Publications

Global fMRI signal topography differs systematically across the lifespan
Jason S. Nomi
Jingwei Li
Taylor Bolt
Catie Chang
Salome Kornfeld
Zachary T. Goodman
B.T. Thomas Yeo
R. Nathan Spreng
Lucina Q. Uddin
H4rm0ny: A Competitive Zero-Sum Two-Player Markov Game for Multi-Agent Learning on Evasive Malware Generation and Detection
Christopher Molloy
Steven H. H. Ding
Philippe Charland
To combat the increasingly versatile and mutable modern malware, Machine Learning (ML) is now a popular and effective complement to the exis… (see more)ting signature-based techniques for malware triage and identification. However, ML is also a readily available tool for adversaries. Recent studies have shown that malware can be modified by deep Reinforcement Learning (RL) techniques to bypass AI-based and signature-based anti-virus systems without altering their original malicious functionalities. These studies only focus on generating evasive samples and assume a static detection system as the enemy.Malware detection and evasion essentially form a two-party cat-and-mouse game. Simulating the real-life scenarios, in this paper we present the first two-player competitive game for evasive malware detection and generation, following the zero-sum Multi-Agent Reinforcement Learning (MARL) paradigm. Our experiments on recent malware show that the produced malware detection agent is more robust against adversarial attacks. Also, the produced malware modification agent is able to generate more evasive samples fooling both AI-based and other anti-malware techniques.
Implications of Topological Imbalance for Representation Learning on Biomedical Knowledge Graphs
Stephen Bonner
Ufuk Kirik
Ola Engkvist
Ian P Barrett
Adoption of recently developed methods from machine learning has given rise to creation of drug-discovery knowledge graphs (KGs) that utiliz… (see more)e the interconnected nature of the domain. Graph-based modelling of the data, combined with KG embedding (KGE) methods, are promising as they provide a more intuitive representation and are suitable for inference tasks such as predicting missing links. One common application is to produce ranked lists of genes for a given disease, where the rank is based on the perceived likelihood of association between the gene and the disease. It is thus critical that these predictions are not only pertinent but also biologically meaningful. However, KGs can be biased either directly due to the underlying data sources that are integrated or due to modelling choices in the construction of the graph, one consequence of which is that certain entities can get topologically overrepresented. We demonstrate the effect of these inherent structural imbalances, resulting in densely connected entities being highly ranked no matter the context. We provide support for this observation across different datasets, models as well as predictive tasks. Further, we present various graph perturbation experiments which yield more support to the observation that KGE models can be more influenced by the frequency of entities rather than any biological information encoded within the relations. Our results highlight the importance of data modelling choices, and emphasizes the need for practitioners to be mindful of these issues when interpreting model outputs and during KG composition.
Revisiting Hotels-50K and Hotel-ID
Aarash Feizi
Arantxa Casanova
In this paper, we propose revisited versions for two recent hotel recognition datasets: Hotels-50K and Hotel-ID. The revisited versions prov… (see more)ide evaluation setups with different levels of difficulty to better align with the intended real-world application, i.e. countering human trafficking. Real-world scenarios involve hotels and locations that are not captured in the current data sets, therefore it is important to consider evaluation settings where classes are truly unseen. We test this setup using multiple state-of-the-art image retrieval models and show that as expected, the models’ performances decrease as the evaluation gets closer to the real-world unseen settings. The rankings of the best performing models also change across the different evaluation settings, which further motivates using the proposed revisited datasets.
Characterizing User Behaviors in Open-Source Software User Forums: An Empirical Study
Jazlyn Hellman
Jiahao Chen
Md. Sami Uddin
Jinghui Cheng
User forums of Open Source Software (OSS) enable end-users to collaboratively discuss problems concerning the OSS applications. Despite deca… (see more)des of research on OSS, we know very little about how end-users engage with OSS communities on these forums, in particular, the challenges that hinder their continuous and meaningful participation in the OSS community. Many previous works are developer-centric and overlook the importance of end-user forums. As a result, end-users' expectations are seldom reflected in OSS development. To better understand user behaviors in OSS user forums, we carried out an empirical study analyzing about 1.3 million posts from user forums of four popular OSS applications: Zotero, Audacity, VLC, and RStudio. Through analyzing the contribution patterns of three common user types (end-users, developers, and organizers), we observed that end-users not only initiated most of the threads (above 96% of threads in three projects, 86% in the other), but also acted as the significant contributors for responding to other users' posts, even though they tended to lack confidence in their activities as indicated by psycho-linguistic analyses. Moreover, we found end-users more open, reflecting a more positive emotion in communication than organizers and developers in the forums. Our work contributes new knowledge about end-users' activities and behaviors in OSS user forums that the vital OSS stakeholders can leverage to improve end-user engagement in the OSS development process.
