AI models need for rigorous and deliberate controls have a wide range of benefits in the general and applied domains. Some benefits include increased transparency (e.g. cost, speed, accuracy, etc.) and predictability for a range of events, especially for less routine tasks, and improve interoperability between competing AI tools.
However, we also know that AI models often do not deliver optimal performance and there is limited expertise to check on their robustness and reproducibility. This means that real-world problems are difficult to test and their results can sometimes differ from published results.
In this blog post, I present a set of techniques for the rigorous and deliberate control of AI systems. These techniques enable companies and developers to ensure AI models work as they are intended.
Machine learning models: More than specific code and algorithms
In a previous post, I argued that machine learning models have an orientation to AI that is very different from the physical hardware that delivers most of the applications in today's enterprise. We also noted that this orientation is highly agnostic of the code that executes on the machine learning models.
The above post prompted a discussion on Twitter, and several people raised the idea that there may be another, less tangible feature of machine learning models. Specifically, some suggested that machine learning models may have the following:
Formal definitions
Information about context of the data
Distribution of data
Choices of what should be learned and what should be set as model parameters
Formal definitions:
Conceptually, machine learning models are very similar to algorithms, but they differ in the way they are designed. With algorithms, the construction is a matter of solving an undecidable problem.
Machine learning models are also often quite different from other models in a way that is not easily identified or understood. For instance, the most widely used machine learning models, such as the CNNs and many visual models, do not share the same architectures or principles.
I think that formal definitions are crucial to understanding machine learning models. We need to get clear on what these models really are, what kinds of data they have, and what kind of world they are trying to learn from. Without formal definitions, it can be difficult for people to implement machine learning models and understand the consequences of those models.
Formal definitions help us:
Know what models are and what they do
Hear their language and understand their semantics
Structure our knowledge of the model and the application domain
Define and design new models
Related resources: