arXiv subjects

Hung Le

Publications and source records attributed to Hung Le.

At least 145 records · Page 8Linked to original sources

arXiv2021

DeepProcess: Supporting business process execution using a MANN-based recommender system

Process-aware Recommender systems can provide critical decision support functionality to aid business process execution by recommending what actions to take next. Based on recent advances in the field of deep learning, we present a novel memory-augmented neural network (MANN) based approach for constructing a process-aware recommender system. We propose a novel network architecture, namely Write-Protected Dual Controller Memory-Augmented Neural Network (DCw-MANN), for building prescriptive models. To evaluate the feasibility and usefulness of our approach, we consider three real-world datasets and show that our approach leads to better performance on several baselines for the task of suffix recommendation and next task prediction.

cs.NE↗

arXiv2021

Robust Deep Reinforcement Learning for Extractive Legal Summarization

Automatic summarization of legal texts is an important and still a challenging task since legal documents are often long and complicated with unusual structures and styles. Recent advances of deep models trained end-to-end with differentiable losses can well-summarize natural text, yet when applied to legal domain, they show limited results. In this paper, we propose to use reinforcement learning to train current deep summarization models to improve their performance on the legal domain. To this end, we adopt proximal policy optimization methods and introduce novel reward functions that encourage the generation of candidate summaries satisfying both lexical and semantic criteria. We apply our method to training different summarization backbones and observe a consistent and significant performance gain across 3 public legal datasets.

cs.CL↗

arXiv2021

Model-Based Episodic Memory Induces Dynamic Hybrid Controls

Episodic control enables sample efficiency in reinforcement learning by recalling past experiences from an episodic memory. We propose a new model-based episodic memory of trajectories addressing current limitations of episodic control. Our memory estimates trajectory values, guiding the agent towards good policies. Built upon the memory, we construct a complementary learning model via a dynamic hybrid control unifying model-based, episodic and habitual learning into a single architecture. Experiments demonstrate that our model allows significantly faster and better learning than other strong reinforcement learning agents across a variety of environments including stochastic and non-Markovian settings.

cs.LG↗

arXiv2021

Optimal Approximate Distance Oracle for Planar Graphs

A $(1+ε)$-approximate distance oracle of an edge-weighted graph is a data structure that returns an approximate shortest path distance between any two query vertices up to a $(1+ε)$ factor. Thorup (FOCS 2001, JACM 2004) and Klein (SODA 2002) independently constructed a $(1+ε)$-approximate distance oracle with $O(n\log n)$ space, measured in number of words, and $O(1)$ query time when $G$ is an undirected planar graph with $n$ vertices and $ε$ is a fixed constant. Many follow-up works gave $(1+ε)$-approximate distance oracles with various trade-offs between space and query time. However, improving $O(n\log n)$ space bound without sacrificing query time remains an open problem for almost two decades. In this work, we resolve this problem affirmatively by constructing a $(1+ε)$-approximate distance oracle with optimal $O(n)$ space and $O(1)$ query time for undirected planar graphs and fixed $ε$. We also make substantial progress for planar digraphs with non-negative edge weights. For fixed $ε> 0$, we give a $(1+ε)$-approximate distance oracle with space $o(n\log(Nn))$ and $O(\log\log(Nn)$ query time; here $N$ is the ratio between the largest and smallest positive edge weight. This improves Thorup's (FOCS 2001, JACM 2004) $O(n\log(Nn)\log n)$ space bound by more than a logarithmic factor while matching the query time of his structure. This is the first improvement for planar digraphs in two decades, both in the weighted and unweighted setting.

cs.DS↗

arXiv2021

Balanced Q-learning: Combining the Influence of Optimistic and Pessimistic Targets

The optimistic nature of the Q-learning target leads to an overestimation bias, which is an inherent problem associated with standard $Q-$learning. Such a bias fails to account for the possibility of low returns, particularly in risky scenarios. However, the existence of biases, whether overestimation or underestimation, need not necessarily be undesirable. In this paper, we analytically examine the utility of biased learning, and show that specific types of biases may be preferable, depending on the scenario. Based on this finding, we design a novel reinforcement learning algorithm, Balanced Q-learning, in which the target is modified to be a convex combination of a pessimistic and an optimistic term, whose associated weights are determined online, analytically. We prove the convergence of this algorithm in a tabular setting, and empirically demonstrate its superior learning performance in various environments.

