Finally, one more interesting aspect of our training infrastructure is that we use a multi-cloud setup in practice. As it was told earlier, GCP is our main vendor for training instances for cost and powerful machines availability-related reasons, while our default production infrastructure is AWS. It means that sometimes we need to combine the two: e.g. We use Flyte to orchestrate this process. Flyte is a workflow management system that allows us to define a pipeline as a DAG of tasks. It is useful for us because it allows us to define a pipeline once and run its steps on different machines with different computational resources allocation, and it also provides a nice UI for monitoring the progress of the pipeline.

Domain-adaptive pre-training (or DA-training for short), also known as post-training, aims to train a pre-trained general-purpose language model (LM) using an unlabeled corpus of a particular domain to adapt the LM so that end-tasks in the domain can give improved performances. However, existing DA-training methods are in some sense blind as they do not explicitly identify what knowledge in the LM should be preserved and what should be changed by the domain corpus. This paper shows that the existing methods are suboptimal and proposes a novel method to perform a more informed adaptation of the knowledge in the LM by (1) soft-masking the attention heads based on their importance to best preserve the general knowledge in the LM and (2) contrasting the representations of the general and the full (both general and domain knowledge) to learn an integrated representation with both general and domain-specific knowledge. Experimental results will demonstrate the effectiveness of the proposed approach.

With the increasing availability of data for Prognostics and Health Management (PHM), Deep Learning (DL) techniques are now the subject of considerable attention for this application, often achieving more accurate Remaining Useful Life (RUL) predictions. However, one of the major challenges for DL techniques resides in the difficulty of obtaining large amounts of labelled data on industrial systems. To overcome this lack of labelled data, an emerging learning technique is considered in our work: Self-Supervised Learning, a sub-category of unsupervised learning approaches. This paper aims to investigate whether pre-training DL models in a self-supervised way on unlabelled sensors data can be useful for RUL estimation with only Few-Shots Learning, i.e. with scarce labelled data. In this research, a fatigue damage prognostics problem is addressed, through the estimation of the RUL of aluminum alloy panels (typical of aerospace structures) subject to fatigue cracks from strain gauge data. Synthetic datasets composed of strain data are used allowing to extensively investigate the influence of the dataset size on the predictive performance. Results show that the self-supervised pre-trained models are able to significantly outperform the non-pre-trained models in downstream RUL prediction task, and with less computational expense, showing promising results in prognostic tasks when only limited labelled data is available.

A NOtice To AirMen (NOTAM) contains important flight route related information. To search and filter them, NOTAMs are grouped into categories called QCodes. In this paper, we develop a tool to predict, with some explanations, a Qcode for a NOTAM. We present a way to extend the interpretable binary classification using column generation proposed in Dash, Gunluk, and Wei (2018) to a multiclass text classification method. We describe the techniques used to tackle the issues related to one vs-rest classification, such as multiple outputs and class imbalances. Furthermore, we introduce some heuristics, including the use of a CP-SAT solver for the subproblems, to reduce the training time. Finally, we show that our approach compares favorably with state-of-the-art machine learning algorithms like Linear SVM and small neural networks while adding the needed interpretability component.

Building user trust in dialogue agents requires smooth and consistent dialogue exchanges. However, agents can easily lose conversational context and generate irrelevant utterances. These situations are called dialogue breakdown, where agent utterances prevent users from continuing the conversation. Building systems to detect dialogue breakdown allows agents to recover appropriately or avoid breakdown entirely. In this paper we investigate the use of semi-supervised learning methods to improve dialogue breakdown detection, including continued pre-training on the Reddit dataset and a manifold-based data augmentation method. We demonstrate the effectiveness of these methods on the Dialogue Breakdown Detection Challenge (DBDC) English shared task. Our submissions to the 2020 DBDC5 shared task place first, beating baselines and other submissions by over 12\% accuracy. In ablations on DBDC4 data from 2019, our semi-supervised learning methods improve the performance of a baseline BERT model by 2\% accuracy. These methods are applicable generally to any dialogue task and provide a simple way to improve model performance.

Data augmentation has recently seen increased interest in graph machine learning given its demonstrated ability to improve model performance and generalization by added training data. Despite this recent surge, the area is still relatively under-explored, due to the challenges brought by complex, non-Euclidean structure of graph data, which limits the direct analogizing of traditional augmentation operations on other types of image, video or text data. Our work aims to give a necessary and timely overview of existing graph data augmentation methods; notably, we present a comprehensive and systematic survey of graph data augmentation approaches, summarizing the literature in a structured manner. We first introduce three different taxonomies for categorizing graph data augmentation methods from the data, task, and learning perspectives, respectively. Next, we introduce recent advances in graph data augmentation, differentiated by their methodologies and applications. We conclude by outlining currently unsolved challenges and directions for future research. Overall, our work aims to clarify the landscape of existing literature in graph data augmentation and motivates additional work in this area, providing a helpful resource for researchers and practitioners in the broader graph machine learning domain. Additionally, we provide a continuously updated reading list at https://github.com/zhao-tong/graph-data-augmentation-papers.

This is also the official implementation for FlexMatch: boosting semi-supervised learning using curriculum pseudo labeling published at NeurIPS 2021. TorchSSL is an all-in-one toolkit based on PyTorch for semi-supervised learning (SSL). Currently, we implmented 9 popular SSL algorithms to enable fair comparison and boost the development of SSL algorithms. Besides, we implement our Curriculum Pseudo Labeling (CPL) method for Pseudo-Label (Flex-Pseudo-Label) and UDA (Flex-UDA). The results are best accuracies with standard errors.

Abstract: Contrastive learning has become a prominent ingredient in learning representations from unlabeled data. However, existing methods primarily consider pairwise relations. This paper proposes a new approach towards self-supervised contrastive learning based on Group Ordering Constraints (GroCo). Building on the recent success of differentiable sorting algorithms, group ordering constraints enforce that the distances of all positive samples (a positive group) are smaller than the distances of all negative images (a negative group); thus, enforcing positive samples to gather around an anchor. This leads to a more holistic optimization of the local neighborhoods.

Abstract: Contrastive learning has become a prominent ingredient in learning representations from unlabeled data. However, existing methods primarily consider pairwise relations. This paper proposes a new approach towards self-supervised contrastive learning based on Group Ordering Constraints (GroCo). Building on the recent success of differentiable sorting algorithms, group ordering constraints enforce that the distances of all positive samples (a positive group) are smaller than the distances of all negative images (a negative group); thus, enforcing positive samples to gather around an anchor. This leads to a more holistic optimization of the local neighborhoods.

Electroencephalogram (EEG) signals are effective tools towards seizure analysis where one of the most important challenges is accurate detection of seizure events and brain regions in which seizure happens or initiates. However, all existing machine learning-based algorithms for seizure analysis require access to the labeled seizure data while acquiring labeled data is very labor intensive, expensive, as well as clinicians dependent given the subjective nature of the visual qualitative interpretation of EEG signals. In this paper, we propose to detect seizure channels and clips in a self-supervised manner where no access to the seizure data is needed. The proposed method considers local structural and contextual information embedded in EEG graphs by employing positive and negative sub-graphs. We train our method through minimizing contrastive and generative losses. The employ of local EEG sub-graphs makes the algorithm an appropriate choice when accessing to the all EEG channels is impossible due to complications such as skull fractures. We conduct an extensive set of experiments on the largest seizure dataset and demonstrate that our proposed framework outperforms the state-of-the-art methods in the EEG-based seizure study. The proposed method is the only study that requires no access to the seizure data in its training phase, yet establishes a new state-of-the-art to the field, and outperforms all related supervised methods.