A5027310054- Reinforcement Learning in Robotics
- Adversarial Robustness in Machine Learning
- Domain Adaptation and Few-Shot Learning
- Machine Learning and Algorithms
- Neural Networks and Applications
- Computational Drug Discovery Methods
- Machine Learning in Materials Science
- Topic Modeling
- Stochastic Gradient Optimization Techniques
- Generative Adversarial Networks and Image Synthesis
- Chemical Reactions and Isotopes
- Explainable Artificial Intelligence (XAI)
- Receptor Mechanisms and Signaling
- Machine Learning and Data Classification
- Data Stream Mining Techniques
- vaccines and immunoinformatics approaches
- RNA and protein synthesis mechanisms
- Bacillus and Francisella bacterial research
- Process Optimization and Integration
- Human Pose and Action Recognition
- Cell Image Analysis Techniques
- Statistics Education and Methodologies
- Bioinformatics and Genomic Networks
- Text Readability and Simplification
- Advanced Graph Neural Networks
Mila - Quebec Artificial Intelligence Institute
2023
McGill University
2015-2021
Université de Montréal
2013
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 examine the role of memorization in deep learning, drawing connections to capacity, generalization, and adversarial robustness. While networks are capable memorizing noise data, our results suggest that they tend prioritize learning simple patterns first. In experiments, we expose qualitative differences gradient-based optimization neural (DNNs) on vs. real data. also demonstrate for appropriately tuned explicit regularization (e.g., dropout) can degrade DNN training performance datasets...
Deep learning has become the state-of-art tool in many applications, but evaluation and training of deep models can be time-consuming computationally expensive. The conditional computation approach been proposed to tackle this problem (Bengio et al., 2013; Davis & Arel, 2013). It operates by selectively activating only parts network at a time. In paper, we use reinforcement as optimize policies. More specifically, cast activation-dependent policies for dropping out blocks units problem. We...
This paper is about the problem of learning a stochastic policy for generating an object (like molecular graph) from sequence actions, such that probability proportional to given positive reward object. Whereas standard return maximization tends converge single return-maximizing sequence, there are cases where we would like sample diverse set high-return solutions. These arise, example, in black-box function optimization when few rounds possible, each with large batches queries, should be...
It has been postulated that a good representation is one disentangles the underlying explanatory factors of variation. However, it remains an open question what kind training framework could potentially achieve that. Whereas most previous work focuses on static setting (e.g., with images), we postulate some causal be discovered if learner allowed to interact its environment. The agent can experiment different actions and observe their effects. More specifically, hypothesize these correspond...
It has been postulated that a good representation is one disentangles the underlying explanatory factors of variation. However, it remains an open question what kind training framework could potentially achieve that. Whereas most previous work focuses on static setting (e.g., with images), we postulate some causal be discovered if learner allowed to interact its environment. The agent can experiment different actions and observe their effects. More specifically, hypothesize these correspond...
Finding features that disentangle the different causes of variation in real data is a difficult task, has nonetheless received considerable attention static domains like natural images. Interactive environments, which an agent can deliberately take actions, offer opportunity to tackle this task better, because experiment with actions and observe their effects. We introduce idea interactive latent factors control observed be identified by figuring out what control. propose naive method find...
Design of de novo biological sequences with desired properties, like protein and DNA sequences, often involves an active loop several rounds molecule ideation expensive wet-lab evaluations. These experiments can consist multiple stages, increasing levels precision cost evaluation, where candidates are filtered. This makes the diversity proposed a key consideration in phase. In this work, we propose learning algorithm leveraging epistemic uncertainty estimation recently GFlowNets as generator...
The COVID-19 pandemic continues to pose a substantial threat human lives and is likely do so for years come. Despite the availability of vaccines, searching efficient small-molecule drugs that are widely available, including in low- middle-income countries, an ongoing challenge. In this work, we report results open science community effort, "Billion molecules against challenge", identify inhibitors SARS-CoV-2 or relevant receptors. Participating teams used wide variety computational methods...
Humans interpret texts with respect to some background information, or world knowledge, and we would like develop automatic reading comprehension systems that can do the same. In this paper, introduce a task several models drive progress towards goal. particular, propose of rare entity prediction: given web document entities removed, are tasked predicting correct missing conditioned on context lexical resources. This is challenging due diversity language styles extremely large number...
