Publications

Adult neurogenesis acts as a neural regularizer
Lina M. Tran
Adam Santoro
Lulu Liu
Sheena A. Josselyn
Blake A. Richards
Paul W. Frankland
New neurons are continuously generated in the subgranular zone of the dentate gyrus throughout adulthood. These new neurons gradually integr… (voir plus)ate into hippocampal circuits, forming new naive synapses. Viewed from this perspective, these new neurons may represent a significant source of “wiring” noise in hippocampal networks. In machine learning, such noise injection is commonly used as a regularization technique. Regularization techniques help prevent overfitting training data and allow models to generalize learning to new, unseen data. Using a computational modeling approach, here we ask whether a neurogenesis-like process similarly acts as a regularizer, facilitating generalization in a category learning task. In a convolutional neural network (CNN) trained on the CIFAR-10 object recognition dataset, we modeled neurogenesis as a replacement/turnover mechanism, where weights for a randomly chosen small subset of hidden layer neurons were reinitialized to new values as the model learned to categorize 10 different classes of objects. We found that neurogenesis enhanced generalization on unseen test data compared to networks with no neurogenesis. Moreover, neurogenic networks either outperformed or performed similarly to networks with conventional noise injection (i.e., dropout, weight decay, and neural noise). These results suggest that neurogenesis can enhance generalization in hippocampal learning through noise injection, expanding on the roles that neurogenesis may have in cognition.
2022-11-01
Proceedings of the National Academy of Sciences of the United States of America (publié)
doi.org
Automatic measure and normalization of spinal cord cross-sectional area using the pontomedullary junction
Sandrine Bédard
Julien Cohen‐Adad
Spinal cord cross-sectional area (CSA) is a relevant biomarker to assess spinal cord atrophy in neurodegenerative diseases. However, the con… (voir plus)siderable inter-subject variability among healthy participants currently limits its usage. Previous studies explored factors contributing to the variability, yet the normalization models required manual intervention and used vertebral levels as a reference, which is an imprecise prediction of the spinal levels. In this study we implemented a method to measure CSA automatically from a spatial reference based on the central nervous system (the pontomedullary junction, PMJ), we investigated factors to explain variability, and developed normalization strategies on a large cohort (N = 804). Following automatic spinal cord segmentation, vertebral labeling and PMJ labeling, the spinal cord CSA was computed on T1w MRI scans from the UK Biobank database. The CSA was computed using two methods. For the first method, the CSA was computed at the level of the C2–C3 intervertebral disc. For the second method, the CSA was computed at 64 mm caudally from the PMJ, this distance corresponding to the average distance between the PMJ and the C2–C3 disc across all participants. The effect of various demographic and anatomical factors was explored, and a stepwise regression found significant predictors; the coefficients of the best fit model were used to normalize CSA. CSA measured at C2–C3 disc and using the PMJ differed significantly (paired t-test, p-value = 0.0002). The best normalization model included thalamus, brain volume, sex and the interaction between brain volume and sex. The coefficient of variation went down for PMJ CSA from 10.09 (without normalization) to 8.59%, a reduction of 14.85%. For CSA at C2–C3, it went down from 9.96 to 8.42%, a reduction of 15.13 %. This study introduces an end-to-end automatic pipeline to measure and normalize cord CSA from a neurological reference. This approach requires further validation to assess atrophy in longitudinal studies. The inter-subject variability of CSA can be partly accounted for by demographics and anatomical factors.
2022-11-01
Frontiers in Neuroimaging (publié)
doi.org
BERTPerf: Inference Latency Predictor for BERT on ARM big.LITTLE Multi-Core Processors
M. Abdelgawad
S. H. Mozafari
J. J. Clark
B. H. Meyer
W. J. Gross
Hardware-aware Neural Architecture Search (NAS) and mapping & scheduling optimization methods are being used to find efficient implementatio… (voir plus)ns of computationally-intense language models such as BERT. This requires measuring real hardware inference latency: good design decisions simply cannot be made with proxy metrics such as FLOPs or the number of parameters. However, the time required to perform on-device latency measurements is prohibitive (e.g., a few days to a few weeks over the course of an optimization run). To address this, we present BERTPerf, a low-cost, highly-accurate method to predict the inference time of BERT on ARM big.LITTLE multi-core processors. BERTPerf exploits latency patterns at the layer-level to reduce on-device latency measurements, and captures the effect of caching and intermediate tensor allocations to reduce latency prediction error. BERTPerf reduces the maximum prediction error by 7–11% compared to the state-of-the-art, and requires 75% less on-device measurements compared to existing work at the same prediction error.
