# Using a Cross-Task Grid of Linear Probes to Interpret CNN Model Predictions On Retinal Images

arXiv: https://arxiv.org/abs/2107.11468 · HTML: https://mlsanity.com/paper/2107.11468/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Interpreting vision and multimodal models, one of 13 directions in Mechanistic interpretability.
- Builds on 8 of its 19 references that are in the map.
- First posted on arXiv 23 Jul 2021.

## Placed in

- Mechanistic interpretability / [Interpreting vision and multimodal models](https://mlsanity.com/mechinterp/interpreting-vision-and-multimodal-models/)

## Builds on (most cited first, 8 of 8)

- [Representation Learning with Contrastive Predictive Coding](https://mlsanity.com/paper/1807.03748/) (2018)
- [Understanding intermediate layers using linear classifier probes](https://mlsanity.com/paper/1610.01644/) (2016)
- [Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)](https://mlsanity.com/paper/1711.11279/) (2017)
- [Designing and Interpreting Probes with Control Tasks](https://mlsanity.com/paper/1909.03368/) (2019)
- [Self-labelling via simultaneous clustering and representation learning](https://mlsanity.com/paper/1911.05371/) (2019)
- [Predicting Cardiovascular Risk Factors from Retinal Fundus Photographs using Deep Learning](https://mlsanity.com/paper/1708.09843/) (2017)
- [Deep learning for predicting refractive error from retinal fundus images](https://mlsanity.com/paper/1712.07798/) (2017)
- [Internal representation dynamics and geometry in recurrent neural networks](https://mlsanity.com/paper/2001.03255/) (2020)

## Abstract (arXiv)

We analyze a dataset of retinal images using linear probes: linear regression models trained on some "target" task, using embeddings from a deep convolutional (CNN) model trained on some "source" task as input. We use this method across all possible pairings of 93 tasks in the UK Biobank dataset of retinal images, leading to ~164k different models. We analyze the performance of these linear probes by source and target task and by layer depth. We observe that representations from the middle layers of the network are more generalizable. We find that some target tasks are easily predicted irrespective of the source task, and that some other target tasks are more accurately predicted from correlated source tasks than from embeddings trained on the same task.