THE
WESTERN INSTITUTE
FOR
ADVANCED STUDY

ENERGY EFFICIENT COMPUTATION
The human brain is 99.9% energy efficient and runs on 20 Watts. If such efficient computation is possible, why does ChatGPT use up the energy resources of a small city?
Like a lightbulb, the human brain’s energy consumption can be measured in watts (W). The greater the number of watts, the more energy is consumed for a device to function. Incredibly, the brain has an energy budget of just 20W. Somehow, the most powerful computer in the world - the human brain - runs on the same amount of energy as a lightbulb.
​
Neuroscientists calculated the brain’s energy use by measuring the amount of caloric energy entering the brain. Neuroscientists also measured the total amount of work done by the brain, measured by energy turnover. Since these two numbers – representing energy intake and energy use – are about equal, the brain is understood to be an energy-efficient system [1, 2].
​
Generative AI does not share this energy efficiency. Despite applications like ChatGPT becoming more prevalent, it consumes enough daily energy to power 180,000 American households, or a small city [3]. To understand this discrepancy in computational resource management, we first have to understand how generative AI works.
​
Generative AI is a type of artificial intelligence that can create new content based on the types of information it was previously exposed to [4]. To create new content, many hours of training and exposure to large amounts of data are necessary to teach a generative AI model, and models are typically trained for a specific purpose, like generating text or photos. For example, if an application was only trained in comic book art, it may be able to develop a new comic strip but would be unlikely to produce poetry.
​
This type of learning has many parallels to the human brain. Generative AI and our brains are constantly making predictions using previously learned information [5-7]. By using prior knowledge, an optimal outcome is selected from all possible available outcomes, known as a probability distribution, in a process called predictive processing [8]. The way each entity approaches predictive processing is where generative AI differs from the human brain.
​
Generative AI engages in statistical computation [4]. Using this approach, the model will select the statistically most likely outcome from a probability distribution. Importantly, it can only select from options it has learned, and when options are learned they are encoded in the system. For example, if you were to ask ChatGPT to finish this sentence: “A player earns points in American Football by scoring a __________,” the expected response would be “touchdown” or “field goal.” However, if this model hasn’t encoded the word “touchdown” it would not be able to produce this response. Programs that create new content must also encode the patterns in the data they are exposed to, like grammar and tone of voice. Considering that generative AI models are trained on billions of data points, the encoding process requires significant amounts of energy [5].
​
​
​
​
​
​
​
​
​
​​​​​​​
​
​​
​

In contrast, the brain is thought to engage in thermodynamic computation [8-12]. Like generative AI, our biological neural networks select the most likely outcome from a probability distribution, but the brain does not need to spend energy on encoding all possible alternatives. In this case, when the brain has selected the optimal outcome, all other possibilities are discarded, and the energy previously being used to encode those possibilities is available for the brain to use for other meaningful work. By modeling the brain as a thermodynamic computer, as researchers at the Western Institute for Advanced Study have done, we can understand why the brain is so energy efficient.
Until recently, it was challenging to explain how exactly the brain functions with such efficiency. Historically, neuroscientists have used classical models to describe the brain, but these models may not be the most useful in explaining the energy efficiency the brain exhibits during predictive processing. ​
In classical models, neurons are modeled as binary computational units. A binary computational unit is described as either being on or off, like a light switch or a transistor. Here, a neuron is either firing or not. This signal can travel from one neuron to the next, causing other neurons to fire too. While this classical binary model effectively describes some neurons, like spinal neurons, it does not fully describe the behavior of cortical neurons, like those in the brain. ​
Because classical models do not fully explain cortical neurons, the Western Institute for Advanced Study embarked on its first Grand Challenge: to devise a theoretical framework that more accurately describes the energy efficiency of the brain. In this new framework, cortical neurons are not described using a binary measure (either off or on), but instead as a probabilistic two-state system, meaning that cortical neurons can exist in one of two states, an off-state or an on-state. However, instead of using a binary measure to describe these states, like classical models, this new framework uses the probability, or likelihood, of a neuron switching from an off-state to an on-state. This difference allows for the incorporation of random electrical noise into the model.
