As discussed on a previous post, the training og AI systems is one of the most challenging aspects of modern AI solutions. To some extend, the technologies and capabilities fro training AI systems have been evolving at a slower pace than AI frameworks themselves.
While the AI market has been exploding with new frameworks, cloud services and industry-specific solutions, the tools and platforms for training AI systems are still on very early stages. Just recently, AI industry leaders such as DeepMind or OpenAI have started efforts to try to streamline the training of AI systems.
The process of building more complete and sophisticated AI training solutions usually requires many building blocks. Some of these foundational pieces are just now being explored by the AI community. I’ve put together a few ideas that I consider could be relevant when comes to the training of AI systems.
Some Ideas of Improve the Training of AI Systems
1-General-Purpose AI Training Tools
Building general-purpose tools and frameworks for training AI systems is absolutely essential to streamline the knowledge building mechanisms for those type of solutions. By general-purpose, I am referring to tools that can be used to train different AI algorithms sometimes even on different domains.
2-Semi-Supervised Learning Models
Semi-supervised learning models seem to be a practical approach to simplify the training of AI solutions in the short term. These type of models an infer relationships between knowledge concepts by using proximity inference techniques. Tools such as Google Expander are a great example of this approach.
Most AI frameworks use similar data structures [graphs] to represent training knowledge used on AI algorithms. Leveraging these similarities to create standards that can be used across different AI platforms can drastically improve the training of AI systems.
4-General Human Behavior Training
One of the most efficient ways to train AI systems is by inferring knowledge from “observing” human experts operate a generic system. Examples of AI problems that can be solved with that type of training strategy are tasks such as playing video games, manipulating Excel spreadsheets or interacting with CAD files. Creating general-purpose AI tools that infer knowledge derived from observing experts can cover a large spectrum of the training requirements of AI systems.
Knowledge is the essential element for training AI systems. When comes to AI, more knowledge is almost always better. Created highly curated knowledge repositories that can be programmatically accessed by AI training tools and frameworks can exponentially improve the processes and algorithms for training AI systems.
6-Training Monitoring and Evaluation
Just like we humans have created many mechanisms for validating areas of knowledge, AI training solutions should provide the infrastructure and tooling to evaluate the efficiency of a specific training process. Additionally, AI platforms should constantly monitor the behavior of AI systems, identify knowledge gaps and adjust the training mechanisms accordingly. Tools for monitoring and evaluating the training of AI solutions are essential for the mainstream adoption of AI training models.