Counterfactual explanations for land cowl mapping: interview with Cassio Dantas

Of their paper Counterfactual Explanations for Land Cowl Mapping in a Multi-class Setting, Cassio Dantas, Diego Marcos and Dino Ienco apply counterfactual explanations to distant sensing time collection information for land-cover mapping classification. On this interview, Cassio inform us extra about explainable AI and counterfactuals, the group’s analysis methodology, and their fundamental findings.

What’s the subject of the analysis in your paper?

Our paper falls into the rising subject of explainable synthetic intelligence (XAI). Regardless of the performances achieved by current deep studying approaches, they continue to be black-box fashions with restricted understanding of their inner habits. To enhance common acceptability and trustworthiness of such fashions, there’s a rising want to enhance their interpretability and make their decision-making processes extra clear. Extra exactly, to make the black field extra gray.

On this analysis work we discover a specific kind of method known as counterfactual explanations, which goals to explain the habits of a mannequin by offering minimal modifications to the enter information that will end in sensible samples belonging to a unique class.

In regards to the area of utility, we concentrate on distant sensing time collection information within the context of a land-cover mapping classification. Of explicit significance, this process gives essential data for making knowledgeable coverage, growth, planning, and useful resource administration selections.

Might you inform us concerning the implications of your analysis and why it’s an attention-grabbing space for examine?

The proposed strategy contributes to elevating the interpretability of present deep studying strategies. This line of analysis turns into more and more necessary as deep studying fashions grow to be widespread. Within the distant sensing area, as an example, such fashions have proven spectacular outcomes on a wide range of duties similar to picture super-resolution, picture restoration, biophysical variable estimation and land cowl classification from satellite tv for pc picture time collection (SITS) information. For such strategic functions, it’s important {that a} mannequin’s prediction comes accompanied with human-interpretable explanations that can assist enhance reliability on the mannequin’s selections in addition to serving to understanding its weaknesses and limitations.

Counterfactual explanations are a strong software on this context, as they assist characterize a mannequin’s resolution boundary, in addition to offering actionable pointers on methods to modify a given prediction with minimal effort, which could be fairly helpful in a number of utility instances.

Nevertheless, offering beneficial counterfactual explanations just isn’t trivial in apply, as they need to ideally fulfil a number of standards, similar to: 1) proximity, i.e. being as shut as doable to the enter information pattern. 2) plausibility, i.e. resembling some actual information that might truly happen in apply; 3) sparsity, i.e. selling modifications to some options solely, in an effort to be extra amenable to motion. Within the case of time collection, as an example, it’s extremely fascinating for the counterfactual clarification to perturb solely brief and contiguous segments of the timeline in an effort to make the interpretation simpler for the area professional.

It may be difficult to offer an computerized counterfactual technology strategy that fulfils all of the above-mentioned necessities. Because of this, it stays an energetic analysis space these days. Our proposed strategy is a brand new try to effectively deal with such technical challenges.

Might you clarify your methodology?

We suggest a counterfactual technology strategy in a multi-class land cowl classification setting for satellite tv for pc picture time collection information. The proposed strategy generates counterfactual explanations which can be believable (i.e. belong as a lot as doable to the information distribution) and near the unique information (modifying solely a restricted and contiguous set of time entries by a small quantity).

One other distinctive function of the proposed strategy (in comparison with different current counterfactual clarification approaches for time collection information) is the shortage of prior assumption on the focused class for a given counterfactual clarification. This inherent flexibility permits for the invention of attention-grabbing data on the connection between land cowl lessons.

To make sure the plausibility of the generated explanations, we depend on a GAN (generative adversarial community)-inspired structure which is proven in Fig.1.

Fig.1: Schematic illustration of the proposed strategy.

