Physics-based character animation has seen significant advances in recent years with the adoption of Deep Reinforcement Learning (DRL). However, DRL-based learning methods are usually computationally expensive and their performance crucially depends on the choice of hyperparameters. Tuning hyperparameters for these methods often requires repetitive training of control policies, which is even more computationally prohibitive. In this work, we propose a novel Curriculum-based Multi-Fidelity Bayesian Optimization framework (CMFBO) for efficient hyperparameter optimization of DRL-based character control systems. Using curriculum-based task difficulty as fidelity criterion, our method improves searching efficiency by gradually pruning search space through evaluation on easier motor skill tasks. We evaluate our method on two physics-based character control tasks: character morphology optimization and hyperparameter tuning of DeepMimic. Our algorithm significantly outperforms state-of-the-art hyperparameter optimization methods applicable for physics-based character animation. In particular, we show that hyperparameters optimized through our algorithm result in at least 5x efficiency gain comparing to author-released settings in DeepMimic.
Bayesian optimization (BO) is a widely popular approach for the hyperparameter optimization (HPO) in machine learning. At its core, BO iteratively evaluates promising configurations until a user-defined budget, such as wall-clock time or number of iterations, is exhausted. While the final performance after tuning heavily depends on the provided budget, it is hard to pre-specify an optimal value in advance. In this work, we propose an effective and intuitive termination criterion for BO that automatically stops the procedure if it is sufficiently close to the global optimum. Our key insight is that the discrepancy between the true objective (predictive performance on test data) and the computable target (validation performance) suggests stopping once the suboptimality in optimizing the target is dominated by the statistical estimation error. Across an extensive range of real-world HPO problems and baselines, we show that our termination criterion achieves a better trade-off between the test performance and optimization time. Additionally, we find that overfitting may occur in the context of HPO, which is arguably an overlooked problem in the literature, and show how our termination criterion helps to mitigate this phenomenon on both small and large datasets.
Recently, the study of graph neural network (GNN) has attracted much attention and achieved promising performance in molecular property prediction. Most GNNs for molecular property prediction are proposed based on the idea of learning the representations for the nodes by aggregating the information of their neighbor nodes (e.g. atoms). Then, the representations can be passed to subsequent layers to deal with individual downstream tasks. Therefore, the architectures of GNNs can be considered as being composed of two core parts: graph-related layers and task-specific layers. Facing real-world molecular problems, the hyperparameter optimization for those layers are vital. Hyperparameter optimization (HPO) becomes expensive in this situation because evaluating candidate solutions requires massive computational resources to train and validate models. Furthermore, a larger search space often makes the HPO problems more challenging. In this research, we focus on the impact of selecting two types of GNN hyperparameters, those belonging to graph-related layers and those of task-specific layers, on the performance of GNN for molecular property prediction. In our experiments. we employed a state-of-the-art evolutionary algorithm (i.e., CMA-ES) for HPO. The results reveal that optimizing the two types of hyperparameters separately can gain the improvements on GNNs' performance, but optimising both types of hyperparameters simultaneously will lead to predominant improvements. Meanwhile, our study also further confirms the importance of HPO for GNNs in molecular property prediction problems.
Considerable research effort has been guided towards algorithmic fairness but real-world adoption of bias reduction techniques is still scarce. Existing methods are either metric- or model-specific, require access to sensitive attributes at inference time, or carry high development or deployment costs. This work explores the unfairness that emerges when optimizing ML models solely for predictive performance, and how to mitigate it with a simple and easily deployed intervention: fairness-aware hyperparameter optimization (HO). We propose and evaluate fairness-aware variants of three popular HO algorithms: Fair Random Search, Fair TPE, and Fairband. We validate our approach on a real-world bank account opening fraud case-study, as well as on three datasets from the fairness literature. Results show that, without extra training cost, it is feasible to find models with 111% mean fairness increase and just 6% decrease in performance when compared with fairness-blind HO.
