Developing agents for complex and underspecified tasks, where no clear objective exists, remains challenging but offers many opportunities. This is especially true in video games, where simulated players (bots) need to play realistically, and there is no clear reward to evaluate them. While imitation learning has shown promise in such domains, these methods often fail when agents encounter out-of-distribution scenarios during deployment. Expanding the training dataset is a common solution, but it becomes impractical or costly when relying on human demonstrations. This article addresses active imitation learning, aiming to trigger expert intervention only when necessary, reducing the need for constant expert input along training. We introduce Random Network Distillation DAgger (RND-DAgger), a new active imitation learning method that limits expert querying by using a learned state-based out-of-distribution measure to trigger interventions. This approach avoids frequent expert-agent action comparisons, thus making the expert intervene only when it is useful. We evaluate RND-DAgger against traditional imitation learning and other active approaches in 3D video games (racing and third-person navigation) and in a robotic locomotion task and show that RND-DAgger surpasses previous methods by reducing expert queries. https://sites.google.com/view/rnd-dagger

Offline Reinforcement Learning (RL) has emerged as a powerful alternative to imitation learning for behavior modeling in various domains, particularly in complex long range navigation tasks. An existing challenge with Offline RL is the signal-to-noise ratio, i.e. how to mitigate incorrect policy updates due to errors in value estimates. Towards this, multiple works have demonstrated the advantage of hierarchical offline RL methods, which decouples high-level path planning from low-level path following. In this work, we present a novel hierarchical transformer-based approach leveraging a learned quantizer of the state space to tackle long horizon navigation tasks. This quantization enables the training of a simpler zone-conditioned low-level policy and simplifies planning, which is reduced to discrete autoregressive prediction. Among other benefits, zone-level reasoning in planning enables explicit trajectory stitching rather than implicit stitching based on noisy value function estimates. By combining this transformer-based planner with recent advancements in offline RL, our proposed approach achieves state-of-the-art results in complex long-distance navigation environments.

Deep reinforcement learning (DRL) techniques have become increasingly used in various fields for decision-making processes. However, a challenge that often arises is the trade-off between both the computational efficiency of the decision-making process and the ability of the learned agent to solve a particular task. This is particularly critical in real-time settings such as video games where the agent needs to take relevant decisions at a very high frequency, with a very limited inference timeIn this work, we propose a generic offline learning approach where the computation cost of the input features is taken into account. We derive the Budgeted Decision Transformer as an extension of the Decision Transformer that incorporates cost constraints to limit its cost at inference. As a result, the model can dynamically choose the best input features at each timestep. We demonstrate the effectiveness of our method on several tasks, including D4RL benchmarks and complex 3D environments similar to those found in video games, and show that it can achieve similar performance while using significantly fewer computational resources compared to classical approaches.

Standard cooperative multi-agent reinforcement learning (MARL) methods aim to find the optimal team cooperative policy to complete a task. However there may exist multiple different ways of cooperating, which usually are very needed by domain experts. Therefore, identifying a set of significantly different policies can alleviate the task complexity for them. Unfortunately, there is a general lack of effective policy diversity approaches specifically designed for the multi-agent domain. In this work, we propose a method called Moment-Matching Policy Diversity to alleviate this problem. This method can generate different team policies to varying degrees by formalizing the difference between team policies as the difference in actions of selected agents in different policies. Theoretically, we show that our method is a simple way to implement a constrained optimization problem that regularizes the difference between two trajectory distributions by using the maximum mean discrepancy. The effectiveness of our approach is demonstrated on a challenging team-based shooter.

Existing imitation learning methods mainly focus on making an agent effectively mimic a demonstrated behavior, but do not address the potential contradiction between the behavior style and the objective of a task. There is a general lack of efficient methods that allow an agent to partially imitate a demonstrated behavior to varying degrees, while completing the main objective of a task. In this paper we propose a method called Regularized Soft Actor-Critic which formulates the main task and the imitation task under the Constrained Markov Decision Process framework (CMDP). The main task is defined as the maximum entropy objective used in Soft Actor-Critic (SAC) and the imitation task is defined as a constraint. We evaluate our method on continuous control tasks relevant to video games applications.

