Synthetic intelligence is consuming monumental quantities of electrical energy in the US. In response to the Worldwide Power Company, AI methods and knowledge facilities used about 415 terawatt hours of energy in 2024. That accounts for greater than 10% of the nation’s complete electrical energy manufacturing, and demand is projected to double by 2030.
This speedy development has raised considerations about sustainability. In response, researchers at a College of Engineering have created a proof-of-concept AI system designed to be much more environment friendly. Their strategy may cut back power use by as much as 100 occasions whereas additionally enhancing efficiency on duties.
A Hybrid Method Known as Neuro-Symbolic AI
The analysis comes from the laboratory of Matthias Scheutz, Karol Household Utilized Expertise Professor. His crew is growing neuro-symbolic AI, which mixes conventional neural networks with symbolic reasoning. This technique mirrors how individuals strategy issues by breaking them into steps and classes.
The work shall be introduced on the Worldwide Convention of Robotics and Automation in Vienna in Might and can seem within the convention proceedings.
Instructing Robots to See, Perceive, and Act
In contrast to acquainted massive language fashions (LLMs) reminiscent of ChatGPT and Gemini, the crew focuses on AI methods utilized in robotics. These methods are often known as visual-language-action (VLA) fashions. They prolong LLM capabilities by incorporating imaginative and prescient and bodily motion.
VLA fashions absorb visible knowledge from cameras and directions from language, then translate that data into real-world actions. For instance, they’ll management a robotic’s wheels, arms, or fingers to finish a process.
Why Conventional AI Struggles With Easy Duties
Standard VLA methods rely closely on knowledge and trial-and-error studying. If a robotic is requested to stack blocks right into a tower, it should first analyze the scene, determine every block, and decide the right way to place them appropriately.
This course of typically results in errors. Shadows might confuse the system a couple of block’s form, or the robotic might place items incorrectly, inflicting the construction to break down.
These errors are just like the issues seen in LLMs. Simply as robots can misplace blocks, chatbots can generate false or deceptive outputs. Examples embrace fabricating authorized circumstances or producing photographs with unrealistic particulars reminiscent of additional fingers.
How Symbolic Reasoning Improves Accuracy and Effectivity
Symbolic reasoning affords a special technique. As a substitute of relying solely on patterns from knowledge, it makes use of guidelines and summary ideas reminiscent of form and stability. This permits the system to plan extra successfully and keep away from pointless trial and error.
“Like an LLM, VLA fashions act on statistical outcomes from massive coaching units of comparable situations, however that may result in errors,” stated Scheutz. “A neuro-symbolic VLA can apply guidelines that restrict the quantity of trial and error throughout studying and get to an answer a lot quicker. Not solely does it full the duty a lot quicker, however the time spent on coaching the system is considerably diminished.”
Robust Ends in Puzzle Exams
The researchers examined their system utilizing the Tower of Hanoi puzzle, a basic drawback that requires cautious planning.
The neuro-symbolic VLA achieved a 95% success price, in contrast with simply 34% for traditional methods. When given a extra complicated model of the puzzle that it had not encountered earlier than, the hybrid system nonetheless succeeded 78% of the time. Conventional fashions failed each try.
Coaching time additionally dropped sharply. The brand new system discovered the duty in solely 34 minutes, whereas typical fashions required greater than a day and a half.
Huge Power Financial savings in Coaching and Use
Power consumption was diminished dramatically as nicely. Coaching the neuro-symbolic mannequin required just one% of the power utilized by a typical VLA system. Throughout operation, it used simply 5% of the power wanted by typical approaches.
Scheutz in contrast this inefficiency to on a regular basis AI instruments. “These methods are simply attempting to foretell the following phrase or motion in a sequence, however that may be imperfect, and so they can provide you with inaccurate outcomes or hallucinations. Their power expense is usually disproportionate to the duty. For instance, once you search on Google, the AI abstract on the prime of the web page consumes as much as 100 occasions extra power than the era of the web site listings.”
The Rising Pressure of AI on Energy Infrastructure
As AI adoption accelerates throughout industries, demand for computing energy continues to climb. Corporations are constructing more and more massive knowledge facilities, a few of which require tons of of megawatts of electrical energy. That degree of consumption can exceed the wants of complete small cities.
This development has sparked a race to broaden infrastructure, elevating considerations about long-term power limits.
A Extra Sustainable Path for AI
The researchers counsel that present approaches primarily based on LLMs and VLAs will not be sustainable in the long term. Whereas these methods are highly effective, they devour massive quantities of power and might nonetheless produce unreliable outcomes.
In distinction, neuro-symbolic AI affords a special route. By combining studying with structured reasoning, it might present a extra environment friendly and reliable basis for future AI methods.
