Data-centric AI

In the context of the implementation of »meta-learning strategies« among others in the »field of imaging medicine« methods like Few-Shot or Transfer-Learning are developed and researched.

Especially in the field of tissue classification in medicine, the annotation of large data sets is particularly difficult and the data situation often varies from hospital to hospital or from imaging device to imaging device. Therefore, methods of Few-Shot Learning, such as Prototypical Network, are suitable here to generalize between different applications. Transfer learning methods, whereby models are pre-trained on comparable data sets with many annotated data points, in order to then apply them to the actual problem, are also used in this context for the practical evaluation of medical CTs.

Overall, the application »Robust AI for Digital Pathology« uses methods of Few-Shot Learning for interactive tissue classification. Due to the possibility to interact with the models, possible new tissue classes can also be considered.

In the application »AI Framework for Autonomous Systems«, meta-learning strategies for autonomous driving are being researched. Reinforcement learning models are pre-trained in a simulation environment and then adapted for real-world application via transfer learning. In another part of this project, continuous learning is used to train models that can quickly and flexibly adapt to emerging scenarios in autonomous driving.

Semi-supervised learning methods are applied in scenarios where a lot of training data is available, but only a minority of it is annotated. These methods also incorporate latent information from the non-annotated data, such as similarities, into the model training in order to train high-performance machine learning models.

In a subproject, methods from the field of »Consistency Regularization« are researched and developed for the application on sequential sensor data. Furthermore, semi-supervised learning strategies are used in a project in the area of camera-based automated garbage sorting to circumvent the problem of less annotated training data. 

The annotation of training data is technically challenging and time-consuming in the area of localization. Becaue of that it is important to develop the right methods to support data mining.
In the application »Data-Driven Localization«, a measurement platform is developed for efficient data annotation using Active Learning and incorporating Predictive Uncertainty in order to suggest measurement points to the user that are expected to yield the greatest information gain.
On the other hand, the generation of the measurement or training data during data-driven localization also requires the inclusion of the statistical distribution or non-linearities of the signal propagation in the environment. Therefore, already during the data annotation phase, attention is paid to a non-uniform distribution of the class sets.
By methods like »SMOTE«, where either training data of the overrepresented class are discarded or additional data from the underrepresented class are augmented, such effects should be compensated.

Digitization and the steady development in artificial intelligence (AI) research, especially in the field of machine learning (ML), are currently benefiting many companies by enabling them to develop new data-driven business models or reduce process costs (e.g. in manufacturing). An important building block for this AI development is the availability of a large amount of high-quality data.

However, the data is not always available in sufficient quantity, complete, error-free or up-to-date, so that it cannot really be used meaningfully for ML applications. Through methods of data augmentation (DA for short) it is possible to significantly improve data quality and quantity. This allows ML models to be used for the first time in special use cases furthermore the results of existing ML models to be optimized.

Few Data Learning is used in application areas where a very small database is available: for example, in the field of image recognition, especially in medical technology for the diagnosis of tissue anomalies, for computer vision applications in image and video production, or for forecasting and optimization applications in production and logistics.

Existing Few Data Learning methods have been developed for very specific data problems and have different objectives. Therefore, the challenge in research and application is to select, combine and further develop the right Few Data Learning methods for a specific use case.

Within the ADA Lovelace Center, the work of the Few Data Learning competency pillar is closely linked to the Few Labels Learning pillar, which focuses on the annotation of large data sets. This is because in practice, the two problems often occur together: When data is missing, incorrect, or not available in adequate quantity, the corresponding annotations are often missing as well. Therefore, methods from the two competence pillars »Few Data Learning« and »Few labels Learning« are often combined.

If a time series (here demand in units for a product) is to be extended with a few data points for ML modeling, a search can be made for similar time series that already exist. Subsequently, these data points are taken over directly to extend the short time series. The search for similar data is done, for example, using clustering methods.

If such a direct assignment to similar time series is not possible, difficult or not desired, a synthetic data series can be generated from several available time series instead, which represents the information of the many data series. In this case, redundancies in a high-dimensional data set are used in order to generate a low-dimensional representation of the entire data set.

In some applications, there may even be no data at all for analysis. In this case, simulation models can be used to generate completely synthetic data. For example in production: If an existing production is to be converted to a completely new product, there are no empirical values or data for this yet. In this case, a simulation environment can be created based on the existing data and process modeling, which simulates data for the new products and then analyzes them using ML models. It is particularly important, but also difficult, to simulate data that is as realistic as possible, which does not overestimate or underestimate certain characteristics (e.g. production errors or machine downtimes), so that the ML model can also deliver good results with real data during operation.

In the application for Robust AI for Digital Pathology, AI methods are being developed to automatically detect colorectal cancer on CT tissue scans. In this process, often only very little data is available. Therefore, within the former competence pillar Few Data Learning, different methods of data augmentation were compared using a multi-scanner database. The goal is to ensure robust tissue analysis in intestinal sections of adenocarcinomas. Different convolutional network architectures were compared with regard to their execution speed and robustness on a multi-scanner database. Robustness is achieved by applying data augmentation, especially color augmentation.

In the application AI-based diagnoses in wireless systems a toolchain for automated detection and prediction of transmission faults in wireless networks is developed. In doing so, a contribution is made to the improvement of fault analysis tools in wireless networks, using spectrum analyzers. Using machine learning based image processing algorithms, individual frames of different wireless technologies as well as interference collisions between frames are detected in real time and classified according to their communication standard. To improve the database, the wireless signals generated with a vector signal generator were further processed and recombined in a specially developed simulation pipeline. Using this approach, an extensive labeled training data set could be accessed.

In the application Self-optimization in adaptive logistics networks, a method was developed that uses clustering to detect similarities in a large data set based on incomplete consumption data of spare parts in logistics and uses available consumption data over a longer time horizon to forecast new spare parts (without a long data history).