When users initiate search sessions, their query are often ambiguous or might lack of context; this resulting in non-efficient document ranking. Multiple approaches have been proposed by the Information Retrieval community to add context and retrieve documents aligned with users' intents. While some work focus on query disambiguation using users' browsing history, a recent line of work proposes to interact with users by asking clarification questions or/and proposing clarification panels. However, these approaches count either a limited number (i.e., 1) of interactions with user or log-based interactions. In this paper, we propose and evaluate a fully simulated query clarification framework allowing multi-turn interactions between IR systems and user agents.

In reinforcement learning, the graph Laplacian has proved to be a valuable tool in the task-agnostic setting, with applications ranging from skill discovery to reward shaping. Recently, learning the Laplacian representation has been framed as the optimization of a temporally-contrastive objective to overcome its computational limitations in large (or continuous) state spaces. However, this approach requires uniform access to all states in the state space, overlooking the exploration problem that emerges during the representation learning process. In this work, we propose an alternative method that is able to recover, in a non-uniform-prior setting, the expressiveness and the desired properties of the Laplacian representation. We do so by combining the representation learning with a skill-based covering policy, which provides a better training distribution to extend and refine the representation. We also show that a simple augmentation of the representation objective with the learned temporal abstractions improves dynamics-awareness and helps exploration. We find that our method succeeds as an alternative to the Laplacian in the non-uniform setting and scales to challenging continuous control environments. Finally, even if our method is not optimized for skill discovery, the learned skills can successfully solve difficult continuous navigation tasks with sparse rewards, where standard skill discovery approaches are no so effective.

In many applications that rely on machine learning, the availability of labelled data is a matter of primary importance. However, when tackling new tasks, labels are usually missing and must be collected from scratch by the users. In this work, we address the problem of learning classifiers when the amount of labels is very scarce. We do so by learning multiple vectors, called prototypes, that represent relevant semantic concepts for the task at hand. We propose a theoretically inspired mechanism that computes probabilities of matching between the prototypes and the input elements, and we combine these probabilities to increase the expressiveness of the classifier. Moreover, by leveraging low-cost extra annotations in the training data, a simple error-boosting technique guides the learning process and provides substantial performance improvements. Empirical results confirm the benefits of the proposed approach in both balanced and unbalanced datasets. Our methodology is thus of practical use when gathering and labelling new examples is more expensive than annotating what we already have.

Obtaining training data for Question Answering (QA) is time-consuming and resourceintensive, and existing QA datasets are only available for limited domains and languages.In this work, we explore to what extent high quality training data is actually required for Extractive QA, and investigate the possibility of unsupervised Extractive QA.We approach this problem by first learning to generate context, question and answer triples in an unsupervised manner, which we then use to synthesize Extractive QA training data automatically.To generate such triples, we first sample random context paragraphs from a large corpus of documents and then random noun phrases or named entity mentions from these paragraphs as answers.Next we convert answers in context to "fill-in-the-blank" cloze questions and finally translate them into natural questions.We propose and compare various unsupervised ways to perform cloze-tonatural question translation, including training an unsupervised NMT model using nonaligned corpora of natural questions and cloze questions as well as a rule-based approach.We find that modern QA models can learn to answer human questions surprisingly well using only synthetic training data.We demonstrate that, without using the SQuAD training data at all, our approach achieves 56.4 F1 on SQuAD v1 (64.5 F1 when the answer is a Named entity mention), outperforming early supervised models.

The problem of keyword spotting i.e. identifying keywords in a real-time audio stream is mainly solved by applying a neural network over successive sliding windows. Due to the difficulty of the task, baseline models are usually large, resulting in a high computational cost and energy consumption level. We propose a new method called SANAS (Stochastic Adaptive Neural Architecture Search) which is able to adapt the architecture of the neural network on-the-fly at inference time such that small architectures will be used when the stream is easy to process (silence, low noise, ...) and bigger networks will be used when the task becomes more difficult. We show that this adaptive model can be learned end-to-end by optimizing a trade-off between the prediction performance and the average computational cost per unit of time. Experiments on the Speech Commands dataset [1] show that this approach leads to a high recognition level while being much faster (and/or energy saving) than classical approaches where the network architecture is static.

