Zhengying Liu

Zhengying Liu

Auto-generated profile, based on publicly available information

Zhengying Liu

Noah Ark's Lab, Hong Kong Hong Kong

Research Focus & Metrics

Academic Career Length

7Years (2020-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
11 April 2025 in ACM Computing Surveys
Verified
Journal Article
Open access
JSJiankai Sun
Jiankai Sun
CZChuanyang Zheng
Chuanyang Zheng
EXEnze Xie
Enze Xie
ZLZhengying Liu
Zhengying Liu
RCRuihang Chu
Ruihang Chu
ZLZhenguo Li
Zhenguo Li
Reasoning, a crucial ability for complex problem-solving, plays a pivotal role in various real-world settings such as negotiation, medical diagnosis, and criminal investigation. It serves as a fundamental methodology in the field of Artificial General Intelligence (AGI). With the ongoing development of foundation models, there is a growing interest in exploring their abilities in reasoning tasks. In this paper, we introduce seminal foundation models proposed or adaptable for reasoning, highlighting the latest advancements in various reasoning tasks, methods, and benchmarks. We then delve into the potential future directions behind the emergence of reasoning abilities within foundation models. We also discuss the relevance of multimodal learning, autonomous agents, and super alignment in the context of reasoning. By discussing these future research directions, we hope to inspire researchers in their exploration of this field, stimulate further advancements in reasoning with foundation models, e.g. Large Language Models (LLMs), and contribute to the development of AGI.

23 April 2021 in IEEE Transactions on Pattern Analysis and Machine Intelligence
Verified
Journal Article
Closed access
ZLZhengying Liu
Zhengying Liu
APAdrien Pavao
Adrien Pavao
ZXZhen Xu
Zhen Xu
Sergio EscaleraSergio Escalera
Sergio Escalera
FFFabio Ferreira
Fabio Ferreira
YZYang Zhang
Yang Zhang
This paper reports the results and post-challenge analyses of ChaLearn's AutoDL challenge series, which helped sorting out a profusion of AutoML solutions for Deep Learning (DL) that had been introduced in a variety of settings, but lacked fair comparisons. All input data modalities (time series, images, videos, text, tabular) were formatted as tensors and all tasks were multi-label classification problems. Code submissions were executed on hidden tasks, with limited time and computational resources, pushing solutions that get results quickly. In this setting, DL methods dominated, though popular Neural Architecture Search (NAS) was impractical. Solutions relied on fine-tuned pre-trained networks, with architectures matching data modality. Post-challenge tests did not reveal improvements beyond the imposed time limit. While no component is particularly original or novel, a high level modular organization emerged featuring a ‘`meta-learner’', ‘`data ingestor’', ‘`model selector’', ‘`model/learner’', and ‘`evaluator’'. This modularity enabled ablation studies, which revealed the importance of (off-platform) meta-learning, ensembling, and efficient data management. Experiments on heterogeneous module combinations further confirm the (local) optimality of the winning solutions. Our challenge legacy includes an ever-lasting benchmark (http://autodl.chalearn.org), the open-sourced code of the winners, and a free 'AutoDL self-service''.

11 July 2022 in arXiv
Verified
Preprint
Open access
PDF file
ABAdrian El Baz
Adrian El Baz
IUIhsan Ullah
Ihsan Ullah
EAEdesio Alcobaça
Edesio Alcobaça
ACAndré C. P. L. F. Carvalho
André C. P. L. F. Carvalho
HCHong Chen
Hong Chen
WZWenwu Zhu
Wenwu Zhu
Although deep neural networks are capable of achieving performance superior to humans on various tasks, they are notorious for requiring large amounts of data and computing resources, restricting their success to domains where such resources are available. Metalearning methods can address this problem by transferring knowledge from related tasks, thus reducing the amount of data and computing resources needed to learn new tasks. We organize the MetaDL competition series, which provide opportunities for research groups all over the world to create and experimentally assess new meta-(deep)learning solutions for real problems. In this paper, authored collaboratively between the competition organizers and the top-ranked participants, we describe the design of the competition, the datasets, the best experimental results, as well as the top-ranked methods in the NeurIPS 2021 challenge, which attracted 15 active teams who made it to the final phase (by outperforming the baseline), making over 100 code submissions during the feedback phase. The solutions of the top participants have been open-sourced. The lessons learned include that learning good representations is essential for effective transfer learning.

