Sap Se

David Kernert

Wolfgang Lehner

Frank Koehler

Joachim Fitzer

Boris Gruschko

Simon Seif

Erick David Santillán Perez

Disclosed herein are system, method, and computer program product embodiments for adapting machine learning models for use in additional applications. For example, feature extraction models are readily available for use in applications such as image detection. These feature extraction models can be used to label inputs (such as images) in conjunction with other deep neural network models. However, in adapting the feature extraction models to these uses, it becomes problematic to improve the quality of their results on target data sets, as these feature extraction models are large and resistant to retraining. Approaches disclosed herein include a transfer layer for providing fast retraining of machine learning models.

Domain adaptation of deep neural networks

A method for executing a data processing pipeline may be provided. The method may include identifying a file providing a runtime environment required for executing a series of data processing operations comprising the data processing pipeline. The file may be identified based on one or more tags associated with the data processing pipeline. The one or more tags may specify at least one runtime requirement for the series of data processing operations. The file may be executed to generate an executable package that includes a plurality of components required for executing the series of data processing operations. The series of data processing operations included in the data processing pipeline may be executed by at least executing the executable package to provide the runtime environment required for executing the series of data processing operations. Related systems and articles of manufacture, including computer program products, are also provided.

Containerized runtime environments

Embodiments relate to storing sparse matrices in an in-memory column-oriented database system. Specifically, recent hardware shifts of primary storage from disc into memory, allow execution of linear algebra queries directly in the database engine. Dynamic matrix manipulation operations (like online insertion or deletion of elements) are not covered by most linear algebra frameworks. Therefore a hybrid architecture comprises a read-optimized main structure, and a write-optimized delta structure. The resulting system layout derived from the Compressed Sparse Row (CSR) representation, integrates well with a columnar database design. Moreover, the resulting architecture is amenable to a wide range of non-numerical use cases when dictionary encoding is used. Performance in specific examples is evaluated for dynamic sparse matrix workloads, by applying work flows of nuclear science and network graphs. Embodiments allow performing linear algebra operations on large, sparse matrices commonly associated with scientific computations and analytical business applications.

Sparse linear algebra in column-oriented in-memory database

According to some embodiments, matrix A data may be loaded into a temporary, unordered starting representation that contains coordinates and values for each element of matrix A. Z-curve ordering of matrix A may be performed to create a two-dimensional density map of matrix A by counting matrix elements that are contained in logical two-dimensional block cells of a given size. A quad-tree recursion may be executed on the two-dimensional density map structure in reduced Z-space to identify areas of different densities in the two dimensional matrix space. An adaptive tile matrix representation of input matrix A may then be created. According to some embodiments, an adaptive tile matrix multiplication operation may perform dynamic tile-granular optimization based on density estimates and a cost model.

Growing community of inventors

Heidelberg, Germany

David Kernert