Whitebox Machine Learning
- All the IP is exclusively our own.
- Unlike most other AI companies, we have no dependency on Google, OpenAI, Facebook, or any kind of pre-trained open-source models.
- Complete transparency of learned results
- Compatibility with human-generated models
- Extensibility: trained and human generated models can be combined with no loss of uncertainty information.
- Compliance: Our technology is considered to be compliant with proposed EU AI regulations.
Machine learned models you can understand
Much of the recent attention in Machine Learning and AI has focused on Deep Learning. While this is very effective for certain kinds of data, the models created are "black boxes", i.e. you cannot immediately understand what has been learned.
This is a very old subject in Machine Learning. Our solution some years ago was to create a Fuzzy Logic Rule Induction algorithm, which has been improved and refined ever since.
From this we have created a whole set of algorithms to handle various kinds of learning.
We have a Genetic Programming algorithm that handles reinforcement learning, otherwise known as "learning with a critic" which can be used wherever some part of a system needs to be optimized. We also have an Unsupervised learning algorithm and even an Association learning algorithm. All of these create DARL code as the generated model.
As you can see elsewhere on this site, DARL code is very similar to English language type rules, and so very easy to understand.
It's important to see what your Machine learning algorithms have learned.
Imagine, for instance, that you have created an insurance risk model for home owners that accidentaly rejects households with two adults of the same sex....
It's important to understand, and be able to test the effect of removing any implicit biases of this kind.
This is something you can only do with Whitebox machine learning models like DARL.
Whitebox learning examples
A few examples are shown below of supervised machine learning to fuzzy logic rules.
The initial DARL code contains a skeleton ruleset that just defines the data values to mine.
Training the model injects the fuzzy rules that the ML algorithm has generated into the rule set.
Select a model, view the training data, train the model and see how the DARL code changes.