4 April 2026
Verified
Thesis
Closed access
ZLZhengying Liu
Zhengying Liu
Apprentissage profond automatisé : principes et pratique: L'apprentissage automatique automatisé (AutoML) vise à rendre l'application des méthodes d'apprentissage automatique (ML) aussi dépourvue d'intervention humaine que possible. Cet objectif ambitieux a fait l'objet de nombreuses recherches depuis les débuts du ML. L'objectif de cette thèse est de mettre un cadre formel autour de ce problème aux multiples facettes, de comparer les méthodes existantes et d'explorer de nouvelles directions. Pour formuler le problème AutoML de manière rigoureuse, nous introduisons d'abord un cadre mathématique qui: (1) catégorise tous les algorithmes impliqués en trois niveaux (niveaux alpha, beta et gamma); (2) définit concrètement le concept de tâche (en particulier dans un cadre d'apprentissage supervisé); (3) définit formellement HPO et méta-apprentissage; (4) introduit une métrique d'any-time learning qui permet d'évaluer les algorithmes d'apprentissage non seulement par leur précision, mais également par leur vitesse d'apprentissage. Ce cadre mathématique unifie différents sous-domaines du ML, nous permet de classer systématiquement les méthodes et nous fournit des outils formels pour faciliter les développements théoriques et de futures recherches empiriques. En particulier, il sert de base théorique à une série de challenges que nous avons organisés. En effet, notre principale approche méthodologique pour aborder AutoML avec Deep Learning a été de mettre en place un benchmark étendu, dans le cadre d'une série de challenges sur l'Automated Deep Learning (AutoDL). Ces challenges fournissent une suite de référence de solutions AutoML de base avec un référentiel d'environ 100 datasets, dont plus de la moitié sont publiés sous forme de datasets publics pour permettre la recherche sur le méta-apprentissage. À la fin de ces challenges, nous avons effectué des analyses post-challenge approfondies qui ont révélé que : (1) les solutions gagnantes se généralisent à de nouveaux datasets invisibles, ce qui valide les progrès vers la solution universelle AutoML ; (2) Malgré nos efforts pour encourager des solutions génériques, les participants ont adopté des flux de travail spécifiques pour chaque modalité ; (3) L'any-time learning a été abordé avec succès, sans sacrifier la performance finale ; (4) Bien que certaines solutions se soient améliorées par rapport à la baseline fournie, elles en ont fortement influencé plusieurs ; (5) Les solutions d'apprentissage en profondeur dominaient, mais la recherche d'architecture neuronale n'était pas pratique dans les délais impartis ; (6) Les études d'ablation ont révélé l'importance du méta-apprentissage, de l'assemblage et du chargement efficace des données, tandis que l'augmentation des données n'est pas critique. Tous les codes et données sont disponibles sur autodl.chalearn.org. Outre l'introduction d'une nouvelle formulation générale du problème AutoML, la mise en place et l'analyse du challenge AutoDL, les contributions de cette thèse comprennent : (1) Développer nos propres solutions aux problèmes que nous avons posés aux participants. Notre travail GramNAS s'attaque au problème de la recherche d'architecture neuronale (NAS) en utilisant une grammaire formelle pour encoder les architectures neuronales. Deux stratégies de recherche alternatives ont été étudiées expérimentalement: une basée sur Monte-Carlo Tree Search (MCTS), qui atteint une précision de 94% sur le dataset CIFAR-10, et une autre basée sur un algorithme évolutif qui bat les packages de pointe AutoGluon et AutoPytorch sur 4 grands datasets bien connus ; (2) Former la base d'un futur challenge sur le méta-apprentissage ; (3) Apporter plusieurs contributions théoriques. Au cours de cette thèse, plusieurs collaborations ont été engagées pour aborder les problèmes de transfer learning et d'expressivité des réseaux de neurones. Les enquêtes sur le théorème d'approximation universelle nous ont aidés à comprendre la garantie théorique derrière les systèmes d'apprentissage profond que nous déployons.

1 February 2022 in arXiv
Verified
Preprint
Open access
PDF file
ABAdrian El Baz
Adrian El Baz
IGIsabelle Guyon
Isabelle Guyon
ZLZhengying Liu
Zhengying Liu
JRJan van Rijn
Jan van Rijn
STSebastien Treguer
Sebastien Treguer
JVJoaquin Vanschoren
Joaquin Vanschoren
To stimulate advances in metalearning using deep learning techniques (MetaDL), we organized in 2021 a challenge and an associated workshop. This paper presents the design of the challenge and its results, and summarizes presentations made at the workshop. The challenge focused on few-shot learning classification tasks of small images. Participants' code submissions were run in a uniform manner, under tight computational constraints. This put pressure on solution designs to use existing architecture backbones and/or pre-trained networks. Winning methods featured various classifiers trained on top of the second last layer of popular CNN backbones, fined-tuned on the meta-training data (not necessarily in an episodic manner), then trained on the labeled support and tested on the unlabeled query sets of the meta-test data.

21 October 2020 in HAL
Verified
Preprint
Open access
PDF file
RERomain Egele
Romain Egele
PBPrasanna Balaprakash
Prasanna Balaprakash
VVVenkatram Vishwanath
Venkatram Vishwanath
IGIsabelle Guyon
Isabelle Guyon
ZLZhengying Liu
Zhengying Liu
MAMsc&T Aivic
Msc&T Aivic
Developing high-performing predictive models for large tabular data sets is a challenging task. The state-of-the-art methods are based on expert-developed model ensembles from different supervised learning methods. Recently, automated machine learning (AutoML) is emerging as a promising approach to automate predictive model development. Neural architecture search (NAS) is an AutoML approach that generates and evaluates multiple neural network architectures concurrently and improves the accuracy of the generated models iteratively. A key issue in NAS, particularly for large data sets, is the large computation time required to evaluate each generated architecture. While data-parallel training is a promising approach that can address this issue, its use within NAS is difficult. For different data sets, the data-parallel training settings such as the number of parallel processes, learning rate, and batch size need to be adapted to achieve high accuracy and reduction in training time. To that end, we have developed AgEBO-Tabular, an approach to combine aging evolution (AgE), a parallel NAS method that searches over neural architecture space, and an asynchronous Bayesian optimization method for tuning the hyperparameters of the data-parallel training simultaneously. We demonstrate the efficacy of the proposed method to generate high-performing neural network models for large tabular benchmark data sets. Furthermore, we demonstrate that the automatically discovered neural network models using our method outperform the state-of-the-art AutoML ensemble models in inference speed by two orders of magnitude while reaching similar accuracy values.