Optimizing the hyperparameters and architecture of a neural network is a long yet necessary phase in the development of any new application. This consuming process can benefit from the elaboration of strategies designed to quickly discard low quality configurations and focus on more promising candidates. This work aims at enhancing HyperNOMAD, a library that adapts a direct search derivative-free optimization algorithm to tune both the architecture and the training of a neural network simultaneously, by targeting two keys steps of its execution and exploiting cheap approximations in the form of static surrogates to trigger the early stopping of the evaluation of a configuration and the ranking of pools of candidates. These additions to HyperNOMAD are shown to improve on its resources consumption without harming the quality of the proposed solutions.
Convolutional neural networks (CNNs) are widely used in image recognition. Numerous CNN models, such as LeNet, AlexNet, VGG, ResNet, and GoogLeNet, have been proposed by increasing the number of layers, to improve the performance of CNNs. However, performance deteriorates beyond a certain number of layers. Hence, hyperparameter optimisation is a more efficient way to improve CNNs. To validate this concept, a new algorithm based on simplified swarm optimisation is proposed to optimise the hyperparameters of the simplest CNN model, which is LeNet. The results of experiments conducted on the MNIST, Fashion MNIST, and Cifar10 datasets showed that the accuracy of the proposed algorithm is higher than the original LeNet model and PSO-LeNet and that it has a high potential to be extended to more complicated models, such as AlexNet.
Hyperparameter optimization is a challenging problem in developing deep neural networks. Decision of transfer layers and trainable layers is a major task for design of the transfer convolutional neural networks (CNN). Conventional transfer CNN models are usually manually designed based on intuition. In this paper, a genetic algorithm is applied to select trainable layers of the transfer model. The filter criterion is constructed by accuracy and the counts of the trainable layers. The results show that the method is competent in this task. The system will converge with a precision of 97% in the classification of Cats and Dogs datasets, in no more than 15 generations. Moreover, backward inference according the results of the genetic algorithm shows that our method can capture the gradient features in network layers, which plays a part on understanding of the transfer AI models.
Recommender Systems have shown to be an effective way to alleviate the over-choice problem and provide accurate and tailored recommendations. However, the impressive number of proposed recommendation algorithms, splitting strategies, evaluation protocols, metrics, and tasks, has made rigorous experimental evaluation particularly challenging. Puzzled and frustrated by the continuous recreation of appropriate evaluation benchmarks, experimental pipelines, hyperparameter optimization, and evaluation procedures, we have developed an exhaustive framework to address such needs. Elliot is a comprehensive recommendation framework that aims to run and reproduce an entire experimental pipeline by processing a simple configuration file. The framework loads, filters, and splits the data considering a vast set of strategies (13 splitting methods and 8 filtering approaches, from temporal training-test splitting to nested K-folds Cross-Validation). Elliot optimizes hyperparameters (51 strategies) for several recommendation algorithms (50), selects the best models, compares them with the baselines providing intra-model statistics, computes metrics (36) spanning from accuracy to beyond-accuracy, bias, and fairness, and conducts statistical analysis (Wilcoxon and Paired t-test). The aim is to provide the researchers with a tool to ease (and make them reproducible) all the experimental evaluation phases, from data reading to results collection. Elliot is available on GitHub (https://github.com/sisinflab/elliot).
Model-based Reinforcement Learning (MBRL) is a promising framework for learning control in a data-efficient manner. MBRL algorithms can be fairly complex due to the separate dynamics modeling and the subsequent planning algorithm, and as a result, they often possess tens of hyperparameters and architectural choices. For this reason, MBRL typically requires significant human expertise before it can be applied to new problems and domains. To alleviate this problem, we propose to use automatic hyperparameter optimization (HPO). We demonstrate that this problem can be tackled effectively with automated HPO, which we demonstrate to yield significantly improved performance compared to human experts. In addition, we show that tuning of several MBRL hyperparameters dynamically, i.e. during the training itself, further improves the performance compared to using static hyperparameters which are kept fixed for the whole training. Finally, our experiments provide valuable insights into the effects of several hyperparameters, such as plan horizon or learning rate and their influence on the stability of training and resulting rewards.