Near-Optimal Glimpse Sequences for Improved Hard Attention Neural Network Training
William Harvey
Michael Teng
Hard visual attention is a promising approach to reduce the computational burden of modern computer vision methodologies. However, hard atte… (see more)ntion mechanisms can be difficult and slow to train, which is especially costly for applications like neural architecture search where multiple networks must be trained. We introduce a method to amortise the cost of training by generating an extra supervision signal for a subset of the training data. This supervision is in the form of sequences of ‘good’ locations to attend to for each image. We find that the best method to generate supervision sequences comes from framing hard attention for image classification as a Bayesian optimal experimental design (BOED) problem. From this perspective, the optimal locations to attend to are those which provide the greatest expected reduction in the entropy of the classification distribution. We introduce methodology from the BOED literature to approximate this optimal behaviour and generate ‘near-optimal’ supervision sequences. We then present a hard attention network training objective that makes use of these sequences and show that it allows faster training than prior work. We finally demonstrate the utility of faster hard attention training by incorporating supervision sequences in a neural architecture search, resulting in hard attention architectures which can outperform networks with access to the entire image.
On the Effectiveness of Interpretable Feedforward Neural Network
Miles Q. Li
Adel Abusitta
Deep learning models have achieved state-of-the-art performance in many classification tasks. However, most of them cannot provide an explan… (see more)ation for their classification results. Machine learning models that are interpretable are usually linear or piecewise linear and yield inferior performance. Non-linear models achieve much better classification performance, but it is usually hard to explain their classification results. As a counter-example, an interpretable feedforward neural network (IFFNN) is proposed to achieve both high classification performance and interpretability for malware detection. If the IFFNN can perform well in a more flexible and general form for other classification tasks while providing meaningful explanations, it may be of great interest to the applied machine learning community. In this paper, we propose a way to generalize the interpretable feedforward neural network to multi-class classification scenarios and any type of feedforward neural networks, and evaluate its classification performance and interpretability on interpretable datasets. We conclude by finding that the generalized IFFNNs achieve comparable classification performance to their normal feedforward neural network counterparts and provide meaningful explanations. Thus, this kind of neural network architecture has great practical use.
Generative Models of Brain Dynamics
Mahta Ramezanian-Panahi
Germán Abrevaya
Jean-Christophe Gagnon-Audet
Vikram Voleti
Predicting Adverse Radiation Effects in Brain Tumors After Stereotactic Radiotherapy With Deep Learning and Handcrafted Radiomics
Simon A. Keek
Manon Beuque
Sergey Primakov
Henry C. Woodruff
Avishek Chatterjee
Janita E. van Timmeren
Lizza E. L. Hendriks
Johannes Kraft
Nicolaus Andratschke
Steve E. Braunstein
Olivier Morin
Philippe Lambin
Introduction There is a cumulative risk of 20–40% of developing brain metastases (BM) in solid cancers. Stereotactic radiotherapy (SRT) en… (see more)ables the application of high focal doses of radiation to a volume and is often used for BM treatment. However, SRT can cause adverse radiation effects (ARE), such as radiation necrosis, which sometimes cause irreversible damage to the brain. It is therefore of clinical interest to identify patients at a high risk of developing ARE. We hypothesized that models trained with radiomics features, deep learning (DL) features, and patient characteristics or their combination can predict ARE risk in patients with BM before SRT. Methods Gadolinium-enhanced T1-weighted MRIs and characteristics from patients treated with SRT for BM were collected for a training and testing cohort (N = 1,404) and a validation cohort (N = 237) from a separate institute. From each lesion in the training set, radiomics features were extracted and used to train an extreme gradient boosting (XGBoost) model. A DL model was trained on the same cohort to make a separate prediction and to extract the last layer of features. Different models using XGBoost were built using only radiomics features, DL features, and patient characteristics or a combination of them. Evaluation was performed using the area under the curve (AUC) of the receiver operating characteristic curve on the external dataset. Predictions for individual lesions and per patient developing ARE were investigated. Results The best-performing XGBoost model on a lesion level was trained on a combination of radiomics features and DL features (AUC of 0.71 and recall of 0.80). On a patient level, a combination of radiomics features, DL features, and patient characteristics obtained the best performance (AUC of 0.72 and recall of 0.84). The DL model achieved an AUC of 0.64 and recall of 0.85 per lesion and an AUC of 0.70 and recall of 0.60 per patient. Conclusion Machine learning models built on radiomics features and DL features extracted from BM combined with patient characteristics show potential to predict ARE at the patient and lesion levels. These models could be used in clinical decision making, informing patients on their risk of ARE and allowing physicians to opt for different therapies.
