MLwell lab

The xgbGAMView allows learning GAMs with a familiar scikit-learn interface. The GAMs use xgboost as the underlying engine to learn the model and offer visualization as well as graph smoothing options. The library can be installed from PyPi (pip install xgbGAMView).

The impact of long-term Coronavirus disease 2019 (COVID-19) on the pediatric population is still not well understood. This study was designed to estimate the magnitude of COVID-19 long-term morbidity 3–6 months after the date of diagnosis.

When dealing with tabular data, models based on decision trees are a popular choice due to their high accuracy on these data types, their ease of application, and explainability properties. However, when it comes to graph-structured data, it is not clear how to apply them effectively, in a way that in- corporates the topological information with the tabular data available on the vertices of the graph. To address this challenge, we introduce TREE-G. TREE-G modifies standard decision trees, by introducing a novel split function that is specialized for graph data. Not only does this split function incorporate the node features and the topological information, but it also uses a novel pointer mechanism that allows split nodes to use information computed in previous splits. Therefore, the split function adapts to the predictive task and the graph at hand. We analyze the theoretical properties of TREE-G and demonstrate its benefits empirically on multiple graph and vertex prediction benchmarks. In these experiments, TREE-G consistently outperforms other tree-based models and often outperforms other graph-learning algorithms such as Graph Neural Networks (GNNs) and Graph Kernels, sometimes by large margins. Moreover, TREE-Gs models and their predic tions can be explained and visualized.

Congratulations to Belle Kriger, Hagar Rosenblatt, Chaya Ben-Yehuda, and Yarin Udi for completing their Master's degrees

Graphs, such is protein interaction graphs, social contact graphs, and molecular structure graphs, are commonly used. Learning over graphs is challenging and is done using Graph Neural Networks (GNNs) in most cases. We introduce a new method that uses trees and gradient boosting for the task. It outperforms GNNs in many cases. Try it yourself

The field of Deep Neural Nets (DNNs) is still evolving and new architectures are emerging to better extract information from available data. The Information Bottleneck, IB, offers an optimal information theoretic framework for data modeling. However, IB is intractable in most settings. In recent years attempts were made to combine deep learning with IB both for optimization and to explain the inner workings of deep neural nets. VAE inspired variational approximations such as VIB became a popular method to approximate bounds on the required mutual information computations. This work continues this direction by introducing a new tractable variational upper bound for the IB functional which is empirically tighter than previous bounds. When used as an objective function it enhances the performance of previous IB-inspired DNNs in terms of test accuracy and robustness to adversarial attacks across several challenging tasks. Furthermore, the utilization of information theoretic tools allows us to analyze experiments and confirm theoretical predictions in real world problems.

How should courts act upon deep-fakes? should legislators modify existing laws to address it? In a new paper (in Hebrew) we study these issues and suggest that adding clauses to existing legislation specific to deep-fake might create problems in laws that do not add such clauses due to the principle of narrow interpretation.

Congratulations to Daniel Sarusi, Neta Biran, Lotan Hacohen, Yuval Reingold, and Yuval Argoetti on very successful presentations of their undergrad projects. Click the logo ← to see a short video (in Hebrew) about Yuval A.'s project.

Explainability has been proposed as a possible solution to some of the risks emerging from recent advances in AI. In this paper, we study explanations from legal point of view and show that many of the reasons for requiring explanations cannot be fulfilled by AI systems. Moreover, in some cases, these explanations can increase risks instead of mitigating them.

Congratulations to Nimrod Harel for successfully defending his Master's thesis

Remote assessment of the gait of older adults (OAs) during daily living using wrist-worn sensors has the potential to augment clinical care and mobility research. However, hand movements can degrade gait detection from wrist-sensor recordings. To address this challenge, we developed an anomaly detection algorithm and compared its performance to four previously published gait detection algorithms. Multiday accelerometer recordings from a wrist-worn and lower-back sensor (i.e., the “gold-standard” reference) were obtained in 30 OAs, 60% with Parkinson’s disease (PD). The area under the receiver operator curve (AUC) and the area under the precision–recall curve (AUPRC) were used to evaluate the performance of the algorithms. The anomaly detection algorithm obtained AUCs of 0.80 and 0.74 for OAs and PD, respectively, but AUPRCs of 0.23 and 0.31 for OAs and PD, respectively. The best performing detection algorithm, a deep convolutional neural network (DCNN), exhibited high AUCs (i.e., 0.94 for OAs and 0.89 for PD) but lower AUPRCs (i.e., 0.66 for OAs and 0.60 for PD), indicating trade-offs between precision and recall. When choosing a classification threshold of 0.9 (i.e., opting for high precision) for the DCNN algorithm, strong correlations (r > 0.8) were observed between daily living walking time estimates based on the lower-back (reference) sensor and the wrist sensor. Further, gait quality measures were significantly different in OAs and PD compared to healthy adults. These results demonstrate that daily living gait can be quantified using a wrist-worn sensor.

While working on her Ph.D. in computer science, Maya recently graduated from her under grad studies in Mathematics.