cs.LG↗

arXiv2021

Plug and Play, Model-Based Reinforcement Learning

Sample-efficient generalisation of reinforcement learning approaches have always been a challenge, especially, for complex scenes with many components. In this work, we introduce Plug and Play Markov Decision Processes, an object-based representation that allows zero-shot integration of new objects from known object classes. This is achieved by representing the global transition dynamics as a union of local transition functions, each with respect to one active object in the scene. Transition dynamics from an object class can be pre-learnt and thus would be ready to use in a new environment. Each active object is also endowed with its reward function. Since there is no central reward function, addition or removal of objects can be handled efficiently by only updating the reward functions of objects involved. A new transfer learning mechanism is also proposed to adapt reward function in such cases. Experiments show that our representation can achieve sample-efficiency in a variety of set-ups.

cs.LG↗

arXiv2021

Near-Optimal Spanners for General Graphs in (Nearly) Linear Time

Let $G = (V,E,w)$ be a weighted undirected graph on $|V| = n$ vertices and $|E| = m$ edges, let $k \ge 1$ be any integer, and let $ε< 1$ be any parameter. We present the following results on fast constructions of spanners with near-optimal sparsity and lightness, which culminate a long line of work in this area. (By near-optimal we mean optimal under Erdős' girth conjecture and disregarding the $ε$-dependencies.) - There are (deterministic) algorithms for constructing $(2k-1)(1+ε)$-spanners for $G$ with a near-optimal sparsity of $O(n^{1/k} \log(1/ε)/ε))$. The first algorithm can be implemented in the pointer-machine model within time $O(mα(m,n) \log(1/ε)/ε) + SORT(m))$, where $α( , )$ is the two-parameter inverse-Ackermann function and $SORT(m)$ is the time needed to sort $m$ integers. The second algorithm can be implemented in the WORD RAM model within time $O(m \log(1/ε)/ε))$. - There is a (deterministic) algorithm for constructing a $(2k-1)(1+ε)$-spanner for $G$ that achieves a near-optimal bound of $O(n^{1/k}\mathrm{poly}(1/ε))$ on both sparsity and lightness. This algorithm can be implemented in the pointer-machine model within time $O(mα(m,n) \mathrm{poly}(1/ε) + SORT(m))$ and in the WORD RAM model within time $O(m α(m,n) \mathrm{poly}(1/ε))$. The previous fastest constructions of $(2k-1)(1+ε)$-spanners with near-optimal sparsity incur a runtime of is $O(\min\{m(n^{1+1/k}) + n\log n,k n^{2+1/k}\})$, even regardless of the lightness. Importantly, the greedy spanner for stretch $2k-1$ has sparsity $O(n^{1/k})$ -- with no $ε$-dependence whatsoever, but its runtime is $O(m(n^{1+1/k} + n\log n))$. Moreover, the state-of-the-art lightness bound of any $(2k-1)$-spanner is poor, even regardless of the sparsity and runtime.

cs.DS↗

arXiv2021

A New Representation of Successor Features for Transfer across Dissimilar Environments

Transfer in reinforcement learning is usually achieved through generalisation across tasks. Whilst many studies have investigated transferring knowledge when the reward function changes, they have assumed that the dynamics of the environments remain consistent. Many real-world RL problems require transfer among environments with different dynamics. To address this problem, we propose an approach based on successor features in which we model successor feature functions with Gaussian Processes permitting the source successor features to be treated as noisy measurements of the target successor feature function. Our theoretical analysis proves the convergence of this approach as well as the bounded error on modelling successor feature functions with Gaussian Processes in environments with both different dynamics and rewards. We demonstrate our method on benchmark datasets and show that it outperforms current baselines.

cs.LG↗

arXiv2021

Memory and attention in deep learning

Intelligence necessitates memory. Without memory, humans fail to perform various nontrivial tasks such as reading novels, playing games or solving maths. As the ultimate goal of machine learning is to derive intelligent systems that learn and act automatically just like human, memory construction for machine is inevitable. Artificial neural networks model neurons and synapses in the brain by interconnecting computational units via weights, which is a typical class of machine learning algorithms that resembles memory structure. Their descendants with more complicated modeling techniques (a.k.a deep learning) have been successfully applied to many practical problems and demonstrated the importance of memory in the learning process of machinery systems. Recent progresses on modeling memory in deep learning have revolved around external memory constructions, which are highly inspired by computational Turing models and biological neuronal systems. Attention mechanisms are derived to support acquisition and retention operations on the external memory. Despite the lack of theoretical foundations, these approaches have shown promises to help machinery systems reach a higher level of intelligence. The aim of this thesis is to advance the understanding on memory and attention in deep learning. Its contributions include: (i) presenting a collection of taxonomies for memory, (ii) constructing new memory-augmented neural networks (MANNs) that support multiple control and memory units, (iii) introducing variability via memory in sequential generative models, (iv) searching for optimal writing operations to maximise the memorisation capacity in slot-based memory networks, and (v) simulating the Universal Turing Machine via Neural Stored-program Memory-a new kind of external memory for neural networks.

