A5102922377- Bayesian Modeling and Causal Inference
- Topic Modeling
- Domain Adaptation and Few-Shot Learning
- Cancer-related Molecular Pathways
- Speech Recognition and Synthesis
- Reinforcement Learning in Robotics
- Advanced Graph Neural Networks
- Explainable Artificial Intelligence (XAI)
- Advanced Breast Cancer Therapies
- Machine Learning and Algorithms
- Natural Language Processing Techniques
- Generative Adversarial Networks and Image Synthesis
- Gaussian Processes and Bayesian Inference
- Machine Learning in Healthcare
- Neural Networks and Applications
- Multimodal Machine Learning Applications
- Protein Degradation and Inhibitors
- Lung Cancer Research Studies
- Advanced Bandit Algorithms Research
- Data Stream Mining Techniques
- Machine Learning and Data Classification
- Advanced Image and Video Retrieval Techniques
- Adversarial Robustness in Machine Learning
- Cancer Genomics and Diagnostics
- Ovarian cancer diagnosis and treatment
DeepMind (United Kingdom)
2023
Northeastern University
2023
Peking University
2021-2022
Chongqing University
2022
Polytechnique Montréal
2017-2021
Université de Montréal
2015-2021
China University of Geosciences
2021
Mila - Quebec Artificial Intelligence Institute
2021
Harvard University
2011-2018
Dana-Farber Cancer Institute
2011-2018
The two fields of machine learning and graphical causality arose are developed separately. However, there is, now, cross-pollination increasing interest in both to benefit from the advances other. In this article, we review fundamental concepts causal inference relate them crucial open problems learning, including transfer generalization, thereby assaying how can contribute modern research. This also applies opposite direction: note that most work starts premise variables given. A central...
We present MILABOT: a deep reinforcement learning chatbot developed by the Montreal Institute for Learning Algorithms (MILA) Amazon Alexa Prize competition. MILABOT is capable of conversing with humans on popular small talk topics through both speech and text. The system consists an ensemble natural language generation retrieval models, including template-based bag-of-words sequence-to-sequence neural network latent variable models. By applying to crowdsourced data real-world user...
The use of dialogue systems as a medium for human-machine interaction is an increasingly prevalent paradigm. A growing number conversation strategies that are learned from large datasets. There well documented instances where interactions with these system have resulted in biased or even offensive conversations due to the data-driven training process. Here, we highlight potential ethical issues arise research, including: implicit biases systems, rise adversarial examples, sources privacy...
Abstract Recent studies suggest that targeting transcriptional machinery can lead to potent and selective anticancer effects in cancers dependent on high constant expression of certain transcription factors for growth survival. Cyclin-dependent kinase 7 (CDK7) is the catalytic subunit CDK-activating complex. Its function required both cell-cycle regulation control gene expression. CDK7 has recently emerged as an attractive cancer target because its inhibition leads decreased transcript...
We propose zoneout, a novel method for regularizing RNNs. At each timestep, zoneout stochastically forces some hidden units to maintain their previous values. Like dropout, uses random noise train pseudo-ensemble, improving generalization. But by preserving instead of dropping units, gradient information and state are more readily propagated through time, as in feedforward stochastic depth networks. perform an empirical investigation various RNN regularizers, find that gives significant...
The two fields of machine learning and graphical causality arose developed separately. However, there is now cross-pollination increasing interest in both to benefit from the advances other. In present paper, we review fundamental concepts causal inference relate them crucial open problems learning, including transfer generalization, thereby assaying how can contribute modern research. This also applies opposite direction: note that most work starts premise variables are given. A central...
Many efforts have been devoted to training generative latent variable models with autoregressive decoders, such as recurrent neural networks (RNN). Stochastic successful in capturing the variability observed natural sequential data speech. We unify ideas from recently proposed architectures into a stochastic model: each step sequence is associated that used condition dynamics for future steps. Training performed amortized variational inference where approximate posterior augmented RNN runs...
CDK7 has emerged as an exciting target in oncology due to its roles two important processes that are misregulated cancer cells: cell cycle and transcription. This report describes the discovery of
Deep Neural Network (DNN) acoustic models have yielded many state-of-the-art results in Automatic Speech Recognition (ASR) tasks. More recently, Recurrent (RNN) been shown to outperform DNNs counterparts. However, DNN and RNN tend be impractical deploy on embedded systems with limited computational capacity. Traditionally, the approach for platforms is either train a small directly, or that learns output distribution of large DNN. In this paper, we utilize transfer knowledge We use model...
E-type cyclins (cyclins E1 and E2) are components of the core cell cycle machinery overexpressed in many human tumor types. E thought to drive proliferation by activating cyclin-dependent kinase 2 (CDK2). The cyclin gene represents site recurrent integration hepatitis B virus pathogenesis hepatocellular carcinoma, this event is associated with strong up-regulation expression. Regardless underlying mechanism tumorigenesis, majority liver cancers overexpress cyclins. Here we used conditional...
Most recommender systems recommend a list of items. The user examines the list, from first item to last, and often chooses attractive does not examine rest. This type behavior can be modeled by cascade model. In this work, we study cascading bandits, an online learning variant model where goal is $K$ most items large set $L$ candidate We propose two algorithms for solving problem, which are based on idea linear generalization. key in our solutions that learn predictor attraction...
The fundamental challenge in causal induction is to infer the underlying graph structure given observational and/or interventional data. Most existing algorithms operate by generating candidate graphs and evaluating them using either score-based methods (including continuous optimization) or independence tests. In our work, we instead treat inference process as a black box design neural network architecture that learns mapping from both data structures via supervised training on synthetic...
We propose a simple technique for encouraging generative RNNs to plan ahead. train "backward" recurrent network generate given sequence in reverse order, and we encourage states of the forward model predict cotemporal backward model. The is used only during training, plays no role sampling or inference. hypothesize that our approach eases modeling long-term dependencies by implicitly forcing hold information about longer-term future (as contained states). show empirically achieves 9%...
Often, the performance on a supervised machine learning task is evaluated with emph{task loss} function that cannot be optimized directly. Examples of such loss functions include classification error, edit distance and BLEU score. A common workaround for this problem to instead optimize emph{surrogate function, as instance cross-entropy or hinge loss. In order remedy effective, it important ensure minimization surrogate results in loss, condition we call emph{consistency loss}. work, propose...
Learning long-term dependencies in extended temporal sequences requires credit assignment to events far back the past. The most common method for training recurrent neural networks, back-propagation through time (BPTT), information be propagated backwards every single step of forward computation, potentially over thousands or millions steps. This becomes computationally expensive even infeasible when used with long sequences. Importantly, biological brains are unlikely perform such detailed...