Rapid Operational Deployment for Deep Learning
Deep Learning is, without doubt, the hottest topic in machine learning and data science. It has fueled our imagination on what “artificial intelligence” will be able to achieve in the near future, from self-driving cars to personal digital assistants, medical diagnosis and robotics.
Deep learning commonly refers to a class of neural network algorithms that use many layers and learn from vast amounts of training data. In just a few years, solutions based on deep learning algorithms have broken many records, e.g., in image processing and speech recognition. While the fundamental concepts behind deep learning algorithms have been known for some time, recent advances in high-performance computing and “big data” now put deep learning models in reach for many organizations.
Today, we already have a choice of various deep learning toolkits to train models, e.g., Caffe, Theano, Tensor Flow, DeepLearning4j, Torch, to mention just a few of the popular packages. Until now, however, using such models as part of operational IT systems and every-day business processes has been a huge challenge and often required extensive custom coding.
No more! Zementis offers a standards-based platform to deploy deep learning models side-by side with traditional data mining algorithms and models build in R, Python, SAS, IBM SPSS and many others.
Zementis solutions facilitate the rapid operational deployment and integration of deep learning models.
The video here illustrates ADAPA executing two deep convolutional neural networks based on images received from a mobile phone over the network. The first example is processing handwritten characters, the second example recognizes objects in images. It showcases real-time scoring (no GPU needed) and flexible model management with a scalable, cloud-based ADAPA server.
Immediate Benefits of Using ADAPA or UPPI
Once a model has been built in a deep learning library and saved as a PMML file, it can be directly deployed in ADAPA or in UPPI for scoring in whatever production environment the business utilizes. With ADAPA and UPPI, you can:
- Execute deep learning models on any target platform
- Support a wide range of deep learning model types
- Deliver real-time scoring for high-volume transactions
- Deploy deep learning models in minutes or hours, not months
- Ensure consistent model performance in one common operational platform
- Eliminate the need for custom code or GPU acceleration
- Remove dependencies on the development environment
Deep Learning PMML Support
Zementis offers support for deep learning models through a PMML converter that takes your models and converts them automatically into a representation based on the PMML industry standard. This allows for great flexibility to export models from different deep learning packages into PMML.
- DeepPMML Converter: Please contact Zementis to learn more about the converter, supported libraries and model types as well as our deep learning expertise.
A Common Industry Standard
PMML allows for the decoupling of two very important data mining tasks: model development and operational model deployment. With PMML, data scientists can focus on data analysis and model building using best-of-breed model development tools, whereas ADAPA and UPPI make operational deployment and actual use of the model extremely fast and simple.
For example, if a data scientist develops a deep convolutional neural network model using Caffe, effectively deploying the model operationally only requires saving it as a PMML file using the Zementis DeepPMML converter and then deploying it either in ADAPA or UPPI for scoring.
Once deployed, the model is available for all to use. ADAPA makes models available for scoring on-premise through ADAPA On-site or in the cloud through ADAPA in the Cloud. UPPI makes models available for scoring in Hadoop.
PMML allows for the model development environment to be used just for that, model development. Scoring, in real-time or for batch jobs, is handled by ADAPA or UPPI.
To speak with a Zementis representative and get more information about our predictive analytics solutions, please contact us today.