The Unreasonable Effectiveness Of Early Discarding After One Epoch In Neural Network Hyperparameter Optimization

To reach high performance with deep learning, hyperparameter optimization (HPO) is essential. This process is usually time-consuming due to costly evaluations of neural networks. Early discarding techniques limit the resources granted to unpromising candidates by observing the empirical learning curves and canceling neural network training as soon as the lack of competitiveness of a candidate becomes evident. Despite two decades of research, little is understood about the trade-off between the aggressiveness of discarding and the loss of predictive performance. Our paper studies this trade-off for several commonly used discarding techniques such as successive halving and learning curve extrapolation. Our surprising finding is that these commonly used techniques offer minimal to no added value compared to the simple strategy of discarding after a constant number of epochs of training. The chosen number of epochs depends mostly on the available compute budget. We call this approach i-Epoch (i being the constant number of epochs with which neural networks are trained) and suggest to assess the quality of early discarding techniques by comparing how their Pareto-Front (in consumed training epochs and predictive performance) complement the Pareto-Front of i-Epoch.

RRR-Net: Reusing, Reducing, and Recycling a Deep Backbone Network

It has become mainstream in computer vision and other machine learning domains to reuse backbone networks pre-trained on large datasets as preprocessors. Typically, the last layer is replaced by a shallow learning machine of sorts; the newly-added classification head and (optionally) deeper layers are fine-tuned on a new task. Due to its strong performance and simplicity, a common pre-trained backbone network is ResNet152.However, ResNet152 is relatively large and induces inference latency. In many cases, a compact and efficient backbone with similar performance would be preferable over a larger, slower one. This paper investigates techniques to reuse a pre-trained backbone with the objective of creating a smaller and faster model. Starting from a large ResNet152 backbone pre-trained on ImageNet, we first reduce it from 51 blocks to 5 blocks, reducing its number of parameters and FLOPs by more than 6 times, without significant performance degradation. Then, we split the model after 3 blocks into several branches, while preserving the same number of parameters and FLOPs, to create an ensemble of sub-networks to improve performance. Our experiments on a large benchmark of $40$ image classification datasets from various domains suggest that our techniques match the performance (if not better) of ``classical backbone fine-tuning'' while achieving a smaller model size and faster inference speed.

Meta-Album: Multi-domain Meta-Dataset for Few-Shot Image Classification

We introduce Meta-Album, an image classification meta-dataset designed to facilitate few-shot learning, transfer learning, meta-learning, among other tasks. It includes 40 open datasets, each having at least 20 classes with 40 examples per class, with verified licences. They stem from diverse domains, such as ecology (fauna and flora), manufacturing (textures, vehicles), human actions, and optical character recognition, featuring various image scales (microscopic, human scales, remote sensing). All datasets are preprocessed, annotated, and formatted uniformly, and come in 3 versions (Micro $\subset$ Mini $\subset$ Extended) to match users' computational resources. We showcase the utility of the first 30 datasets on few-shot learning problems. The other 10 will be released shortly after. Meta-Album is already more diverse and larger (in number of datasets) than similar efforts, and we are committed to keep enlarging it via a series of competitions. As competitions terminate, their test data are released, thus creating a rolling benchmark, available through OpenML.org. Our website https://meta-album.github.io/ contains the source code of challenge winning methods, baseline methods, data loaders, and instructions for contributing either new datasets or algorithms to our expandable meta-dataset.

PyExperimenter: Easily distribute experiments and track results

PyExperimenter is a tool to facilitate the setup, documentation, execution, and subsequent evaluation of results from an empirical study of algorithms and in particular is designed to reduce the involved manual effort significantly. It is intended to be used by researchers in the field of artificial intelligence, but is not limited to those.

Lessons learned from the NeurIPS 2021 MetaDL challenge: Backbone fine-tuning without episodic meta-learning dominates for few-shot learning image classification

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.

