The structure of artificial intelligence systems should be very different depending on the tasks being solved. To simplify the creation of complex AI systems for prediction and classification, Scorch.AI develop a software solution - Scorch Computation Framework.
The Framework performs 2 important functions:Combination: combines various machine learning modules, as well as auxiliary modules into a single ecosystem.
Integration: provides integration of input and output data with business formats.
You can use Computation Framework for different purposes:
- Define layers and data flows
- Identify data sources
- Divide layers into units, each unit can receive data from another unit or from a previous layer
- Combine neural networks, ML classifiers and other machine learning algorithms, as well as customizable filters and networks with predefined weights
- Train each unit separately according to its methodology
- Use memory to store temporary data
- Abstract from the data format
- Receive data via asynchronous requests (HTTP RESTFull) from client applications, defining the client application as one of the Framework layers
- Transmit decisions and predictions data via asynchronous requests (HTTP RESTFul)
- Implement any AI prediction, classification (or regression) task
The Framework can solve complex classification problems when it is necessary to involve a lot of algorithms (and not just a neural network) and data sources. Data sources could be video, sound, data from motion sensors, radar. Analysis algorithms could be neural networks, contour analysis, various mathematical filters and decompositions (for example, sound can be decomposed into Fourier transforms), ML classifiers (probabilistic, vector and others). Moreover, the Computation Framework provides additional functionality for working with memory and various configuration tools, such as sum of weights configuration modules from different units.
Combination function of the Computation Framework is realized in this way:
Framework consists of several layers. The layers are arranged in the data processing order, the data sources are located at the bottom level, and the output data adaptors after the highest layer.
Layers contain computation units that have the implementation of one or many data processing algorithms and make decisions. It could be:
- Neuro Net ( CNN, RNN, RBM ... )
- Memory Augment Net or Time Series Networks
- Mashine Learning Classifier (SVM,RF,SOM ..)
- Data Endoder, Decoder or Transformer
- Genetic Selection Algorithm
- Other Data Processing Algo
Each algorithm can be trained or tuned separately with its own train data sets. This approach speeds up the training of the system as a whole.
Integration function of the Computation Framework provides the following options:
- Multiple data sources and various input formats
- Data analysis throughmultiple layers by various instruments (neural networks, mat filters and signal processing, etc.)
- Additional functions for temporal memory
- Unification of formats and protocols for transferring knowledge and data from layer to layer from unit to unit
- Additional options for configuring layers and transferring information (predefined weights and confidence factors)
- Multiple output points with different formats suitable for business
Using Computation Framework to create systems and services with artificial intelligence can save time by:
- Deploying pre-engineered modules into the overall ecosystem
- Implementation of machine learning modules using machine learning units and utility units from the Framework predefined units set
- Integration with input and output data of external systems
- Configuration of unit-modules and data flows between them
- Implementation of management tools