cs.LG↗

arXiv2021

DVD: A Diagnostic Dataset for Multi-step Reasoning in Video Grounded Dialogue

A video-grounded dialogue system is required to understand both dialogue, which contains semantic dependencies from turn to turn, and video, which contains visual cues of spatial and temporal scene variations. Building such dialogue systems is a challenging problem, involving various reasoning types on both visual and language inputs. Existing benchmarks do not have enough annotations to thoroughly analyze dialogue systems and understand their capabilities and limitations in isolation. These benchmarks are also not explicitly designed to minimise biases that models can exploit without actual reasoning. To address these limitations, in this paper, we present DVD, a Diagnostic Dataset for Video-grounded Dialogues. The dataset is designed to contain minimal biases and has detailed annotations for the different types of reasoning over the spatio-temporal space of video. Dialogues are synthesized over multiple question turns, each of which is injected with a set of cross-turn semantic relationships. We use DVD to analyze existing approaches, providing interesting insights into their abilities and limitations. In total, DVD is built from $11k$ CATER synthetic videos and contains $10$ instances of $10$-round dialogues for each video, resulting in more than $100k$ dialogues and $1M$ question-answer pairs. Our code and dataset are publicly available at https://github.com/facebookresearch/DVDialogues.

cs.AI↗

arXiv2021

Clan Embeddings into Trees, and Low Treewidth Graphs

In low distortion metric embeddings, the goal is to embed a host "hard" metric space into a "simpler" target space while approximately preserving pairwise distances. A highly desirable target space is that of a tree metric. Unfortunately, such embedding will result in a huge distortion. A celebrated bypass to this problem is stochastic embedding with logarithmic expected distortion. Another bypass is Ramsey-type embedding, where the distortion guarantee applies only to a subset of the points. However, both these solutions fail to provide an embedding into a single tree with a worst-case distortion guarantee on all pairs. In this paper, we propose a novel third bypass called \emph{clan embedding}. Here each point $x$ is mapped to a subset of points $f(x)$, called a \emph{clan}, with a special \emph{chief} point $χ(x)\in f(x)$. The clan embedding has multiplicative distortion $t$ if for every pair $(x,y)$ some copy $y'\in f(y)$ in the clan of $y$ is close to the chief of $x$: $\min_{y'\in f(y)}d(y',χ(x))\le t\cdot d(x,y)$. Our first result is a clan embedding into a tree with multiplicative distortion $O(\frac{\log n}ε)$ such that each point has $1+ε$ copies (in expectation). In addition, we provide a "spanning" version of this theorem for graphs and use it to devise the first compact routing scheme with constant size routing tables. We then focus on minor-free graphs of diameter prameterized by $D$, which were known to be stochastically embeddable into bounded treewidth graphs with expected additive distortion $εD$. We devise Ramsey-type embedding and clan embedding analogs of the stochastic embedding. We use these embeddings to construct the first (bicriteria quasi-polynomial time) approximation scheme for the metric $ρ$-dominating set and metric $ρ$-independent set problems in minor-free graphs.

