Artificial intelligence (AI) is very energy intensive, particularly during its training phase. This is a problem of such magnitude thatElon MuskElon Musk predicted that the world would run out of electricity for AI within a year. But researchers from Cold Spring Harbor Laboratory in the United States have perhaps found the solution, taking inspiration from brainbrain human.
“ As impressive as they are ChatGPTChatGPT and all these current AI technologies, in terms of interacting with the world physicalphysical, they are still very limited. Even in areas like solving math problems and writing essays, they need to practice on billions and billions of examples before they can do them well. “, explained Kyle Daruwalla, the lead author of the article published in the journalFrontiers in Computational Neuroscience.
A working memory to allow synapses to be updated
For him, the problem lies in the movement of data. Artificial neural networks are made up of billions of connections, and data must pass through entire circuits before the model updates. With this new approach, each artificial neuron receives feedback through an auxiliary memory network and can be modified in real time.
This creates a direct connection between the working memoryworking memory and synaptic updates, providing new evidence for an unproven theory in neuroscience that links working memory to learning and academic performance. This new approach could therefore considerably reduce the computing power and therefore theenergyenergy necessary for the functioning of artificial intelligence.