The sample efficiency of Bayesian optimization(BO) is often boosted by Gaussian Process(GP) surrogate models. However, on mixed variable spaces, surrogate models other than GPs are prevalent, mainly due to the lack of kernels which can model complex dependencies across different types of variables. In this paper, we propose the frequency modulated (FM) kernel flexibly modeling dependencies among different types of variables, so that BO can enjoy the further improved sample efficiency. The FM kernel uses distances on continuous variables to modulate the graph Fourier spectrum derived from discrete variables. However, the frequency modulation does not always define a kernel with the similarity measure behavior which returns higher values for pairs of more similar points. Therefore, we specify and prove conditions for FM kernels to be positive definite and to exhibit the similarity measure behavior. In experiments, we demonstrate the improved sample efficiency of GP BO using FM kernels (BO-FM).On synthetic problems and hyperparameter optimization problems, BO-FM outperforms competitors consistently. Also, the importance of the frequency modulation principle is empirically demonstrated on the same problems. On joint optimization of neural architectures and SGD hyperparameters, BO-FM outperforms competitors including Regularized evolution(RE) and BOHB. Remarkably, BO-FM performs better even than RE and BOHB using three times as many evaluations.
Modern machine learning algorithms usually involve tuning multiple (from one to thousands) hyperparameters which play a pivotal role in terms of model generalizability. Black-box optimization and gradient-based algorithms are two dominant approaches to hyperparameter optimization while they have totally distinct advantages. How to design a new hyperparameter optimization technique inheriting all benefits from both approaches is still an open problem. To address this challenging problem, in this paper, we propose a new hyperparameter optimization method with zeroth-order hyper-gradients (HOZOG). Specifically, we first exactly formulate hyperparameter optimization as an A-based constrained optimization problem, where A is a black-box optimization algorithm (such as deep neural network). Then, we use the average zeroth-order hyper-gradients to update hyperparameters. We provide the feasibility analysis of using HOZOG to achieve hyperparameter optimization. Finally, the experimental results on three representative hyperparameter (the size is from 1 to 1250) optimization tasks demonstrate the benefits of HOZOG in terms of simplicity, scalability, flexibility, effectiveness and efficiency compared with the state-of-the-art hyperparameter optimization methods.
Respiratory ailments such as asthma, chronic obstructive pulmonary disease (COPD), pneumonia, and lung cancer are life-threatening. Respiration rate (RR) is a vital indicator of the wellness of a patient. Continuous monitoring of RR can provide early indication and thereby save lives. However, a real-time continuous RR monitoring facility is only available at the intensive care unit (ICU) due to the size and cost of the equipment. Recent researches have proposed Photoplethysmogram (PPG) and/ Electrocardiogram (ECG) signals for RR estimation however, the usage of ECG is limited due to the unavailability of it in wearable devices. Due to the advent of wearable smartwatches with built-in PPG sensors, it is now being considered for continuous monitoring of RR. This paper describes a novel approach to RR estimation using machine learning (ML) models with the PPG signal features. Feature selection algorithms were used to reduce computational complexity and the chance of overfitting. The best ML model and the best feature selection algorithm combination was fine-tuned to optimize its performance using hyperparameter optimization. Gaussian Process Regression (GPR) with fitrgp feature selection algorithm outperformed all other combinations and exhibits a root mean squared error (RMSE), mean absolute error (MAE), and two-standard deviation (2SD) of 2.57, 1.91, and 5.13 breaths per minute, respectively. This ML model based RR estimation can be embedded in wearable devices for real-time continuous monitoring of the patient.
Conventional hyperparameter optimization methods are computationally intensive and hard to generalize to scenarios that require dynamically adapting hyperparameters, such as life-long learning. Here, we propose an online hyperparameter optimization algorithm that is asymptotically exact and computationally tractable, both theoretically and practically. Our framework takes advantage of the analogy between hyperparameter optimization and parameter learning in recurrent neural networks (RNNs). It adapts a well-studied family of online learning algorithms for RNNs to tune hyperparameters and network parameters simultaneously, without repeatedly rolling out iterative optimization. This procedure yields systematically better generalization performance compared to standard methods, at a fraction of wallclock time.