cs.DS↗

arXiv2021

Truly Optimal Euclidean Spanners

Euclidean spanners are important geometric structures, having found numerous applications over the years. Cornerstone results in this area from the late 80s and early 90s state that for any $d$-dimensional $n$-point Euclidean space, there exists a $(1+ε)$-spanner with $nO(ε^{-d+1})$ edges and lightness $O(ε^{-2d})$. Surprisingly, the fundamental question of whether or not these dependencies on $ε$ and $d$ for small $d$ can be improved has remained elusive, even for $d = 2$. This question naturally arises in any application of Euclidean spanners where precision is a necessity. The state-of-the-art bounds $nO(ε^{-d+1})$ and $O(ε^{-2d})$ on the size and lightness of spanners are realized by the {\em greedy} spanner. In 2016, Filtser and Solomon proved that, in low dimensional spaces, the greedy spanner is near-optimal. The question of whether the greedy spanner is truly optimal remained open to date. The contribution of this paper is two-fold. We resolve these longstanding questions by nailing down the exact dependencies on $ε$ and $d$ and showing that the greedy spanner is truly optimal. Specifically, for any $d= O(1), ε= Ω({n}^{-\frac{1}{d-1}})$: - We show that any $(1+ε)$-spanner must have $n Ω(ε^{-d+1})$ edges, implying that the greedy (and other) spanners achieve the optimal size. - We show that any $(1+ε)$-spanner must have lightness $Ω(ε^{-d})$, and then improve the upper bound on the lightness of the greedy spanner from $O(ε^{-2d})$ to $O(ε^{-d})$. We then complement our negative result for the size of spanners with a rather counterintuitive positive result: Steiner points lead to a quadratic improvement in the size of spanners! Our bound for the size of Steiner spanners is tight as well (up to lower-order terms).

cs.CG↗

arXiv2020

Waves of maximal height for a class of nonlocal equations with inhomogeneous symbols

In this paper, we consider a class of nonlocal equations where the convolution kernel is given by a Bessel potential symbol of order $α$ for $α> 1$. Based on the properties of the convolution operator, we apply a global bifurcation technique to show the existence of a highest, even, $2π$-periodic traveling-wave solution. The regularity of this wave is proved to be exactly Lipschitz.

math.AP↗

arXiv2020

UniConv: A Unified Conversational Neural Architecture for Multi-domain Task-oriented Dialogues

Building an end-to-end conversational agent for multi-domain task-oriented dialogues has been an open challenge for two main reasons. First, tracking dialogue states of multiple domains is non-trivial as the dialogue agent must obtain complete states from all relevant domains, some of which might have shared slots among domains as well as unique slots specifically for one domain only. Second, the dialogue agent must also process various types of information across domains, including dialogue context, dialogue states, and database, to generate natural responses to users. Unlike the existing approaches that are often designed to train each module separately, we propose "UniConv" -- a novel unified neural architecture for end-to-end conversational systems in multi-domain task-oriented dialogues, which is designed to jointly train (i) a Bi-level State Tracker which tracks dialogue states by learning signals at both slot and domain level independently, and (ii) a Joint Dialogue Act and Response Generator which incorporates information from various input components and models dialogue acts and target responses simultaneously. We conduct comprehensive experiments in dialogue state tracking, context-to-text, and end-to-end settings on the MultiWOZ2.1 benchmark, achieving superior performance over competitive baselines.

cs.CL↗

arXiv2020

BiST: Bi-directional Spatio-Temporal Reasoning for Video-Grounded Dialogues

Video-grounded dialogues are very challenging due to (i) the complexity of videos which contain both spatial and temporal variations, and (ii) the complexity of user utterances which query different segments and/or different objects in videos over multiple dialogue turns. However, existing approaches to video-grounded dialogues often focus on superficial temporal-level visual cues, but neglect more fine-grained spatial signals from videos. To address this drawback, we propose Bi-directional Spatio-Temporal Learning (BiST), a vision-language neural framework for high-resolution queries in videos based on textual cues. Specifically, our approach not only exploits both spatial and temporal-level information, but also learns dynamic information diffusion between the two feature spaces through spatial-to-temporal and temporal-to-spatial reasoning. The bidirectional strategy aims to tackle the evolving semantics of user queries in the dialogue setting. The retrieved visual cues are used as contextual information to construct relevant responses to the users. Our empirical results and comprehensive qualitative analysis show that BiST achieves competitive performance and generates reasonable responses on a large-scale AVSD benchmark. We also adapt our BiST models to the Video QA setting, and substantially outperform prior approaches on the TGIF-QA benchmark.

