- Scientists have developed a model of artificial intelligence that can directly communicate with each other through recurrent neural networks (RNN) and teach tasks.
- An artificial intelligence model communicated with its “brother” and passed on its information using natural language processing (NLP).
- This development paves the way for the development of complex networks that can be integrated into humanoid robots and that artificial intelligence can have human-like learning and communication skills.
Research shows that the next evolution in artificial intelligence may lie in actors that can communicate directly and teach each other to perform tasks.
Scientists have modeled an artificial intelligence network capable of learning and performing tasks based solely on written instructions. This artificial intelligence then told what it had learned to a “brother” artificial intelligence, which performed the same task despite having no prior training or experience. The scientists published March 18 in the journal Nature. in your articlessaid that the first artificial intelligence communicated with its brother using natural language processing (NLP).
NLP is a subfield of artificial intelligence that aims to recreate human language on computers. In this way, machines can understand and reproduce written texts or speech in a natural way. These are built on neural networks, which are collections of machine learning algorithms modeled to mimic the layout of neurons in the brain.
Alexandre Pouget, leader of the University of Geneva Neuro-Centre and lead author of the paper, said: “Once these tasks were learned, the network was able to identify them to a second network, thus reproducing the information. “To our knowledge, this is the first time two AIs have been able to talk to each other in a fully linguistic way.”
The scientists achieved this knowledge transfer by starting with an NLP model called “S-Bert” that was pre-trained to understand human language. They connected S-Bert to a smaller neural network focused on interpreting sensory input and simulating motor actions in response.
This composite artificial intelligence (sensorimotor-recurrent neural network) [RNN]) was then trained on a set of 50 psychophysical tasks. These focused on responding to a stimulus, such as responding to light, through instructions fed through the S-Bert language model. Thanks to the embedded language model, RNN understood complete written sentences. In this way, it performed tasks from natural language instructions 83% correctly on average, even though it had never seen training images or performed the tasks before. Then RNN; Using linguistic instructions, it was able to communicate the results of its sensorimotor learning to an identical sibling AI that sequentially performed tasks it had never performed before.
The inspiration for this research is the way humans learn to perform tasks by following verbal or written instructions, even if we have never performed such actions before. This cognitive function distinguishes humans from animals. For example, before you can train a dog to respond to verbal instructions, you need to show him something. Artificial intelligence-supported chat robots; While they can interpret linguistic instructions to create an image or text, they cannot translate written or spoken instructions into physical actions, let alone explain the instructions to another AI.
Researchers; They created an artificial intelligence with human-like learning and communication skills by simulating the regions of the human brain responsible for perceiving language, interpreting it, and taking actions based on instructions. This alone will not lead to the rise of artificial general intelligence (AGI), where a single AI actor can reason as well as a human and perform tasks across multiple domains. Researchers stated that artificial intelligence models like the one they created could help us understand how the human brain works.
In their statement, the researchers said, “The network we developed is very small. “On this basis, there are now no obstacles to developing much more complex networks that can be integrated into humanoid robots that can understand us and also understand each other.” they said.
Compiled by: Esin Özcan