Revisiting Transfer Functions: Learning About a Lagged Exposure-Outcome Association in Time-Series Data
Hiroshi Mamiya
Alexandra M. Schmidt
Erica E. M. Moodie
Challenging Common Assumptions about Catastrophic Forgetting
Timothee LESORT
Oleksiy Ostapenko
Pau Rodriguez
Md Rifat Arefin
Diganta Misra
Building learning agents that can progressively learn and accumulate knowledge is the core goal of the continual learning (CL) research fiel… (see more)d. Unfortunately, training a model on new data usually compromises the performance on past data. In the CL literature, this effect is referred to as catastrophic forgetting (CF). CF has been largely studied, and a plethora of methods have been proposed to address it on short sequences of non-overlapping tasks. In such setups, CF always leads to a quick and significant drop in performance in past tasks. Nevertheless, despite CF, recent work showed that SGD training on linear models accumulates knowledge in a CL regression setup. This phenomenon becomes especially visible when tasks reoccur. We might then wonder if DNNs trained with SGD or any standard gradient-based optimization accumulate knowledge in such a way. Such phenomena would have interesting consequences for applying DNNs to real continual scenarios. Indeed, standard gradient-based optimization methods are significantly less computationally expensive than existing CL algorithms. In this paper, we study the progressive knowledge accumulation (KA) in DNNs trained with gradient-based algorithms in long sequences of tasks with data re-occurrence. We propose a new framework, SCoLe (Scaling Continual Learning), to investigate KA and discover that catastrophic forgetting has a limited effect on DNNs trained with SGD. When trained on long sequences with data sparsely re-occurring, the overall accuracy improves, which might be counter-intuitive given the CF phenomenon. We empirically investigate KA in DNNs under various data occurrence frequencies and propose simple and scalable strategies to increase knowledge accumulation in DNNs.
Leveraging Structure Between Environments: Phylogenetic Regularization Incentivizes Disentangled Representations
Elliot Layne
Jason Hartford
Sébastien Lachapelle
Recently, learning invariant predictors across varying environments has been shown to improve the generalization of supervised learning meth… (see more)ods. This line of investigation holds great potential for application to biological problem settings, where data is often naturally heterogeneous. Biological samples often originate from different distributions, or environments. However, in biological contexts, the standard "invariant prediction" setting may not completely fit: the optimal predictor may in fact vary across biological environments. There also exists strong domain knowledge about the relationships between environments, such as the evolutionary history of a set of species, or the differentiation process of cell types. Most work on generic invariant predictors have not assumed the existence of structured relationships between environments. However, this prior knowledge about environments themselves has already been shown to improve prediction through a particular form of regularization applied when learning a set of predictors. In this work, we empirically evaluate whether a regularization strategy that exploits environment-based prior information can be used to learn representations that better disentangle causal factors that generate observed data. We find evidence that these methods do in fact improve the disentanglement of latent embeddings. We also show a setting where these methods can leverage phylogenetic information to estimate the number of latent causal features.