Explanations from Humans for Humans: The Potential of Transparent AI for Trust Calibration and Didactics

Table of Contents

• Installation
• Introduction
• Data Acquisition
• Data Annotation
• Data Preprocessing
• Training and Validation
• Preliminary XAI Experiments without Humans
• Concept-based Prototypical Nearest Neighbors (CoProNN)

Installation

For reading the segmentation masks annotated in Label Studio, you will need:

• labelstudio-converter

For building the Deep Models for image classification and plotting the feature attribution maps:

• wingbeats

For computing localisation and faithfullness metrics of the saliency maps:

• quantus

For computing the saliency maps and TCAV-scores (depending on the experiment/notebook):

• tf-explain, innvestigate and xplique

For generating concept images (used in the Human-in-the-Loop (HIL) experiments):

• Stable Diffusion v1.5 from diffusers

Only for running the diffuser you will need PyTorch 1.10. The rest of the code is built on TensorFlow 2.8 and Python 3.10. (see requirements).

Introduction

A study published in 2017 revealed a shocking 75% decline in insect biomass in German nature protected areas over the past 27 years, mainly because of pesticides, intensive farming and climatic changes. But not only the overall number of insects has dramatically sunken. The 33,000 different species that used to find their habitat in Germany are rapidly vanishing. Multiple organizations along with the German Ministry of Environment have joined forces to curb the dramatic decrease of insect diversity. Steps have already been taken to determine which areas are primarily affected by insect depopulation and which species face a high risk of extinction (see Insektensommer and DINA - articles in German). This is called insect monitoring and, given the current availability of technology, the goal is to construct automatic systems able to recognize insect species without needing to trap and kill any insects.

Automated insect monitoring requires accurate computer vision (CV) models and development of these models, in turn, requires high quality datasets. Building such datasets is difficult, as there are few domain experts (entomologists) that can annotate images of rare species.

In an ideal human-machine collaborative setting, a Machine Learning (ML) algorithm could assist human experts annotating images of rare species by automatically annotating some images and referring others to domain experts. When working with such complex structures as an insect taxonomic tree, model explainability becomes a requirement rather than a nice-to-have feature. The entomologists delegated to supervise the monitoring procedure need to understand what exactly led their model to the given output.

Data Acquisition

The script webscraper_inat.py downloads photos of bees from iNaturalist. The user only needs to specify in the main section the target folder where the photos should be downloaded, a maximum number of photos to search for and the corresponding url index of the species.

For instance, to find the index for the species Anthidium manicatum, type the species name in the search bar of the iNaturalist site and copy the number at the end of the url (it should be 62453).

We checked that the wild bee dataset (both mini and full) did not contain duplicates from the iNat Challenge 2021, since we may be interested in utilizing models pretrained on that task. We also checked for duplicates within our own datasets and it turned out that the scraper did indeed download some images twice (either because of the Internet connection going down or because the iNat users themselves uploaded the same photo more than once). These duplicates were subsequently removed.

Data Annotation

From the downloaded images a subset of roughly 30 samples per species (further also referred to as the mini dataset) has been selected and further annotated in Label Studio. The mini dataset eventually contained 726 images for 25 bee species. More specifically, the insects' main body parts have been segmented through brushing, as in the image below:

The annotated images and segmentation masks can be downloaded from here. Further details on how to decode the mask coordinates follow below.

Data Preprocessing

The scripts create_metafiles_mini.py and create_metafiles_all.py create metafiles from the json-files downloaded from Label Studio similar to the CUB200 format. The files created are:

• classes.txt (for mini and whole dataset) - each class name is allocated a unique identifier from 1 to 22 or 25. The reason why we had a variable number of classes is because of four species that are near to impossible to tell apart based solely on image input. These species are Bombus lucorum, Bombus cryptarum, Bombus terrestris and Bombus magnus. Our tests have revealed that, indeed, models also struggle to differentiate these species. Consequently, we decided to compress the images from the 4 species under a common label - Bombus lucorum, which reduced the number of classes down to 22.
• images.txt (mini and whole) - each jpg-file is given a unique identifier from 1 to length_of_dataset
• image_class_labels.txt (mini and whole) - each file ID is mapped to the corresponding class ID
• parts.txt (mini) - each relevant body part is given an ID from 1 to 3
• part_locs.txt (mini) - each body part from each file is mapped to its RLE coordinates; each row starts with the file ID followed by the body part ID followed by the list of RLE coordinates

Additional annotations have also been created in derive_contours_bboxes.py:

• bounding_boxes.txt - each file ID is associated with the bounding box coordinates of the whole object
• part_contours.npy - contour indexes of the binary masks for every body part. The reason behind it is that decoding RLE coordinates is relatively time-consuming; therefore, it is preferable to compute the masks only once and store them for later training; however, storing large sparse matrices or their positive indexes is not very efficient and since Numpy or OpenCV save files in uint8 at least, the contours of the masks were preferred. At training/inference time, the full object masks can easily be retrieved by refilling the contour lines. See an example. Note that the first entry in the arrays in this file is always a list of [file_id, part_id].

Based on the meta files, crop_mask_images.py derives separate datasets from the original mini set by applying different input transformations such as masking and cropping to bounding box. See examples.

Important Note on the RLE Format

Beware that the RLE format used by Label Studio to store segmentation masks is not the same commonly known RLE format from the COCO dataset. Unlike the COCO standard, the RLE numbers here are not interpreted as pixel counts. Please follow this issue for further details.

Decoding RLE into Segmentation Masks

For translating the RLE coordinates into segmentation matrices, the library label-studio-converter was used. Every image is accompanied by three part annotations:

• Head - includes the whole head of the insect along with its tentacles
• Thorax - middle body together with legs and wings
• Abdomen - lower body including needle, if present

The picture below shows the segmentation masks - binary and overlaid - in the case of an Halictus scabiosae. More examples can be found here. The mask of the whole object has been computed by means of logical union of the part masks. Note that it sometimes occured that small pixel patches from within the part segment were not brushed over by the annotator. This was corrected through a morphological process called closing, which is dilation followed by erosion.

From the object mask, the bounding box can also be easily derived (see annot_computer.py). The convention agreed upon here for storing the bounding box coordinates is the same as in the CUB200 dataset: (x_min, y_min, width, height).

Training and Validation

We have trained a ResNet50 initialized with a backbone pretrained on the iNat 2021 challenge complemented with a linear classifier on the top. The whole dataset has been used for training and cross validation, while the mini dataset was kept as test set. We report 0.78 top-1 and 0.95 top-3 accuracy on the test set, which is competitive with similar state-of-the-art fine-grained models - Horn et al..

More details regarding the model architecture, hypertuning process, as well as training and validation scores can be found in the cooresponding notebooks, plots for training curves, ROC-AUC curves, worst predictions and on our freely online available TensorBoard Experiment.

Preliminary XAI Experiments without Humans

Motivation and Related Work

When working with such complex structures as an insect taxonomic tree, model explainability becomes a requirement rather than a nice-to-have feature. The entomologists delegated to supervise the monitoring procedure need to understand what exactly led their model to the given output. Nowadays, there is a wide array of XAI methods and going into the details of all the possible categorisations would exceed the scope of this preliminary work.

The XAI community has been working arduously in the past years to develop new XAI methods - Guidotti et al. - as well as to formalize the notions of explanation and explainability - Doshi et al. It has been acknowledged that a key requirement to better these methods are quality metrics for XAI, so that model explanations can be evaluated and compared to one another - Adebayo et al. 2018, Adebayop et al. 2020, Kindermans et al. Next to automated evaluations of XAI methods, researchers also investigated XAI methods in human-in-the-loop settings - Hase and Bansal, but it remains unclear to what extent these two types of evaluations are related - Bießmann and Refiano, Borowski et al.