This paper proposes a new algorithm -- the \underline{S}ingle-timescale Do\underline{u}ble-momentum \underline{St}ochastic \underline{A}pprox\underline{i}matio\underline{n} (SUSTAIN) -- for tackling stochastic unconstrained bilevel optimization problems. We focus on bilevel problems where the lower level subproblem is strongly-convex and the upper level objective function is smooth. Unlike prior works which rely on \emph{two-timescale} or \emph{double loop} techniques, we design a stochastic momentum-assisted gradient estimator for both the upper and lower level updates. The latter allows us to control the error in the stochastic gradient updates due to inaccurate solution to both subproblems. If the upper objective function is smooth but possibly non-convex, we show that {\aname}~requires $\mathcal{O}(\epsilon^{-3/2})$ iterations (each using ${\cal O}(1)$ samples) to find an $\epsilon$-stationary solution. The $\epsilon$-stationary solution is defined as the point whose squared norm of the gradient of the outer function is less than or equal to $\epsilon$. The total number of stochastic gradient samples required for the upper and lower level objective functions matches the best-known complexity for single-level stochastic gradient algorithms. We also analyze the case when the upper level objective function is strongly-convex.
Metafeatures, or dataset characteristics, have been shown to improve the performance of hyperparameter optimization (HPO). Conventionally, metafeatures are precomputed and used to measure the similarity between datasets, leading to a better initialization of HPO models. In this paper, we propose a cross dataset surrogate model called Differentiable Metafeature-based Surrogate (DMFBS), that predicts the hyperparameter response, i.e. validation loss, of a model trained on the dataset at hand. In contrast to existing models, DMFBS i) integrates a differentiable metafeature extractor and ii) is optimized using a novel multi-task loss, linking manifold regularization with a dataset similarity measure learned via an auxiliary dataset identification meta-task, effectively enforcing the response approximation for similar datasets to be similar. We compare DMFBS against several recent models for HPO on three large meta-datasets and show that it consistently outperforms all of them with an average 10% improvement. Finally, we provide an extensive ablation study that examines the different components of our approach.
Graph neural networks (GNNs) have been proposed for a wide range of graph-related learning tasks. In particular, in recent years, an increasing number of GNN systems were applied to predict molecular properties. However, a direct impediment is to select appropriate hyperparameters to achieve satisfactory performance with lower computational cost. Meanwhile, many molecular datasets are far smaller than many other datasets in typical deep learning applications. Most hyperparameter optimization (HPO) methods have not been explored in terms of their efficiencies on such small datasets in the molecular domain. In this paper, we conducted a theoretical analysis of common and specific features for two state-of-the-art and popular algorithms for HPO: TPE and CMA-ES, and we compared them with random search (RS), which is used as a baseline. Experimental studies are carried out on several benchmarks in MoleculeNet, from different perspectives to investigate the impact of RS, TPE, and CMA-ES on HPO of GNNs for molecular property prediction. In our experiments, we concluded that RS, TPE, and CMA-ES have their individual advantages in tackling different specific molecular problems. Finally, we believe our work will motivate further research on GNN as applied to molecular machine learning problems in chemistry and materials sciences.
Recent empirical work shows that inconsistent results based on choice of hyperparameter optimization (HPO) configuration are a widespread problem in ML research. When comparing two algorithms J and K searching one subspace can yield the conclusion that J outperforms K, whereas searching another can entail the opposite. In short, the way we choose hyperparameters can deceive us. We provide a theoretical complement to this prior work, arguing that, to avoid such deception, the process of drawing conclusions from HPO should be made more rigorous. We call this process epistemic hyperparameter optimization (EHPO), and put forth a logical framework to capture its semantics and how it can lead to inconsistent conclusions about performance. Our framework enables us to prove EHPO methods that are guaranteed to be defended against deception, given bounded compute time budget t. We demonstrate our framework's utility by proving and empirically validating a defended variant of random search.
