Mindtech Technologies

Mindtech have developed unique and sophisticated methodologies to create data that is closely domain matched

Mindtech are developing the Chameleon Platform, focused on providing the solutions required to create the massive amounts of annotated data required to train Neural Networks.  The use of Synthetic Data gives several key advantages for the implementation of AI systems.

Data - The New Oil

Artificial Intelligence, Machine learning, deep learning, neural networks… by whatever name you know the technology, they have revolutionized the field of image understanding  and recognition. From the victory of the Alexnet neural network in the 2012 Imagenet Large Scale Visual Recognition Challenge (ILSVRC), to the  2019 introduction of Tesla’s FSD SoC into production vehicles, with 72 TOPs of AI compute performance these deep learning solutions have proved their value.  

Neural networks consist of two building blocks: The network itself, consisting of a number of different interconnected layers performing various functions and the data to train this network.  The two are inseparable;  without both elements, the network will not function.

The most useful data to train networks, is annotated data, that is data which is labeled to describe to the machine learning system exactly what the data contains.  These annotations may vary from simple 2D spatial to complex time and 3D space-dependent labels.

 

Synthetic or "Real"?

The answer is actually and  not  or.  In most use cases, the final target for the inference system (the real system that is using the AI implementation), will be based on looking at the natural world.  Therefore, as a minimum, we would expect the AI system to be tested using real-world data.

Our platform is designed to work with both real and synthetic data, enabling the best results to be achieved.

Why synthetic data?

Bias Reduction

Data can be created to specifically target bias in the real data set

GDPR Free

Datasets can be stored and used without privacy concerns

Pixel Perfect Annotation

Data can be perfectly annotated, including advanced annotations such as 3D-Bounding boxes, velocity vectors and distance data

Multi-sensor Data

Synthetic data can be generated to represent multiple different sensor types, simultaneously, RGB data, Bayer data, LIDAR data and thermal data are a few examples

Inference System Modeling

The synthetic data allows you to have inference system independent data, and then apply the relevant biases on a case by case basis

Corner case modeling

Dangerous, difficult and time consuming corner test cases can be modeled with the synthetic data.