This new theoretical framework is built on four independent models of cortical computation: wave mechanics [11], Hamiltonian mechanics [10], matrix mechanics [9], and mean field theory [8]. Each of these mathematical concepts demonstrates the process discussed above, of selecting an optimal signal from a probability distribution whereby all other possibilities are discarded, freeing up energy to do more work. Through this process, we make predictions about our surrounding environment and ultimately select outcomes that most closely match the surrounding environment. In this way, we reconstruct an internal cognitive model of the world around us. ​
While this may seem like a trivial change from classical models of computation, this new theoretical framework may be the foundation for future neuroscience breakthroughs. This new framework more fully describes the energy efficiency of the human brain, compared with previous models. In addition, this new theoretical framework produces quantifiable predictions which can be tested in the laboratory. Evaluating these predictions will open up new avenues for research, and may help us to engineer more efficient, more sustainable, and more generally intelligent systems, who can solve problems alongside us.
​
​
REFERENCES
1.Sokoloff L. Localization of Functional Activity in the Central Nervous System by Measurement of Glucose Utilization with Radioactive Deoxyglucose. Journal of Cerebral Blood Flow & Metabolism. 1981;1(1):7-36. doi:10.1038/jcbfm.1981.4
2.Magistretti PJ, Pellerin L. Cellular mechanisms of brain energy metabolism and their relevance to functional brain imaging. Philosophical Transactions of the Royal Society B Biological Sciences. 1999;354(1387):1155-1163. doi:10.1098/rstb.1999.0471
3.Gordon C. ChatGPT and generative AI innovations are creating sustainability havoc. Forbes. https://www.forbes.com/sites/cindygordon/2024/03/12/chatgpt-and-generative-ai-innovations-are-creating-sustainability-havoc/. Published March 17, 2024.
4.What is generative AI? McKinsey & Company. Published April 2, 2024. https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-generative-ai
5.Clark A. What generative AI reveals about the human mind. TIME. https://time.com/6552233/generative-ai-reveals-human-mind/. Published January 5, 2024.
6.Melloni L, Schwiedrzik CM, Müller N, Rodriguez E, Singer W. Expectations change the signatures and timing of electrophysiological correlates of perceptual awareness. Journal of Neuroscience. 2011;31(4):1386-1396. doi:10.1523/jneurosci.4570-10.2011
7.Kok P, Jehee JFM, De Lange FP. Less is more: expectation sharpens representations in the primary visual cortex. Neuron. 2012;75(2):265-270. doi:10.1016/j.neuron.2012.04.034
8.Stoll EA. An energy-efficient process of non-deterministic computation drives the emergence of predictive models and exploratory behavior. Frontiers in Cognition. 2024;2. doi:10.3389/fcogn.2023.1171273
9.Stoll EA. Random electrical noise drives non-deterministic computation in cortical neural networks. bioRxiv (Cold Spring Harbor Laboratory). Published online December 7, 2022. doi:10.1101/2022.12.03.518981
10.Stoll EA. The mechanics underpinning non-deterministic computation in cortical neural networks. Applied Math. 2024; 4(3): 806-827. doi:10.3390/appliedmath4030043
11.Stoll EA. Modeling electron interference at the neuronal membrane yields a holographic projection of representative information content. bioRxiv (Cold Spring Harbor Laboratory). Published online December 7, 2022. doi:10.1101/2022.12.03.518989
12. Stoll EA. A thermodynamical model of non-deterministic computation in cortical neural networks. Physical Biology. 2023;21(1):016003. doi:10.1088/1478-3975/ad0f2d
13.Grider MH, Jessu R, Kabir R. Physiology, action potential. StatPearls - NCBI Bookshelf. Published May 8, 2023. https://www.ncbi.nlm.nih.gov/books/NBK538143/