For conducting and evaluating the upcoming XAI experiments, the following libraries were used: tf-explain, iNNvestigate (for computing the explanations) and quantus (for evaluating them). In the initial stages of our research, we decided to focus our attention on post hoc feature attribution based methods, also known as saliency maps (see example below).

The maps are normalized from 0 (blue) to 1 (red).

Evaluation

The authors of quantus define a simple set of 'desiderata', that is, properties that good XAI methods are expected to have: faithfulness, localisation, robustness, complexity, randomisation and axiomatic properties. For our work, we decided to first investigate the first two and provide initial results.

Localisation via Segmentation Masks

We have used the segmentation masks drawn manually as the ground truth for our model's explanations, measuring to what extent the 'hot pixels' fall into our masks. The results are shown in the table below. We note that the model achieves higher scores on metrics such as Pointing Game and AUC, that evaluate localisation on a coarse level. On the other hand, the model performs rather poorly on the other 'finer-grained' metrics, which compute a more in-depth assessment of how well the salient regions overlap with the segmentation masks.

Metric	Gradient	Grad. * Input	Int. Grad.
Pointing Game	0.9463	0.9421	0.9298
Attribution Localisation	0.2984	0.2551	0.3168
Top K Intersection	0.3320	0.2807	0.3547
Relevance Rank Accuracy	0.3047	0.2587	0.3231
AUC	0.7157	0.6671	0.7297

Faithfulness via Monte-Carlo Dropout and Pixel Flipping

Inspired by Bykov et al., we evaluated the faithfulness of the explanations via Pixel Flipping (PF) in conjunction with the model's prediction uncertainty via Monte-Carlo Dropout(MCD). By activating the Dropout layer during inference, the network predicts class probability distributions instead of pointwise estimations. For instance, in the figure below, the breadth of the histogram is indicative of the model's predictive uncertainty.

Example of Pixel Flipping w.r.t. 3 Saliency Maps

Pixel FLipping presumes replacing certain pixels in the input and re-evaluating the model's prediction. In the two examples below, we see an example of flipping pixels in a sample image randomly and according to highlighted pixels in saliency maps. The replacement pixels are generated randomly from a normal distribution computed over all the images of the test set.

Example of Quantile Maps generated by MC-Dropout

For every new (Monte-Carlo) prediction, a new saliency maps is constructed by each of the three methods. The pixels that are marked as relevant throughout all the saliency quantiles are supposed to represent stable relevant features.

Pixel Flipping via MC-Dropout for a whole class in the test set

The quantile maps from the figure above serve here as markers of relevant features in the image. By iteratively masking these relevant pixels, we can judge how important they truly are for identifying the correct class. The greater their relevance, the steeper the decrease in the prediction scores. We would expect good XAI methods to exhibit a rapid and steep decrease in the accuracy curves, as more and more 'relevant' pixels are flipped; moreover, the saliency flipping curves would be expected to fall below the random flipping curve already in the early flipping iterations (like in the plot for Xylocopa violacea). Nonetheless, most of our plots were rather similar to Andrena bicolor, where saliency flipping does not appear to be more efficient than random flipping. On the contrary, sometimes it was visibly less efficient.

Concept-based Prototypical Nearest Neighbors (CoProNN)

Paper accepted at WCXAI24 in Valletta. Download here.

We have developed a novel concept-based post-hoc XAI method. It leverages recent progress in generative text2img models i.e. Stable Diffusion to generate high-level concept images to explain the prediction of a model via kNN. The pipeline is shown below:

To validate the utility of our method, we conducted a user study where participants were asked to classify wild bee species either while receiving our model's prediction as a hint (Control Group) or both the model's prediction and a CoProNN-explanation for it (XAI Group). It turns out that explanations helped users classify the bees more accurately and spot wrong model predictions more easily.

A demo of our human-in-the-loop experiment is freely available online. Also refer to our paper (presented at the XAI Workshop at ECML 24, Vilnius) where we describe the experimental setup and the results in greater detail.