Fernando Perez-Cruz

Safeguarding biodiversity requires a detailed understanding of the biosphere, especially regarding the distribution and state of habitats. Citizen science data offer a rich source of useful information, but they usually only contain partial records of species assemblages, as predominantly opportunistic observations of individual species are reported. We explored the capacity of recommender systems—a class of algorithms originally developed for tasks like recommending movies to users on streaming platforms—to fill gaps in citizen science data and assign unreported plant species to partial assemblages, using opportunistic presence‐only observations as training data. We employed a hierarchical Poisson factorization (HPF) model to learn distributions exclusively from pixel‐wise pooled observations, which were synthesized to a site‐by‐species matrix of a dimension of c . 80,000 × 2000 covering Switzerland. Based on about 1500 independent surveys, we assessed HPF's capacity to predict held‐out species and compared these to predictions from traditional species distribution models (SDMs). Moreover, we investigated HPF's latent features, that is, its low‐dimensional summaries of the similarity in species' occurrence across sites and the similarity in site suitability for species. Indicator species of most broad habitat categories distinguished in Switzerland clustered significantly in species‐related latent feature space, indicating a high capacity of HPF to approximate co‐occurrence. Patterns in site‐related latent features were primarily explained by climate, soil, and topography. However, about 40% of the signal could not be attributed to gridded predictors and likely mirrors small‐scale differences, for example, due to disturbances like those experienced in ruderal habitats. Predictive performance of HPF was close to that of SDMs if 20 (or 49% by median) species were observed at a site. However, when ensembling predictions of HPF and SDMs, performance could be improved for sites with ≥6 species observed. HPF predictions performed comparably well in disturbed habitats like roadside vegetation and comparably poorly in habitats experiencing elevated environmental filtering, such as subalpine forests. Given their ability to efficiently learn the distribution of entire floras based on presence‐only data, their unique capacity to obtain generalized co‐occurrence proxies from partial inventories, and their potential to boost distributional predictions, recommender systems can valuably contribute to improving our understanding of the biosphere.

Training high-quality deep learning models is a challenging task due to computational and technical requirements. A growing number of individuals, institutions, and companies increasingly rely on pre-trained, third-party models made available in public repositories. These models are often used directly or integrated in product pipelines with no particular precautions, since they are effectively just data in tensor form and considered safe. In this paper, we raise awareness of a new machine learning supply chain threat targeting neural networks. We introduce MaleficNet 2.0, a novel technique to embed self-extracting, self-executing malware in neural networks. MaleficNet 2.0 uses spread-spectrum channel coding combined with error correction techniques to inject malicious payloads in the parameters of deep neural networks. MaleficNet 2.0 injection technique is stealthy, does not degrade the performance of the model, and is robust against removal techniques. We design our approach to work both in traditional and distributed learning settings such as Federated Learning, and demonstrate that it is effective even when a reduced number of bits is used for the model parameters. Finally, we implement a proof-of-concept self-extracting neural network malware using MaleficNet 2.0, demonstrating the practicality of the attack against a widely adopted machine learning framework. Our aim with this work is to raise awareness against these new, dangerous attacks both in the research community and industry, and we hope to encourage further research in mitigation techniques against such threats.

Self-supervised learning (SSL) is a machine learning approach where the data itself provides supervision, eliminating the need for external labels. The model is forced to learn about the data’s inherent structure or context by solving a pretext task. With SSL, models can learn from abundant and cheap unlabeled data, significantly reducing the cost of training models where labels are expensive or inaccessible. In Computer Vision, SSL is widely used as pre-training followed by a downstream task, such as supervised transfer, few-shot learning on smaller labeled data sets, and/or unsupervised clustering. Unfortunately, it is infeasible to evaluate SSL methods on all possible downstream tasks and objectively measure the quality of the learned representation. Instead, SSL methods are evaluated using in-domain evaluation protocols, such as fine-tuning, linear probing, and k-nearest neighbors (kNN). However, it is not well understood how well these evaluation protocols estimate the representation quality of a pre-trained model for different downstream tasks under different conditions, such as dataset, metric, and model architecture. In this work, we study how classification-based evaluation protocols for SSL correlate and how well they predict downstream performance on different dataset types. Our study includes eleven common image datasets and 26 models that were pre-trained with different SSL methods or have different model backbones. We find that in-domain linear/kNN probing protocols are, on average, the best general predictors for out-of-domain performance. We further investigate the importance of batch normalization for the various protocols and evaluate how robust correlations are for different kinds of dataset domain shifts. In addition, we challenge assumptions about the relationship between discriminative and generative self-supervised methods, finding that most of their performance differences can be explained by changes to model backbones.