Unsupervised representation learning of words from large multilingual corpora is useful for downstream tasks such as word sense disambiguation, semantic text similarity, and information retrieval. The representation precision of log-bilinear fastText models is mostly due to their use of subword information. In previous work, the optimization of fastText's subword sizes has not been fully explored, and non-English fastText models were trained using subword sizes optimized for English and German word analogy tasks. In our work, we find the optimal subword sizes on the English, German, Czech, Italian, Spanish, French, Hindi, Turkish, and Russian word analogy tasks. We then propose a simple n-gram coverage model and we show that it predicts better-than-default subword sizes on the Spanish, French, Hindi, Turkish, and Russian word analogy tasks. We show that the optimization of fastText's subword sizes matters and results in a 14% improvement on the Czech word analogy task. We also show that expensive parameter optimization can be replaced by a simple n-gram coverage model that consistently improves the accuracy of fastText models on the word analogy tasks by up to 3% compared to the default subword sizes, and that it is within 1% accuracy of the optimal subword sizes.
Scope of Reproducibility The authors have introduced a novel method for unsupervised anomaly detection that utilises a newly introduced “Memory Module” in their paper [1]. We validate the authorsʼ claim that this helps improve performance by helping the network learn prototypical patterns, and uses the learnt memory to reduce the representation capacity of Convolutional Neural Networks. Further, we validate the efficacy of two losses introduced by the authors, “Separateness Loss” and “Compactness Loss” presented to increase the discriminative power of the memory items and the deeply learned features. We test the efficacy with the help of t-SNE plots of the memory items. Methodology The authors provide a codebase available at https://github.com/cvlab-yonsei/MNAD with scripts for the Prediction task. We reused the available code to build scripts for the Reconstruction task and variants with and without memory. The completed codebase for all tasks available at https://github.com/alchemi5t/MNADrc. Results We obtain results that are within a 3% range of the reported results for all datasets other than on the ShanghaiTech dataset[2]. The anomalous run stuck out since the ”Non Memory” of the same run could score markedly better than the ”Memory” variant. This led us to investigate the behaviour of the memory module and found valuable insights which are presented in Section 5. What was easy • The authors provide codes to run the Prediction task with memory. • The ideas presented in the paper were clear and easy to understand. • The datasets were easily available, with the exception of the ShanghaiTech dataset[2]. What was difficult • After training the model on the ShanghaiTech dataset[2], we found that the “Memory” variants of the model had issues as described in Sections 4 and 5. We investigated the behaviour of the memory module and introduced additional supervision to solve the problem. • We required hyperparameter optimization to hit the reported scores. • The paper also lacked information as to which parts of the pipeline were actually to be credited for the improvement in performance, as discussed in Section 5. Communication with original authors The codebase has scripts only for the Prediction task ”Memory” variant of the benchmarks. We reused portions of the codebase to create the other variants. We then validated our modifications with the authors. We also communicated the discrepancies we saw in parts of the results and the authors asked us to wait until they update their repository with the missing scripts.
Automated machine learning (AutoML) aims for constructing machine learning (ML) pipelines automatically. Many studies have investigated efficient methods for algorithm selection and hyperparameter optimization. However, methods for ML pipeline synthesis and optimization considering the impact of complex pipeline structures containing multiple preprocessing and classification algorithms have not been studied thoroughly. In this paper, we propose a data-centric approach based on meta-features for pipeline construction and hyperparameter optimization inspired by human behavior. By expanding the pipeline search space incrementally in combination with meta-features of intermediate data sets, we are able to prune the pipeline structure search space efficiently. Consequently, flexible and data set specific ML pipelines can be constructed. We prove the effectiveness and competitiveness of our approach on 28 data sets used in well-established AutoML benchmarks in comparison with state-of-the-art AutoML frameworks.