Language models are at the heart of numerous works, notably in the text mining and information retrieval communities. These statistical models aim at extracting word distributions, from simple unigram models to recurrent approaches with latent variables that capture subtle dependencies in texts. However, those models are learned from word sequences only, and authors' identities, as well as publication dates, are seldom considered. We propose a neural model, based on recurrent language modeling, which aims at capturing language diffusion tendencies in author communities through time. By conditioning language models with author and temporal vector states, we are able to leverage the latent dependencies between the text contexts. This allows us to beat several temporal and non-temporal language baselines on two real-world corpora, and to learn meaningful author representations that vary through time.

Object segmentation is a crucial problem that is usually solved by using supervised learning approaches over very large datasets composed of both images and corresponding object masks. Since the masks have to be provided at pixel level, building such a dataset for any new domain can be very costly. We present ReDO, a new model able to extract objects from images without any annotation in an unsupervised way. It relies on the idea that it should be possible to change the textures or colors of the objects without changing the overall distribution of the dataset. Following this assumption, our approach is based on an adversarial architecture where the generator is guided by an input sample: given an image, it extracts the object mask, then redraws a new object at the same location. The generator is controlled by a discriminator that ensures that the distribution of generated images is aligned to the original one. We experiment with this method on different datasets and demonstrate the good quality of extracted masks.

This paper introduces a structured memory which can be easily integrated into a neural network. The memory is very large by design and significantly increases the capacity of the architecture, by up to a billion parameters with a negligible computational overhead. Its design and access pattern is based on product keys, which enable fast and exact nearest neighbor search. The ability to increase the number of parameters while keeping the same computational budget lets the overall system strike a better trade-off between prediction accuracy and computation efficiency both at training and test time. This memory layer allows us to tackle very large scale language modeling tasks. In our experiments we consider a dataset with up to 30 billion words, and we plug our memory layer in a state-of-the-art transformer-based architecture. In particular, we found that a memory augmented model with only 12 layers outperforms a baseline transformer model with 24 layers, while being twice faster at inference time. We release our code for reproducibility purposes.

We propose to focus on the problem of discovering neural network architectures efficient in terms of both prediction quality and cost. For instance, our approach is able to solve the following tasks: learn a neural network able to predict well in less than 100 milliseconds or learn an efficient model that fits in a 50 Mb memory. Our contribution is a novel family of models called Budgeted Super Networks (BSN). They are learned using gradient descent techniques applied on a budgeted learning objective function which integrates a maximum authorized cost, while making no assumption on the nature of this cost. We present a set of experiments on computer vision problems and analyze the ability of our technique to deal with three different costs: the computation cost, the memory consumption cost and a distributed computation cost. We particularly show that our model can discover neural network architectures that have a better accuracy than the ResNet and Convolutional Neural Fabrics architectures on CIFAR-10 and CIFAR-100, at a lower cost.

Machine translation systems achieve near human-level performance on some languages, yet their effectiveness strongly relies on the availability of large amounts of parallel sentences, which hinders their applicability to the majority of language pairs.This work investigates how to learn to translate when having access to only large monolingual corpora in each language.We propose two model variants, a neural and a phrase-based model.Both versions leverage a careful initialization of the parameters, the denoising effect of language models and automatic generation of parallel data by iterative back-translation.These models are significantly better than methods from the literature, while being simpler and having fewer hyper-parameters.On the widely used WMT'14 English-French and WMT'16 German-English benchmarks, our models respectively obtain 28.1 and 25.2 BLEU points without using a single parallel sentence, outperforming the state of the art by more than 11 BLEU points.On low-resource languages like English-Urdu and English-Romanian, our methods achieve even better results than semisupervised and supervised approaches leveraging the paucity of available bitexts.Our code for NMT and PBSMT is publicly available.

In hierarchical reinforcement learning, the framework of options models sub-policies over a set of primitive actions. In this paper, we address the problem of discovering and learning options from scratch. Inspired by recent works in cognitive science, our approach is based on a new budgeted learning approach in which options naturally arise as a way to minimize the cognitive effort of the agent. In our case, this effort corresponds to the amount of information acquired by the agent at each time step. We propose the Budgeted Hierarchical Neural Network model (BHNN), a hierarchical recurrent neural network architecture that learns latent options as continuous vectors. With respect to existing approaches, BHNN does not need to explicitly predefine sub-goals nor to a priori define the number of possible options. We evaluate this model on different classical RL problems showing the quality of the resulting learned policy.