Learning Curves for Decision Making in Supervised Machine Learning -- A Survey

Learning curves are a concept from social sciences that has been adopted in the context of machine learning to assess the performance of a learning algorithm with respect to a certain resource, e.g. the number of training examples or the number of training iterations. Learning curves have important applications in several contexts of machine learning, most importantly for the context of data acquisition, early stopping of model training and model selection. For example, by modelling the learning curves, one can assess at an early stage whether the algorithm and hyperparameter configuration have the potential to be a suitable choice, often speeding up the algorithm selection process. A variety of approaches has been proposed to use learning curves for decision making. Some models answer the binary decision question of whether a certain algorithm at a certain budget will outperform a certain reference performance, whereas more complex models predict the entire learning curve of an algorithm. We contribute a framework that categorizes learning curve approaches using three criteria: the decision situation that they address, the intrinsic learning curve question that they answer and the type of resources that they use. We survey papers from literature and classify them into this framework.

Naive Automated Machine Learning

An essential task of Automated Machine Learning (AutoML) is the problem of automatically finding the pipeline with the best generalization performance on a given dataset. This problem has been addressed with sophisticated black-box optimization techniques such as Bayesian Optimization, Grammar-Based Genetic Algorithms, and tree search algorithms. Most of the current approaches are motivated by the assumption that optimizing the components of a pipeline in isolation may yield sub-optimal results. We present Naive AutoML, an approach that does precisely this: It optimizes the different algorithms of a pre-defined pipeline scheme in isolation. The finally returned pipeline is obtained by just taking the best algorithm of each slot. The isolated optimization leads to substantially reduced search spaces, and, surprisingly, this approach yields comparable and sometimes even better performance than current state-of-the-art optimizers.

Fast and Informative Model Selection using Learning Curve Cross-Validation

Common cross-validation (CV) methods like k-fold cross-validation or Monte-Carlo cross-validation estimate the predictive performance of a learner by repeatedly training it on a large portion of the given data and testing on the remaining data. These techniques have two major drawbacks. First, they can be unnecessarily slow on large datasets. Second, beyond an estimation of the final performance, they give almost no insights into the learning process of the validated algorithm. In this paper, we present a new approach for validation based on learning curves (LCCV). Instead of creating train-test splits with a large portion of training data, LCCV iteratively increases the number of instances used for training. In the context of model selection, it discards models that are very unlikely to become competitive. We run a large scale experiment on the 67 datasets from the AutoML benchmark and empirically show that in over 90% of the cases using LCCV leads to similar performance (at most 1.5% difference) as using 5/10-fold CV. However, it yields substantial runtime reductions of over 20% on average. Additionally, it provides important insights, which for example allow assessing the benefits of acquiring more data. These results are orthogonal to other advances in the field of AutoML.

Towards Green Automated Machine Learning: Status Quo and Future Directions

Automated machine learning (AutoML) strives for the automatic configuration of machine learning algorithms and their composition into an overall (software) solution - a machine learning pipeline - tailored to the learning task (dataset) at hand. Over the last decade, AutoML has become a hot research topic with hundreds of contributions. While AutoML offers many prospects, it is also known to be quite resource-intensive, which is one of its major points of criticism. The primary cause for a high resource consumption is that many approaches rely on the (costly) evaluation of many ML pipelines while searching for good candidates. This problem is amplified in the context of research on AutoML methods, due to large scale experiments conducted with many datasets and approaches, each of them being run with several repetitions to rule out random effects. In the spirit of recent work on Green AI, this paper is written in an attempt to raise the awareness of AutoML researchers for the problem and to elaborate on possible remedies. To this end, we identify four categories of actions the community may take towards more sustainable research on AutoML, namely approach design, benchmarking, research incentives, and transparency.

Automated Machine Learning, Bounded Rationality, and Rational Metareasoning

The notion of bounded rationality originated from the insight that perfectly rational behavior cannot be realized by agents with limited cognitive or computational resources. Research on bounded rationality, mainly initiated by Herbert Simon, has a longstanding tradition in economics and the social sciences, but also plays a major role in modern AI and intelligent agent design. Taking actions under bounded resources requires an agent to reflect on how to use these resources in an optimal way - hence, to reason and make decisions on a meta-level. In this paper, we will look at automated machine learning (AutoML) and related problems from the perspective of bounded rationality, essentially viewing an AutoML tool as an agent that has to train a model on a given set of data, and the search for a good way of doing so (a suitable "ML pipeline") as deliberation on a meta-level.