azhar Sultan and Lars Schmidt-Thieme},
  title = {{Investigating the Interaction between Data and Algorithms}},
  booktitle = {Proceedings of the World Congress on Electrical Engineering and Computer Systems and Science},
  year = {2022},
  doi = {10.11159/mvml22.106},
}
Figure