Search

A variety of methods exist to explain image classification models. However, whether they provide any benefit to users over simply comparing various inputs and the model's respective predictions remains unclear. We conducted a user study (N=240) to test how such a baseline explanation technique performs against concept-based and counterfactual explanations. To this end, we contribute a synthetic dataset generator capable of biasing individual attributes and quantifying their relevance to the model. In a study, we assess if participants can identify the relevant set of attributes compared to the ground-truth. Our results show that the baseline outperformed concept-based explanations. Counterfactual explanations from an invertible neural network performed similarly as the baseline. Still, they allowed users to identify some attributes more accurately. Our results highlight the importance of measuring how well users can reason about biases of a model, rather than solely relying on technical evaluations or proxy tasks. We open-source our study and dataset so it can serve as a blue-print for future studies. For code see, https://github.com/berleon/do_users_benefit_from_interpretable_vision, Published at ICLR 2022

Human interpretability of deep neural networks' decisions is crucial, especially in domains where these directly affect human lives. Counterfactual explanations of already trained neural networks can be generated by perturbing input features and attributing importance according to the change in the classifier's outcome after perturbation. Perturbation can be done by replacing features using heuristic or generative in-filling methods. The choice of in-filling function significantly impacts the number of artifacts, i.e., false-positive attributions. Heuristic methods result in false-positive artifacts because the image after the perturbation is far from the original data distribution. Generative in-filling methods reduce artifacts by producing in-filling values that respect the original data distribution. However, current generative in-filling methods may also increase false-negatives due to the high correlation of in-filling values with the original data. In this paper, we propose to alleviate this by generating in-fillings with the statistically-grounded Knockoffs framework, which was developed by Barber and Cand\`es in 2015 as a tool for variable selection with controllable false discovery rate. Knockoffs are statistically null-variables as decorrelated as possible from the original data, which can be swapped with the originals without changing the underlying data distribution. A comparison of different in-filling methods indicates that in-filling with knockoffs can reveal explanations in a more causal sense while still maintaining the compactness of the explanations., Comment: 12 pages, 10 figures

This paper aims to synthesize the results of the two-year experience in implementation of the decision support system for urban development. CRISALIDE project (City Replicable and Integrated Smart Actions Leading Innovation to Develop Urban Economies) started in October 2018 and was one of the very few projects financed between E.U. and Russian Federation through the ERA NET RUS PLUS (ENRP) program. It was also the only financed project in ENRP dealing with urban planning topics in this current E.U. programming period (2013-2020). The project became possible due to the rising requests for innovations in urban life resulted in a noticeable shift in a political discourse towards the innovative ICT led economy, the new digital technologies, and the smart city policy in Russia. Though new initiatives seem too focused on technological solutions and lack a comprehensive understanding of smart development, they provoke the appearance of public discussions of the mentioned issues in cities, which themselves are the ecosystems for developing innovation. The missing components might be provided from the bottom level by using the place-based approach and implementing smart planning tools responding to the wicked local problems. This assumption lies in the basis of the CRISALIDE project aiming at bringing together technological, social, and organizational innovations. CRISALIDE was experimenting in the Russian city Rostov-on-Don through a collaborative approach involving E.U. and Russian researchers, creating an innovative digital platform to facilitate the renewal and regeneration of brownfields. The first selected experimental field was the area of the old airport relocated outside of the city. Attracting notable attention of different local and regional actors due to its size, location, marketing potential, and regional significance, the selected brownfield helped involve local experts, activists, and policy-makers to develop an innovation tool and raise a public discussion on urban development. A series of consortium events, new collaborations, and discussions became essential for designing future development scenarios. CRISALIDE system has an open, flexible structure based on urban ontology. It is a variant of artificial intelligence and will serve diverse issues in different cities. Oriented towards support in three main activities in urban development – new construction, reconstruction, and services provision – the system functions in three different modes. Three modes allow implementing required levels of public participation: from the internal communication within city administration to the public discussion involving all citizens., CITIES 20.50 – Creating Habitats for the 3rd Millennium: Smart – Sustainable – Climate Neutral. Proceedings of REAL CORP 2021, 26th International Conference on Urban Development and Regional Planning in the Information Society, 213-221

The approach to decision theory has varied from one scholar and school of thought to the other from the development of mathematically complex systems, models and equations to the disregard of this particular academic endeavor altogether due to randomness and overwhelming subjectivity and complexity. This paper shall briefly cite the main opposing perspectives on the general decision-making model, and in so doing argue whether the study of decisions makes academic sense at all. Concluding as to not only the pertinence, but the necessity of decision science, we do however admit the need for a more comprehensive and integrated approach. We then attempt to draw the guidelines of a general decision model, and later list some of the factors that such a model should integrate in order to constitute a realistic point of reference.

