Machine learning very simple and yet state of the art

DIM-GP

Deep Infinite Mixture of Gaussian Processes (DIM-GP) is PI’s self-developed machine learning algorithm. A unique combination of neural networks (deep learning) and Gaussian processes is used. With it, a large number of problems can be dealt with, without expert knowledge and without time-consuming setting of hyperparameters! All our products are written in Python and are therefore very easy to integrate into existing software platforms and workflows.

dim-gp

(Image-)regression

Classic regression tasks (continuous output variables) can be processed quickly and easily with DIM-GP. As a special feature, it is also possible to perform image regression. Typically, images can only be used as input information in the event of classification problems. However, DIM-GP also allows continuous quantities to be predicted on the basis of image information.

Regression

dim-gp

Klassifizierung

(Image-)classification

Classification tasks (categorical output variables) are also not a problem for DIM-GP. Any number of categories can be learned within a model. As with all predictions from DIM-GP, you always get an associated probability. As with convolutional neural networks (CNN), DIM-GP can use images as input information to carry out the classification.

dim-gp

Multi-Output

In the case of multi-output questions, there are two or more output variables that correlate with one another. Often, however, individual models are trained for these output variables, which cannot also map the correlations. DIM-GP is able to learn these correlations within a single model. This can be used to map signals or field sizes, for example. Thanks to the learned correlations, the predictions are much more accurate. An example from mechanical engineering would be a Live Finite Element Analysis or Computational Fluid Dynamics (CFD) investigation. DIM-GP is able to predict each node element result.
multi output

dim-gp

Konfidenz

Confidence intervals of the predictions

Since DIM-GP is a probabilistic model, the uncertainty of this prediction can always be specified in addition to a prediction. This allows confidence areas to be displayed over the entire design space shown. This means you don’t have to blindly trust the model prediction, but always know how certain the model is.

dim-gp

Automatic noise detection and avoidance of outliers

Since DIM-GP is a probabilistic model, the uncertainty of this forecast can always be given in addition to a forecast. This allows confidence areas to be displayed across the entire depicted design space. That means you don’t have to blindly trust the model prediction, you always know how certain the model is.

Ausreißer

dim-gp

Automatic noise detection and avoidance of outliers

Of course, data is not always perfect. Often there are outliers or noise in the data. If possible, this should not be included in the model. Many models can be very strongly influenced by a few outliers and the user has to clean up his data beforehand in order to avoid an unwanted influence.

DIM-GP is able to identify and avoid outliers independently and fully automatically. The same is true for noise in the data. As a special feature, DIM-GP can even recognize and map different levels of noise. In this way, data of different quality can also be merged in one model.
Sequentially dependent output variables

dim-gp

BIG DATA READY

Like neural networks, DIM-GP can divide the training data into packets of any size, so that DIM-GP can be used successfully on any hardware but also with any large amount of data. Despite the use of Gaussian processes for the prediction, its scalability is comparable to neural networks. In addition to CPUs, DIM-GP can also be trained on GPUs. This allows an even faster training. It is also possible to carry out the training distributed over several computing cores up to a cluster – for really demanding and time-consuming questions.

dim-gp

No adjustments by the user necessary!

Given the wide range of possible applications that DIM-GP offers, for which a large number of different algorithms would normally be required, you do not need to set any parameters when using DIM-GP. This is a huge advantage over most other methods and especially over neural networks, which have a large number of setting options.

DIM-GP can deliver high-quality models that correspond to the state of science and technology and that for a wide variety of questions without setting parameters. You save the hassle of trying out different settings and other algorithms and you can evaluate your data more quickly.