Given a pre-trained Classifier, a counterfactual clarification is obtained for every enter time collection by including a perturbation to the unique sign. Such perturbation is generated by the Noiser module, a easy multilayer perceptron with two hidden layers, which is realized with the purpose to swap the prediction of the Classifier. Lastly, we embrace a Discriminator module which is educated to establish unrealistic counterfactuals in a two-player recreation towards the Noiser module that acts as a the generator on this adversarial coaching scheme. Lastly, to maintain perturbations concentrated in a small and contiguous timeframe we make use of a weighted L1-norm penalization.

What had been your fundamental findings?

We utilized the proposed method to NDVI (Normalized Differential Vegetation Index) time collection derived from Sentinel-2 satellite tv for pc photographs spanning over the yr 2020 and overlaying a 2338 km2 space across the city of Koumbia, within the south-west of Burkina Faso. The enter information consists of eight land cowl lessons, together with three crop-types (cereals, cotton, oleaginous), three vegetation varieties (grassland, shrubland, forest), naked soil and water.

As a result of the proposed strategy doesn’t predefine a goal class for the generated counterfactuals, it permits enter samples from a sure class to freely split-up into a number of goal lessons. Class transitions obtained in such a means (see Fig. 2) as to convey up beneficial insights on the relation between lessons.

For example, in Fig. 2, the three crop-related lessons (cereals, cotton and oleaginous) kind a really coherent cluster, with virtually all transitions staying inside the sub-group. The vegetation lessons shrubland and forest are most frequently despatched to at least one one other, whereas grassland stays a lot nearer to the crop lessons (oleaginous). The naked soil class can also be most frequently reworked into oleaginous. Lastly, the water class could be very not often modified by the counterfactual studying course of, which is considerably anticipated as a consequence of its very distinct attribute (NDVI signature) in comparison with the opposite lessons.

Fig. 2: Class transitions induced by the counterfactuals (B. stands for naked soil and W. for water).

Two illustrative examples of counterfactual explanations are proven in Fig. 3. It’s attention-grabbing to notice the similarity between the generated counterfactual and an actual information pattern from the identical class (on the neighboring plot). To remodel a shrubland pattern into forest, NDVI is added between the months of July and October. The alternative is completed to acquire the reverse transition, which matches the final data of such land cowl lessons on the thought-about examine space.

Fig. 3: Examples of actual time collection with corresponding counterfactual from lessons shrubland and forest.

To quantify to what extent the proposed counterfactual explanations match the unique information distribution, we ran an anomaly detection algorithm on the generated explanations. A powerful fee of 88.9% of the generated counterfactuals had been recognized as inliers, in comparison with solely 72.6% for a non-adversarial variant of the proposed community. The obtained outcomes clearly present that counterfactual plausibility is achieved because of the adversarial coaching course of.

Lastly, we additionally verified that the proposed weighted L1-norm regularization efficiently enforces sparsity of the counterfactual explanations apart from controlling its proximity with the enter information pattern.

What additional work are you planning on this space?

The rising subject of counterfactual explanations nonetheless affords thrilling analysis avenues to be explored. Particularly in the case of much less typical forms of information (examine to photographs), for the reason that early developments on this space principally emanated from the pc imaginative and prescient group. With regards to time collection information, as an example, a lot work stays to be executed, much more so for distant sensing time collection information.

As a doable future work, it could be attention-grabbing to increase our proposed framework to the case of multivariate time collection information and consider the methodology on different functions coming from totally different domains wherein time collection information are outstanding. One other thrilling analysis course can be to leverage the suggestions supplied by the generated counterfactual samples relating to probably the most frequent class confusions to enhance the robustness of the classifier itself.

About Cassio F. Dantas

Cassio F. Dantas is a analysis scientist at INRAE, TETIS, Montpellier, France. Between 2020 and 2022, he was a postdoctoral researcher at IRIT (laptop science laboratory of Toulouse) and IMAG (arithmetic laboratory of the College of Montpellier). He carried out his PhD research at Inria in Rennes, France, and obtained his diploma on sign, picture and imaginative and prescient in 2019. His present analysis pursuits embrace interpretable synthetic intelligence, optimization algorithms and machine studying for distant sensing information with functions to agriculture, ecosystems and atmosphere.