Soils are a major component of the global carbon budget and may serve an important role in mitigating increasing atmospheric CO2 through their capacity to store carbon. In this regard, it is important to evaluate the implications of forest management on changes in carbon cycling and sequestration and to determine the magnitude by which the efflux of CO2 from the soil surface can vary in time and space. For this study, soil CO2 efflux was measured in 5 replicate blocks of naturally regenerated and planted loblolly pine (Pinus taeda), shortleaf pine (Pinus echinata), and eastern white pine (Pinus strobus) in a 50-acre clear-cut on the Virginia Piedmont. Rates of CO2 efflux were measured every 2 weeks immediately adjacent and away (1m) from newly planted seedlings and cut stumps using a dynamic, closed-chamber infrared gas analyzer system. For each measurement date, volumetric water content was taken in the top 17cm, using time domain reflectometry (TDR) and soil-surface temperature was recorded in the top 7cm, using a temperature probe. For the October measurement a 12cm depth soil core (7cm diameter) was collected for each location. Carbon, nitrogen, coarse fragments, roots, surface litter and coarse woody debris were measured separately for each core. Position (near and away from seedling) had a strong effect on soil CO2 efflux rates. For the first measurement date, rates were higher near the newly planted seedlings (3.09μmol/m2/s) than those taken away from the seedlings (2.29 μmol/m2/s).. The same trend was maintained for the CO2 efflux rates measured near a cut stumps (3.51μmol/m2/s) and those taken away from the stump base (2.56μmol/m2/s). Species proved to have no significant effect on respiration rates for any date and no interaction between species and position was observed. Regression analysis was used to model the influence of soil and plant factors on efflux rates. Temperature (29.2%), position (near and away from the seedlings and stumps base)* temperature, (7.7%), soil carbon (4.1%), organic matter (1.6%), and soil moisture (0.7%) proved to be the major drivers for soil respiration (R2 = 0.4329). When only data near seedlings or stumps were modeled, species had a significant effect on soil CO2 efflux rates. The largest seedlings, loblolly pine (100 cm3 seedling value), had on average the highest rates followed by shortleaf pine (30 cm3 seedling value) and eastern white pine, which were the smallest (9 cm3 seedling value). Stumps had the highest efflux rates. The mean soil respiration rate measured over a seven month sampling period was 2.58 μmol/m2/s,, while the calculated carbon loss from the soil over the same period added up to 575 g C/ m2. The annual carbon loss was estimated to be 675 g C/ m2. Master of Science

Decision-making is becoming increasingly complex as the number of stakeholders rises,the number of decision-makers rises as well, the implications to consider are various and far-reaching,risks and uncertainties abound and the regulatory framework is growing increasingly complex.Decision-making is caught between the need to thoroughly document and substantiate a choice in frontof oneself and others, on one hand, and the potential traps that increase in number and impact as thecomplexity of the decision-making process increases, on the other hand. The present paper sought topresent critically the main types of downfalls that a decision-making model may face at the level of anindividual decision-maker. Our approach was to put into relation some of the most notablecontributions from various perspectives in order to obtain an integrated overview of the challengesfaced by decision sciences in formulating a both thorough and life-like decision-making model. Weidentified with this occasion three main potential weaknesses in formal decision modeling:unquantifiable factors, such as feelings or morals, excessive formalism and simplification of modelsthat make them vulnerable to paradoxes, and finally cognitive biases that deviate models from anobjective path. Though scientific literature on each of these topics abounds, it is time to integrate it andset the basis for comprehensive and multi-perspective modeling of decisions.