Several AI Research & Development Centers are working on the development of The Machine-Learning research and more specifically in the development of the Auto Machine Learning program AutoML in which algorithms can be developed that can evolve on their own without human intervention.
In addition, they can cause “mutations” in new generations of algorithms, which follow the principles of Darwin’s evolution, namely the “survival of the fittest.”
Photo by website www.bigthink.com
Machine Learning Research
Machine Learning research has advanced in multiple aspects, including model structures and learning methods. The effort to automate such research is known as AutoML and has focused on the architecture of neural networks, where it has relied on sophisticated expert-designed layers as building blocks – or similarly restrictive search spaces.
Machine learning tools allow the use of algorithms to search for bulk data and quickly identify patterns. Using mathematical equations, machine learning algorithms are developed that synthesizer 100 “prospective algorithms”.
These then compete between using basic machine learning tools, such as neural network image differentiation tests and the best-performing algorithms that then mutated or evolved through random code change.
The system can eliminate tens of thousands of algorithms every second by looking for a solution, while rejecting “evolutionary dead ends” and copies. Over many generations, the process has developed a library of high-performance algorithms.
AutoML-Zero
The program AutoML-Zero is the progress of the AutoML because today it is in position to automatically discover complete machine learning algorithms just using basic mathematical operations as building blocks. This is achieved through a generic search space that significantly reduces human bias.
Despite the enormity of this space, evolutionary search can still discover two-layer neural networks trained by backpropagation. These simple neural networks can then be exceeded by evolving directly on tasks of interest, e.g. CIFAR-10 modifications, where modern techniques emerge in the top algorithms, like bilinear interactions, normalized gradients, and weighted averaging. Evolution adapts algorithms to different task types: dropout like techniques appear when little data is available.
Essentially, AutoML-Zero can “automatically discover” unknown algorithms and develop new artificial intelligence (AI) programs that have not previously been discovered without any human intervention, using only basic mathematical concepts. The future evolution of Artificial Intelligent machines is present.