We address the task of node classification in heterogeneous networks, where the nodes are of different types, each type having its own set of labels, and the relations between nodes may also be of different types. A typical example is provided by social networks where node types may for example be users, content, or films, and relations friendship , like , authorship . Learning and performing inference on such heterogeneous networks is a recent task requiring new models and algorithms. We propose a model, Labeling Heterogeneous Network (LaHNet) , a transductive approach to classification that learns to project the different types of nodes into a common latent space. This embedding is learned so as to reflect different characteristics of the problem such as the correlation between node labels, as well as the graph topology. The application focus is on social graphs, but the algorithm is general and can be used for other domains. The model is evaluated on five datasets representative of different instances of social data.

We introduce a dynamical spatio-temporal model formalized as a recurrent neural network for forecasting time series of spatial processes, i.e. series of observations sharing temporal and spatial dependencies. The model learns these dependencies through a structured latent dynamical component, while a decoder predicts the observations from the latent representations. We consider several variants of this model, corresponding to different prior hypothesis about the spatial relations between the series. The model is evaluated and compared to state-of-the-art baselines, on a variety of forecasting problems representative of different application areas: epidemiology, geo-spatial statistics and car-traffic prediction. Besides these evaluations, we also describe experiments showing the ability of this approach to extract relevant spatial relations.

This paper introduces a new encoder-decoder architecture that is trained to reconstruct images by disentangling the salient information of the image and the values of attributes directly in the latent space. As a result, after training, our model can generate different realistic versions of an input image by varying the attribute values. By using continuous attribute values, we can choose how much a specific attribute is perceivable in the generated image. This property could allow for applications where users can modify an image using sliding knobs, like faders on a mixing console, to change the facial expression of a portrait, or to update the color of some objects. Compared to the state-of-the-art which mostly relies on training adversarial networks in pixel space by altering attribute values at train time, our approach results in much simpler training schemes and nicely scales to multiple attributes. We present evidence that our model can significantly change the perceived value of the attributes while preserving the naturalness of images.

Data privacy is a major problem that has to be considered before releasing datasets to the public or even to a partner company that would compute statistics or make a deep analysis of these data. This is insured by performing data anonymization as required by legislation. In this context, many different anonymization techniques have been proposed in the literature. These methods are usually specific to a particular de-anonymization procedure-or attack-one wants to avoid, and to a particular known set of characteristics that have to be preserved after the anonymization. They are difficult to use in a general context where attacks can be of different types, and where measures are not known to the anonymizer. The paper proposes a novel approach for automatically finding an anonymization procedure given a set of possible attacks and a set of measures to preserve. The approach is generic and based on machine learning techniques. It allows us to learn directly an anonymization function from a set of training data so as to optimize a trade off between privacy risk and utility loss. The algorithm thus allows one to get a good anonymization procedure for any kind of attacks, and any characteristic in a given set. Experiments made on two datasets show the effectiveness and the genericity of the approach.

Analyzing and modeling the temporal diffusion of information on social media has mainly been treated as a diffusion process on known graphs or proximity structures. The underlying phenomenon results however from the interactions of several actors and media and is more complex than what these models can account for and cannot be explained using such limiting assumptions. We introduce here a new approach to this problem whose goal is to learn a mapping of the observed temporal dynamic onto a continuous space. Nodes participating to diffusion cascades are projected in a latent representation space in such a way that information diffusion can be modeled efficiently using a heat diffusion process. This amounts to learning a diffusion kernel for which the proximity of nodes in the projection space reflects the proximity of their infection time in cascades. The proposed approach possesses several unique characteristics compared to existing ones. Since its parameters are directly learned from cascade samples without requiring any additional information, it does not rely on any pre-existing diffusion structure. Because the solution to the diffusion equation can be expressed in a closed form in the projection space, the inference time for predicting the diffusion of a new piece of information is greatly reduced compared to discrete models. Experiments and comparisons with baselines and alternative models have been performed on both synthetic networks and real datasets. They show the effectiveness of the proposed method both in terms of prediction quality and of inference speed.

