Artificial intelligence’s impressive performance comes at the cost of significant energy consumption, with complex tasks demanding even more power. In response, researchers Víctor López-Pastor and Florian Marquardt from the Max Planck Institute for the Science of Light in Erlangen, Germany, have introduced an innovative approach to training AI that offers remarkable energy efficiency. Their method diverges from the conventional use of digital artificial neural networks and relies on physical processes instead. The results of their work are published in the journal Physical Review X.
Training models like GPT-3, the foundation of eloquent chatbots like ChatGPT, demands an astounding amount of energy. While the exact energy consumption for GPT-3 remains undisclosed by Open AI, estimates suggest it could equate to the annual consumption of 200 German households. Despite these energy expenditures, such AI models have yet to grasp the underlying meaning of phrases they analyze.
To address the energy consumption challenges posed by conventional digital computers, researchers have been exploring neuromorphic computing, a concept inspired by the brain’s way of functioning. Unlike artificial neural networks, which run on conventional digital hardware, neuromorphic computing emulates the brain’s parallel processing approach, where neurons and synapses combine processing and memory functions.
Marquardt, the director of the Max Planck Institute for the Science of Light, highlighted that the data transfer between processors and memory in conventional systems consumes substantial energy when training neural networks with billions of parameters. In contrast, the human brain’s parallel processing enables remarkable energy efficiency.
Neuromorphic computing seeks counterparts to biological neurons, with photonic circuits being one potential candidate. These circuits use light instead of electrons to perform calculations and simultaneously serve as switches and memory cells.
López-Pastor and Marquardt have devised an efficient training method for neuromorphic computers, introducing the concept of a “self-learning physical machine.” This approach conducts training as a physical process, where the machine optimizes its parameters autonomously, eliminating the need for external feedback and making training significantly more efficient.
The researchers emphasize that this method can be applied to any physical process as long as it meets specific criteria, including reversibility and non-linearity. Reversible processes can run forwards or backwards with minimal energy loss, while non-linear processes are sufficiently complex to perform intricate transformations between input data and results.
Optics, with its reversible and non-linear processes, is one promising avenue for implementing self-learning physical machines. López-Pastor and Marquardt are collaborating with an experimental team to develop an optical neuromorphic computer that processes information using superimposed light waves and suitable components to regulate interactions.
They aim to present the first self-learning physical machine within three years, catering to neural networks with even more synapses and larger datasets. As AI continues to evolve, the adoption of efficiently trained neuromorphic computers is likely to gain momentum, marking a significant step in the advancement of artificial intelligence.
By Impact Lab