Generative Flow Networks (GFlowNets) have been introduced as a method to sample diverse set of candidates in an active learning context, with training objective that makes them approximately proportion given reward function. In this paper, we show number additional theoretical properties GFlowNets. They can be used estimate joint probability distributions and the corresponding marginal where some variables are unspecified and, particular interest, represent over composite objects like sets...
We study the link between generalization and interference in temporal-difference (TD) learning. Interference is defined as inner product of two different gradients, representing their alignment. This quantity emerges being interest from a variety observations about neural networks, parameter sharing dynamics find that TD easily leads to low-interference, under-generalizing parameters, while effect seems reversed supervised hypothesize cause can be traced back interplay bootstrapping....
Generative flow networks (GFlowNets) are a method for learning stochastic policy generating compositional objects, such as graphs or strings, from given unnormalized density by sequences of actions, where many possible action may lead to the same object. We find previously proposed objectives GFlowNets, matching and detailed balance, which analogous temporal difference learning, be prone inefficient credit propagation across long sequences. thus propose new objective trajectory more...
Generative flow networks (GFlowNets) are a family of algorithms for training sequential sampler discrete objects under an unnormalized target density and have been successfully used various probabilistic modeling tasks. Existing objectives GFlowNets either local to states or transitions, propagate reward signal over entire sampling trajectory. We argue that these alternatives represent opposite ends gradient bias-variance tradeoff propose way exploit this mitigate its harmful effects....
Machine-learning algorithms can be fooled by small well-designed adversarial perturbations. This is reminiscent of cellular decision-making where ligands (called antagonists) prevent correct signaling, like in early immune recognition. We draw a formal analogy between neural networks used machine learning and models (adaptive proofreading). apply attacks from to simple show explicitly the correspondence antagonism weakly bound ligands. Such absent more nonlinear models, which inspires us...
We study the problem of generating diverse candidates in context Multi-Objective Optimization. In many applications machine learning such as drug discovery and material design, goal is to generate which simultaneously optimize a set potentially conflicting objectives. Moreover, these objectives are often imperfect evaluations some underlying property interest, making it important have multiple options for expensive downstream evaluations. propose GFlowNets (MOGFNs), novel method Pareto...
Generative Flow Networks (GFlowNets; GFNs) are a family of reward/energy-based generative methods for combinatorial objects, capable generating diverse and high-utility samples. However, biasing GFNs towards producing samples is non-trivial. In this work, we leverage connections between reinforcement learning (RL) propose to combine the GFN policy with an action-value estimate, $Q$, create greedier sampling policies which can be controlled by mixing parameter. We show that several variants...
Generative Flow Networks (GFlowNets, GFNs) are a generative framework for learning unnormalized probability mass functions over discrete spaces. Since their inception, GFlowNets have proven to be useful models in applications where the majority of space is unvisited during training. This has inspired some hypothesize that GFlowNets, when paired with deep neural networks (DNNs), favourable generalization properties. In this work, we empirically verify hypothesized mechanisms GFlowNets....
Despite substantial progress in machine learning for scientific discovery recent years, truly de novo design of small molecules which exhibit a property interest remains significant challenge. We introduce LambdaZero, generative active approach to search synthesizable molecules. Powered by deep reinforcement learning, LambdaZero learns over the vast space discover candidates with desired property. apply molecular docking novel that inhibit enzyme soluble Epoxide Hydrolase 2 (sEH), while...
Diffusion models have emerged as effective distribution estimators in vision, language, and reinforcement learning, but their use priors downstream tasks poses an intractable posterior inference problem. This paper studies amortized sampling of the over data, $\mathbf{x}\sim p^{\rm post}(\mathbf{x})\propto p(\mathbf{x})r(\mathbf{x})$, a model that consists diffusion generative prior $p(\mathbf{x})$ black-box constraint or likelihood function $r(\mathbf{x})$. We state prove asymptotic...
The Generative Flow Network (GFlowNet) is a probabilistic framework in which an agent learns stochastic policy and flow functions to sample objects with probability proportional unnormalized reward function. GFlowNets share strong resemblance reinforcement learning (RL), that typically aims maximize reward, due their sequential decision-making processes. Recent works have studied connections between maximum entropy (MaxEnt) RL, modifies the standard objective of RL agents by...