2022-11-01
IEEE Workshop on Signal Processing Systems (publié)
doi.org
A General-Purpose Neural Architecture for Geospatial Systems
Nasim Rahaman
Martin Weiss
Frederik Träuble
Francesco Locatello
Alexandre Lacoste
Yoshua Bengio
Chris Pal
Li Erran Li
Bernhard Schölkopf
2022-11-01
OpenReview (inconnu)
doi.org
openreview.net
Active Keyword Selection to Track Evolving Topics on Twitter
Sacha Lévy
Farimah Poursafaei
Kellin Pelrine
Reihaneh Rabbany
How can we study social interactions on evolving topics at a mass scale? Over the past decade, researchers from diverse fields such as econo… (voir plus)mics, political science, and public health have often done this by querying Twitter's public API endpoints with hand-picked topical keywords to search or stream discussions. However, despite the API's accessibility, it remains difficult to select and update keywords to collect high-quality data relevant to topics of interest. In this paper, we propose an active learning method for rapidly refining query keywords to increase both the yielded topic relevance and dataset size. We leverage a large open-source COVID-19 Twitter dataset to illustrate the applicability of our method in tracking Tweets around the key sub-topics of Vaccine, Mask, and Lockdown. Our experiments show that our method achieves an average topic-related keyword recall 2x higher than baselines. We open-source our code along with a web interface for keyword selection to make data collection from Twitter more systematic for researchers.
2022-10-31
2022 IEEE International Conference on Data Mining Workshops (ICDMW) (publié)
doi.org
arxiv.org
Consistent Training via Energy-Based GFlowNets for Modeling Discrete Joint Distributions
Chanakya Ajit Ekbote
Moksh J. Jain
Payel Das
Yoshua Bengio
Generative Flow Networks (GFlowNets) have demonstrated significant performance improvements for generating diverse discrete objects …
2022-10-31
arXiv (prépublication)
doi.org
arxiv.org
Development and Validation of Multiparametric MRI-based Radiomics Models for Preoperative Risk Stratification of Endometrial Cancer
Thierry L. Lefebvre
Yoshiko Ueno
Anthony Dohan
Avishek Chatterjee
Martin Vallières
Eric Winter-Reinhold
Sameh Saif
Ives R. Levesque
Xing Ziggy Zeng
Reza Forghani
Jan Seuntjens
Philippe Soyer
Peter Savadjiev
Caroline Reinhold
Background Stratifying high-risk histopathologic features in endometrial carcinoma is important for treatment planning. Radiomics analysis a… (voir plus)t preoperative MRI holds potential to identify high-risk phenotypes. Purpose To evaluate the performance of multiparametric MRI three-dimensional radiomics-based machine learning models for differentiating low- from high-risk histopathologic markers-deep myometrial invasion (MI), lymphovascular space invasion (LVSI), and high-grade status-and advanced-stage endometrial carcinoma. Materials and Methods This dual-center retrospective study included women with histologically proven endometrial carcinoma who underwent 1.5-T MRI before hysterectomy between January 2011 and July 2015. Exclusion criteria were tumor diameter less than 1 cm, missing MRI sequences or histopathology reports, neoadjuvant therapy, and malignant neoplasms other than endometrial carcinoma. Three-dimensional radiomics features were extracted after tumor segmentation at MRI (T2-weighted, diffusion-weighted, and dynamic contrast-enhanced MRI). Predictive features were selected in the training set with use of random forest (RF) models for each end point, and trained RF models were applied to the external test set. Five board-certified radiologists conducted MRI-based staging and deep MI assessment in the training set. Areas under the receiver operating characteristic curve (AUCs) were reported with balanced accuracies, and radiologists' readings were compared with radiomics with use of McNemar tests. Results In total, 157 women were included: 94 at the first institution (training set; mean age, 66 years ± 11 [SD]) and 63 at the second institution (test set; 67 years ± 12). RF models dichotomizing deep MI, LVSI, high grade, and International Federation of Gynecology and Obstetrics (FIGO) stage led to AUCs of 0.81 (95% CI: 0.68, 0.88), 0.80 (95% CI: 0.67, 0.93), 0.74 (95% CI: 0.61, 0.86), and 0.84 (95% CI: 0.72, 0.92), respectively, in the test set. In the training set, radiomics provided increased performance compared with radiologists' readings for identifying deep MI (balanced accuracy, 86% vs 79%; P = .03), while no evidence of a difference was observed in performance for advanced FIGO stage (80% vs 78%; P = .27). Conclusion Three-dimensional radiomics can stratify patients by using preoperative MRI according to high-risk histopathologic end points in endometrial carcinoma and provide nonsignificantly different or higher performance than radiologists in identifying advanced stage and deep myometrial invasion, respectively. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Kido and Nishio in this issue.