cs.CV↗

arXiv2020

Light Euclidean Spanners with Steiner Points

The FOCS'19 paper of Le and Solomon, culminating a long line of research on Euclidean spanners, proves that the lightness (normalized weight) of the greedy $(1+ε)$-spanner in $\mathbb{R}^d$ is $\tilde{O}(ε^{-d})$ for any $d = O(1)$ and any $ε= Ω(n^{-\frac{1}{d-1}})$ (where $\tilde{O}$ hides polylogarithmic factors of $\frac{1}ε$), and also shows the existence of point sets in $\mathbb{R}^d$ for which any $(1+ε)$-spanner must have lightness $Ω(ε^{-d})$. Given this tight bound on the lightness, a natural arising question is whether a better lightness bound can be achieved using Steiner points. Our first result is a construction of Steiner spanners in $\mathbb{R}^2$ with lightness $O(ε^{-1} \log Δ)$, where $Δ$ is the spread of the point set. In the regime of $Δ\ll 2^{1/ε}$, this provides an improvement over the lightness bound of Le and Solomon [FOCS 2019]; this regime of parameters is of practical interest, as point sets arising in real-life applications (e.g., for various random distributions) have polynomially bounded spread, while in spanner applications $ε$ often controls the precision, and it sometimes needs to be much smaller than $O(1/\log n)$. Moreover, for spread polynomially bounded in $1/ε$, this upper bound provides a quadratic improvement over the non-Steiner bound of Le and Solomon [FOCS 2019], We then demonstrate that such a light spanner can be constructed in $O_ε(n)$ time for polynomially bounded spread, where $O_ε$ hides a factor of $\mathrm{poly}(\frac{1}ε)$. Finally, we extend the construction to higher dimensions, proving a lightness upper bound of $\tilde{O}(ε^{-(d+1)/2} + ε^{-2}\log Δ)$ for any $3\leq d = O(1)$ and any $ε= Ω(n^{-\frac{1}{d-1}})$.

cs.CG↗

arXiv2020

Neurocoder: Learning General-Purpose Computation Using Stored Neural Programs

Artificial Neural Networks are uniquely adroit at machine learning by processing data through a network of artificial neurons. The inter-neuronal connection weights represent the learnt Neural Program that instructs the network on how to compute the data. However, without an external memory to store Neural Programs, they are restricted to only one, overwriting learnt programs when trained on new data. This is functionally equivalent to a special-purpose computer. Here we design Neurocoder, an entirely new class of general-purpose conditional computational machines in which the neural network "codes" itself in a data-responsive way by composing relevant programs from a set of shareable, modular programs. This can be considered analogous to building Lego structures from simple Lego bricks. Notably, our bricks change their shape through learning. External memory is used to create, store and retrieve modular programs. Like today's stored-program computers, Neurocoder can now access diverse programs to process different data. Unlike manually crafted computer programs, Neurocoder creates programs through training. Integrating Neurocoder into current neural architectures, we demonstrate new capacity to learn modular programs, handle severe pattern shifts and remember old programs as new ones are learnt, and show substantial performance improvement in solving object recognition, playing video games and continual learning tasks. Such integration with Neurocoder increases the computation capability of any current neural network and endows it with entirely new capacity to reuse simple programs to build complex ones. For the first time a Neural Program is treated as a datum in memory, paving the ways for modular, recursive and procedural neural programming.

cs.LG↗

arXiv2020

On Light Spanners, Low-treewidth Embeddings and Efficient Traversing in Minor-free Graphs

Understanding the structure of minor-free metrics, namely shortest path metrics obtained over a weighted graph excluding a fixed minor, has been an important research direction since the fundamental work of Robertson and Seymour. A fundamental idea that helps both to understand the structural properties of these metrics and lead to strong algorithmic results is to construct a "small-complexity" graph that approximately preserves distances between pairs of points of the metric. We show the two following structural results for minor-free metrics: 1. Construction of a light subset spanner. Given a subset of vertices called terminals, and $ε$, in polynomial time we construct a subgraph that preserves all pairwise distances between terminals up to a multiplicative $1+ε$ factor, of total weight at most $O_ε(1)$ times the weight of the minimal Steiner tree spanning the terminals. 2. Construction of a stochastic metric embedding into low treewidth graphs with expected additive distortion $εD$. Namely, given a minor free graph $G=(V,E,w)$ of diameter $D$, and parameter $ε$, we construct a distribution $\mathcal{D}$ over dominating metric embeddings into treewidth-$O_ε(\log n)$ graphs such that the additive distortion is at most $εD$. One of our important technical contributions is a novel framework that allows us to reduce \emph{both problems} to problems on simpler graphs of bounded diameter. Our results have the following algorithmic consequences: (1) the first efficient approximation scheme for subset TSP in minor-free metrics; (2) the first approximation scheme for vehicle routing with bounded capacity in minor-free metrics; (3) the first efficient approximation scheme for vehicle routing with bounded capacity on bounded genus metrics.

cs.DS↗