-
Asked by anon-379566 on 10 Jan 2024. This question was also asked by anon-379522.
-
Andrew Maynard answered on 10 Jan 2024:
That’s what some experts are hoping for. And of course, experts like Ray Kurzweil who believe in the singularity (when AI creates superintelligent AI) certainly believe this.
We don’t yet know if this is possible though as it would most likely require AI that can invent, be creative, and innovate — we’re not there yet
-
-
Carl Peter Robinson answered on 10 Jan 2024: last edited 10 Jan 2024 5:45 pm
This is an interesting question as it can be viewed from a number of angles. In one sense, we have a method called transfer learning. This is where an existing, trained machine learning (ML) model that performs a specific task can be further trained with new data to perform a task in the same area, or similar area, as the first task.So, in a sense, you are creating a new AI model based on an existing AI model.
You also have the case of synthetic data, which can be generated by certain types of ML models (e.g., GANs, Generative AI models, etc.). This synthetic data can then be used to supplement (augment) existing real data when training a new ML model; this has been used in research to increase dataset sizes that were too small, to enable more effective deep learning efforts. So, in effect, AI helped train new AI by generating some of the data to train on.
But I think you’re really asking if an ML model or pipeline exists whose purpose is create other ML models, be they of the same architecture, or to explore completely new structures. I’m pretty sure I’ve come across research papers that have talked about doing this (there’s so much in this field to take note of and try to remember at the moment!), but I haven’t seen any concrete announcements and evidence as yet. I’m sure you can write some code that randomly creates ML model structures, or ask ChatGPT to do it based off all the code that’s sucked in during its training. But whether they would be cohesive, sound architectures is another matter.
-
Kevin Tsang answered on 10 Jan 2024:
I’ll give two examples which could be interpreted as a “yes”. In my daily work, I use AI tools like Google Bard, ChatGPT, and Github Copilot to write code to develop new AI models in my field of work.
In reinforcement learning (one type of machine learning and AI), is the idea of self-play. An example of self-play is training an AI to play the game of chess (a two-player game), which requires two AIs to play against each other. The iterative process of training will create new versions of the winning AI as we create more and better AI which can play chess.
Comments
Nora V commented on :
Yes, AI can be used to create more AI through automated or AI-assisted machine learning.There are several ways AI is used in the creation of more AI. Some of these include Automated Machine Learning (AutoML), Hyperparameter optimization (such as genetic algorithms or reinforcement learning), data augmentation (can be used to generate synthetic data or perform data augmentation).
It’s important to note that while AI can automate many aspects of the machine learning process, human expertise is still crucial for defining the problem, selecting the appropriate data, interpreting results, and making decisions about the model’s deployment and use.