This work proposes a Graph Neural Network (GNN) modeling approach to predict the resulting surface from a particle based fabrication process. The latter consists of spray-based printing of cementitious plaster on a wall and is facilitated with the use of a robotic arm. The predictions are computed using the robotic arm trajectory features, such as position, velocity and direction, as well as the printing process parameters. The proposed approach, based on a particle representation of the wall domain and the end effector, allows for the adoption of a graph-based solution. The GNN model consists of an encoder-processor-decoder architecture and is trained using data from laboratory tests, while the hyperparameters are optimized by means of a Bayesian scheme. The aim of this model is to act as a simulator of the printing process, and ultimately used for the generation of the robotic arm trajectory and the optimization of the printing parameters, towards the materialization of an autonomous plastering process. The performance of the proposed model is assessed in terms of the prediction error against unseen ground truth data, which shows its generality in varied scenarios, as well as in comparison with the performance of an existing benchmark model. The results demonstrate a significant improvement over the benchmark model, with notably better performance and enhanced error scaling across prediction steps.

Regulation of chemicals requires knowledge of their toxicological effects on a large number of species, which has traditionally been acquired through in vivo testing. The recent effort to find alternatives based on machine learning, however, has not focused on guaranteeing transparency, comparability and reproducibility, which makes it difficult to assess advantages and disadvantages of these methods. Also, comparable baseline performances are needed. In this study, we trained regression models on the ADORE "t-F2F" challenge proposed in [Schür et al., Nature Scientific data, 2023] to predict acute mortality, measured as LC50 (lethal concentration 50), of organic compounds on fishes. We trained LASSO, random forest (RF), XGBoost, Gaussian process (GP) regression models, and found a series of aspects that are stable across models: (i) using mass or molar concentrations does not affect performances; (ii) the performances are only weakly dependent on the molecular representations of the chemicals, but (iii) strongly on how the data is split. Overall, the tree-based models RF and XGBoost performed best and we were able to predict the log10-transformed LC50 with a root mean square error of 0.90, which corresponds to an order of magnitude on the original LC50 scale. On a local level, on the other hand, the models are not able to consistently predict the toxicity of individual chemicals accurately enough. Predictions for single chemicals are mostly influence by a few chemical properties while taxonomic traits are not captured sufficiently by the models. We discuss technical and conceptual improvements for these challenges to enhance the suitability of in silico methods to environmental hazard assessment. Accordingly, this work showcases state-of-the-art models and contributes to the ongoing discussion on regulatory integration.

Regulation of chemicals requires knowledge of their toxicological effects on a large number of species, which has traditionally been acquired through in vivo testing. The recent effort to find alternatives based on machine learning, however, has not focused on guaranteeing transparency, comparability and reproducibility, which makes it difficult to assess advantages and disadvantages of these methods. Also, comparable baseline performances are needed. In this study, we trained regression models on the ADORE “t-F2F” challenge proposed in [Schür et al ., Nature Scientific data , 2023] to predict acute mortality, measured as LC50 (lethal concentration 50), of organic compounds on fishes. We trained LASSO, random forest (RF), XGBoost, Gaussian process (GP) regression models, and found a series of aspects that are stable across models: (i) using mass or molar concentrations does not affect performances; (ii) the performances are only weakly dependent on the molecular representations of the chemicals, but (iii) strongly on how the data is split. Overall, the tree-based models RF and XGBoost performed best and we were able to predict the log10-transformed LC50 with a root mean square error of 0.90, which corresponds to an order of magnitude on the original LC50 scale. On a local level, on the other hand, the models are not able to consistently predict the toxicity of individual chemicals accurately enough. Predictions for single chemicals are mostly influenced by a few chemical properties while taxonomic traits are not captured sufficiently by the models. We discuss technical and conceptual improvements for these challenges to enhance the suitability of in silico methods to environmental hazard assessment. Accordingly, this work showcases state-of-the-art models and contributes to the ongoing discussion on regulatory integration. Environmental significance: Conventional environmental hazard assessment in its current form will not be able to adapt to the growing need for toxicity testing. Alternative methods, such as toxicity prediction through machine learning, could fulfill that need in an economically and ethically sound manner. Proper implementation, documentation, and the integration into the regulatory process are prerequisites for the usability and acceptance of these models.

Legislative bodies generally keep records of their activities. While the digitization wave spurred the availability of archival documents, their processing remains a challenge. The Swiss parliamentary records are no exception.In this paper we present a supervised pipeline for extracting and structuring of content of archival records. Our pipeline consists of five steps, starting with an assessment of which elements need extraction and how they relate to each other. Step two involves general pre‐processing to prepare the PDF documents and is followed by an element classification step. Step four involves post‐processing and the final step is a validation of the extracted information. With our supervised approach, we are able to process over 200,000 pages of Swiss parliamentary records (spanning the years 1891–1995), a feat that would exceed the budget of most projects using manual curation. We discuss validation of individual steps and offer guidance to researchers engaged in similar data processing efforts.

We propose a novel algorithm for data augmentation in nonlinear over-parametrized regression. Our data augmentation algorithm borrows from the literature on causality and extends the recently proposed Anchor regression (AR) method for data augmentation, which is in contrast to the current state-of-the-art domain-agnostic solutions that rely on the Mixup literature. Our Anchor Data Augmentation (ADA) uses several replicas of the modified samples in AR to provide more training examples, leading to more robust regression predictions. We apply ADA to linear and nonlinear regression problems using neural networks. ADA is competitive with state-of-the-art C-Mixup solutions.