2022-10-31
Radiology (publié)
doi.org
Discrete Factorial Representations as an Abstraction for Goal Conditioned Reinforcement Learning
Riashat Islam
Hongyu Zang
Anirudh Goyal
Alex Lamb
Kenji Kawaguchi
Xin Li
Romain Laroche
Yoshua Bengio
Remi Tachet des Combes
2022-10-31
arXiv (prépublication)
doi.org
arxiv.org
Successive-Cancellation Decoding of Reed-Muller Codes With Fast Hadamard Transform
Nghia Doan
Seyyed Ali Hashemi
Warren J. Gross
A novel permuted fast successive-cancellation list decoding algorithm with fast Hadamard transform (FHT-FSCL) is presented. The proposed dec… (voir plus)oder initializes
2022-10-31
IEEE Transactions on Vehicular Technology (publié)
doi.org
arxiv.org
Type B Ultra Long-Range Interactions in PFAS (TULIPs) Are Recurrent Epigenomic Features of PFA Ependymoma
Michael Johnston
John JY Lee
Bo Hu
Ana Nikolic
Audrey Baguette
Seungil Paik
Haifen Chen
Sachin Kumar
Carol Chen
Selin Jessa
Polina Balin
Vernon Fong
Melissa Zwaig
Kulandaimanuvel MichealRaj
Xun Chen
Yanlin Zhang
Srinidhi Varadharajan
Pierre Billon
Nikoleta Juretic
Craig Daniels … (voir 21 de plus)
Caterina Giannini
Eric Thompson
Peter Hauser
Seung-Ki Kim
Kyu-Chang Wang
Ji Yeoun Lee
Wieslawa Grajkowska
Sameer Agnihotri
Stephen C. Mack
Benjamin Ellezam
Alex Weil
Guillaume Bourque
Jennifer Chan
Mathieu Lupien
Jiannis Ragoussis
Claudia Kleinman
Jacek Majewski
Nada Jabado
Michael Taylor
Mathieu Blanchette
Marco Gallo
Posterior Fossa Group A (PFA) ependymomas are pediatric brain tumors with extremely poor survival outcomes. As protein-coding mutations in P… (voir plus)FA are exceedingly rare, the underlying etiology of these tumors remains elusive. Elevated CpG island methylation and depletion of H3K27me3 have been described in PFA, leading to the hypothesis that PFA may be driven by a dysregulated epigenetic state. In this study, we sought to determine how three-dimensional (3D) genome features (such as DNA loops, domains, and compartments) differ between pediatric brain tumors. We performed Hi-C sequencing on a collection of 64 patient specimens and patient-derived primary cultures that collectively span multiple subgroups of ependymoma, medulloblastoma, high-grade glioma, and non-neoplastic brain. For certain samples, we further performed RNA-seq, histone modification ChIP-seq, or whole-genome bisulfite sequencing to allow multiomic data integration. Overall, the 3D genome organization of PFA samples appeared distinct from other tumor types. We identified and defined TULIPs: a subset of type B compartments, separated by genomic distances greater than 10 Mbp, that exhibit a striking fivefold increase in reciprocal interaction strength. These TULIPs recurred at the same genomic positions across the vast majority of PFA samples with minimal representation among other tumor or non-tumor samples. TULIPs displayed enrichment for heterochromatic features such as H3K9me3 and late replication timing and were depleted of euchromatic features such as H3K27ac and protein-coding genes. By using immuno-fluorescence for H3K9me3 and oligo-FISH to label TULIP regions, we demonstrated that TULIP regions are more compact in PFA than other tumors. Finally, by applying inhibitors of H3K9 lysine methylation to PFA cultures we showed that TULIPs become more diffuse and cell viability is reduced. Altogether, this work defines TULIPs as highly recurrent epigenetic features of PFA tumors.
2022-10-31
Neuro-Oncology (publié)
doi.org
Clipped Stochastic Methods for Variational Inequalities with Heavy-Tailed Noise
Eduard Gorbunov
Marina Danilova
David Dobre
Pavel Dvurechensky
Alexander Gasnikov
Gauthier Gidel
2022-10-30
NeurIPS.cc/2022/Conference (accepté)
doi.org
openreview.net
Flexible Diffusion Modeling of Long Videos
William Harvey
Saeid Naderiparizi
Vaden Masrani
Christian Dietrich Weilbach
Frank N. Wood
We present a framework for video modeling based on denoising diffusion probabilistic models that produces long-duration video completions in… (voir plus) a variety of realistic environments. We introduce a generative model that can at test-time sample any arbitrary subset of video frames conditioned on any other subset and present an architecture adapted for this purpose. Doing so allows us to efficiently compare and optimize a variety of schedules for the order in which frames in a long video are sampled and use selective sparse and long-range conditioning on previously sampled frames. We demonstrate improved video modeling over prior work on a number of datasets and sample temporally coherent videos over 25 minutes in length. We additionally release a new video modeling dataset and semantically meaningful metrics based on videos generated in the CARLA autonomous driving simulator.
2022-10-30
NeurIPS.cc/2022/Conference (accepté)
doi.org
openreview.net