A5089648583- Reinforcement Learning in Robotics
- Generative Adversarial Networks and Image Synthesis
- Model Reduction and Neural Networks
- Neural Networks and Applications
- Online and Blended Learning
- Domain Adaptation and Few-Shot Learning
- Advanced Neural Network Applications
- Machine Learning and Algorithms
- Online Learning and Analytics
- Music and Audio Processing
- Job Satisfaction and Organizational Behavior
- Machine Learning and Data Classification
- French Language Learning Methods
- Evolutionary Algorithms and Applications
- Aging, Elder Care, and Social Issues
- Educational Tools and Methods
- Innovative Teaching and Learning Methods
- Gaussian Processes and Bayesian Inference
- Multimodal Machine Learning Applications
- Neural Networks and Reservoir Computing
- Computational Physics and Python Applications
- Advanced Bandit Algorithms Research
- Management and Organizational Studies
- Digital Media Forensic Detection
- Information Technology and Learning
Université du Québec en Outaouais
2021-2024
Université Laval
2014-2024
Université TÉLUQ
2019-2022
Google (United States)
2014-2021
DeepMind (United Kingdom)
2014-2018
Google (United Kingdom)
2015-2018
Massachusetts Institute of Technology
2017
Australian National University
2014-2015
Université de Montréal
2009-2014
Département d'Informatique
2013
The ability to learn tasks in a sequential fashion is crucial the development of artificial intelligence. Until now neural networks have not been capable this and it has widely thought that catastrophic forgetting an inevitable feature connectionist models. We show possible overcome limitation train can maintain expertise on they experienced for long time. Our approach remembers old by selectively slowing down learning weights important those tasks. demonstrate our scalable effective solving...
Theano is a Python library that allows to define, optimize, and evaluate mathematical expressions involving multi-dimensional arrays efficiently. Since its introduction, it has been one of the most used CPU GPU compilers - especially in machine learning community shown steady performance improvements. being actively continuously developed since 2008, multiple frameworks have built on top produce many state-of-the-art models. The present article structured as follows. Section I provides an...
Learning to solve complex sequences of tasks--while both leveraging transfer and avoiding catastrophic forgetting--remains a key obstacle achieving human-level intelligence. The progressive networks approach represents step forward in this direction: they are immune forgetting can leverage prior knowledge via lateral connections previously learned features. We evaluate architecture extensively on wide variety reinforcement learning tasks (Atari 3D maze games), show that it outperforms common...
Theano is a compiler for mathematical expressions in Python that combines the convenience of NumPy's syntax with speed optimized native machine language. The user composes high-level description mimics and semantics, while being statically typed functional (as opposed to imperative). These allow provide symbolic differentiation. Before performing computation, optimizes choice expressions, translates them into C++ (or CUDA GPU), compiles dynamically loaded modules, all automatically. Common...
In this paper we present the techniques used for University of Montréal's team submissions to 2013 Emotion Recognition in Wild Challenge. The challenge is classify emotions expressed by primary human subject short video clips extracted from feature length movies. This involves analysis acted scenes lasting approximately one-two seconds, including audio track which may contain voices as well background music. Our approach combines multiple deep neural networks different data modalities,...
We present new intuitions and theoretical assessments of the emergence disentangled representation in variational autoencoders. Taking a rate-distortion theory perspective, we show circumstances under which representations aligned with underlying generative factors variation data emerge when optimising modified ELBO bound $β$-VAE, as training progresses. From these insights, propose modification to regime that progressively increases information capacity latent code during training. This...
Here we propose a novel model family with the objective of learning to disentangle factors variation in data. Our approach is based on spike-and-slab restricted Boltzmann machine which generalize include higher-order interactions among multiple latent variables. Seen from generative perspective, multiplicative emulates entangling variation. Inference can be seen as disentangling these factors. Unlike previous attempts at factors, proposed trained using no supervised information regarding We...
We introduce Natural Neural Networks, a novel family of algorithms that speed up convergence by adapting their internal representation during training to improve conditioning the Fisher matrix. In particular, we show specific example employs simple and efficient reparametrization neural network weights implicitly whitening obtained at each layer, while preserving feed-forward computation network. Such networks can be trained efficiently via proposed Projected Gradient Descent algorithm...
Abstract Sustainable human resource management is gaining importance in organizations due to its role developing a sustainable work environment and well‐being. This paper discusses the relationship between employee perceptions of job satisfaction 54 countries. We propose that HRM positively associated with but this moderated by employees' identification organization country‐level individualism–collectivism. Thus, we suggest national culture functions as second‐level moderator organizational...
Policies for complex visual tasks have been successfully learned with deep reinforcement learning, using an approach called Q-networks (DQN), but relatively large (task-specific) networks and extensive training are needed to achieve good performance. In this work, we present a novel method policy distillation that can be used extract the of learning agent train new network performs at expert level while being dramatically smaller more efficient. Furthermore, same consolidate multiple...
Purpose Ichak Adizes has developed original and practical conceptions of executive interaction, change management corporate development, collectively referred to as “symbergetic organisational therapy”. Although his name is celebrated in some circles, it not widely known within mainstream academia. Further, Adizes’ insights into what organisations are how they achieve optimal performance routinely dealt with Western business schools. After exposing ideas, this paper aims investigate reasons...
The spike-and-slab restricted Boltzmann machine (ssRBM) is defined to have both a real-valued "slab" variable and binary "spike" associated with each unit in the hidden layer. model uses its slab variables conditional covariance of observation-thought be important capturing statistical properties natural images. In this paper, we present canonical ssRBM framework together some extensions. These extensions highlight flexibility RBM as platform for exploring more sophisticated probabilistic...
We introduce RecurrentGemma, an open language model which uses Google's novel Griffin architecture. combines linear recurrences with local attention to achieve excellent performance on language. It has a fixed-sized state, reduces memory use and enables efficient inference long sequences. provide pre-trained 2B non-embedding parameters, instruction tuned variant. Both models comparable Gemma-2B despite being trained fewer tokens.