Social networks are heterogeneous systems composed of different types of nodes (e.g. users, content, groups, etc.) and relations (e.g. social or similarity relations). While learning and performing inference on homogeneous networks have motivated a large amount of research, few work exists on heterogeneous networks and there are open and challenging issues for existing methods that were previously developed for homogeneous networks. We address here the specific problem of nodes classification and tagging in heterogeneous social networks, where different types of nodes are considered, each type with its own label or tag set. We propose a new method for learning node representations onto a latent space, common to all the different node types. Inference is then performed in this latent space. In this framework, two nodes connected in the network will tend to share similar representations regardless of their types. This allows bypassing limitations of the methods based on direct extensions of homogenous frameworks and exploiting the dependencies and correlations between the different node types. The proposed method is tested on two representative datasets and compared to state-of-the-art methods and to baselines.

Maximizing the spread and influence of the messages being published is a challenge for many social network users. Selecting the right content according to the information context and the user characteristics is essential for achieving this goal. We propose a model to automatically choose which information to publish on social networks given a set of possible messages. This model will tend to maximize the spread of the published message for a specific audience. The algorithm is based on the use of a contextual bandit model treating each new potential message as an arm to be selected. We conduct experiments on a Twitter dataset, comparing different algorithms and exploring the influence of the content and the characteristics of the messages on the information spread. The results demonstrate the model's ability to maximize the published information flow as well as it's ability to adapt its behavior to each particular audience.

We address the problem of link prediction in collections of objects connected by multiple relation types, where each type may play a distinct role. While traditional link prediction models are limited to single-type link prediction we attempt here to jointly model and predict the multiple relation types, which we refer to as the Link Pattern Prediction (LPP) problem. For that, we propose a tensor decomposition model to solve the LPP problem, which allows to capture the correlations among different relation types and reveal the impact of various relations on prediction performance. The proposed tensor decomposition model is efficiently learned with a conjugate gradient based optimization method. Extensive experiments on real-world datasets demonstrate that this model outperforms the traditional mono-relational model and can achieve better prediction quality.

In this paper we address the problem of temporal link prediction, i.e., predicting the apparition of new links, in time-evolving networks. This problem appears in applications such as recommender systems, social network analysis or citation analysis. Link prediction in time-evolving networks is usually based on the topological structure of the network only. We propose here a model which exploits multiple information sources in the network in order to predict link occurrence probabilities as a function of time. The model integrates three types of information: the global network structure, the content of nodes in the network if any, and the local or proximity information of a given vertex. The proposed model is based on a matrix factorization formulation of the problem with graph regularization. We derive an efficient optimization method to learn the latent factors of this model. Extensive experiments on several real world datasets suggest that our unified framework outperforms state-of-the-art methods for temporal link prediction tasks.

We consider the problem of learning to annotate documents with concepts or keywords in content information networks, where the documents may share multiple relations. The concepts associated to a document will depend both on its content and on its neighbors in the network through the different relations. We formalize this problem as single and multi-label classification in a multi-graph, the nodes being the documents and the edges representing the different relations. The proposed algorithm learns to weight the different relations according to their importance for the annotation task. We perform experiments on different corpora corresponding to different annotation tasks on scientific articles, emails and Flickr images and show how the model may take advantage of the rich relational information.

We consider here the task of multi-label classification for data organized in a multi-relational graph. We propose the IMMCA model - Iterative Multi-label Multi-Relational Classification Algorithm - a general algorithm for solving the inference and learning problems for this task. Inference is performed iteratively by propagating scores according to the multi-relational structure of the data. We detail two instances of this general model, implementing two different label propagation schemes on the multi-graph. This is the first collective classification method able to handle multiple relations and to perform multi-label classification in multi-graphs. The target application is image annotation in large social media sharing web sites (Flickr). The goal is to assign labels for images when users and images are connected through multiple relations - authorship, friendship, or visual/textual similarities. We show that our model is able to deal with both content and social relations and performs well on real datasets. Additional experiments on artificial data allow us analyzing the behavior of our method in different situations.

International audience