- A quantum-tunneling neural network oscillates between illusion interpretations like humans do.
- The network uses quantum physics math as a tool, not actual quantum hardware.
- Being tricked by illusions is what makes the AI more human-like, not a flaw.
A quantum-inspired network does what standard AI can't: flip between interpretations like human perception.
When Ivan Maksymov showed his neural network an optical illusion, something unusual happened. It couldn't make up its mind.
The physicist-turned-AI-researcher at Charles Sturt University in Australia built what he calls a quantum-tunneling deep neural network. When he showed it the Necker cube - that famous wireframe box that seems to flip between orientations - the network did something most conventional AI cannot.
It oscillated between interpretations, just as human perception does.
The Necker cube is a classic ambiguous figure. Human perception oscillates between seeing it from above-left or below-right.
Standard Neural Networks See Too Clearly
Most AI systems have a problem with optical illusions. Show a conventional neural network the Necker cube or Rubin's vase, and it typically picks one interpretation and sticks with it. The network sees a cube facing left, or faces in profile, and that's final.
Key figure
50x faster
Training speed of quantum-tunneling neural networks compared to classical counterparts, according to Maksymov's January 2025 follow-up research
Human brains work differently. We flip between interpretations involuntarily, sometimes holding both possibilities in mind before one wins out.
This perceptual ambiguity has long puzzled researchers trying to build machines that see like we do. The challenge isn't making AI more accurate. It's making AI appropriately uncertain.
Quantum Tunneling Makes the Network Flip
Maksymov's solution borrows mathematics from quantum physics. His network uses a quantum-tunneling activation function - the same equations that describe particles passing through barriers they classically shouldn't cross.
What is quantum tunneling?
In quantum physics, particles can pass through barriers they shouldn't be able to cross - like a ball rolling through a wall instead of bouncing off. Maksymov uses the mathematics of this phenomenon as an activation function in his neural network, allowing it to "tunnel" between different interpretations of an image.
"I trained my quantum-tunneling neural network to recognize the Necker cube and Rubin's vase illusions," Maksymov writes in The Conversation. "When faced with the illusion as an input, it produced an output of one or the other of the two interpretations. Over time, which interpretation it chose oscillated back and forth."
The network doesn't just flip between views. It can also produce outputs representing both interpretations simultaneously - a state Maksymov compares to Schrödinger's famous thought experiment.
Over time, which interpretation it chose oscillated back and forth.
Ivan Maksymov, Principal Research Fellow, Charles Sturt University
"When we see an optical illusion with two possible interpretations, researchers believe we temporarily hold both interpretations at the same time, until our brains decide which picture should be seen," he explains. "This situation resembles the quantum-mechanical thought experiment of Schrödinger's cat."
Being Fooled Is a Feature, Not a Bug
The irony runs deep here. To make artificial intelligence more like human intelligence, Maksymov had to make it worse at a certain kind of seeing. The network's ability to be tricked is precisely what makes it more human-like.
This inverts the usual framing of optical illusions as failures of perception. Neuroscientist Susana Martinez-Conde, co-creator of the Best Illusion of the Year Contest, has argued that illusions reveal something fundamental about cognition. "Illusions are a feature and not a bug," she told NPR.
Illusions are a feature and not a bug.
Susana Martinez-Conde, Neuroscientist, co-creator of Best Illusion of the Year Contest
If that's true, then modeling how we misperceive might matter as much as modeling how we perceive correctly. Maksymov suggests the approach could eventually help train pilots to recognize visual disorientation, or aid in screening for cognitive conditions like dementia - though neither application has been tested yet.
The Code Is Public, the Questions Remain Open
Maksymov has made his code available on GitHub, and the original work was published in the peer-reviewed journal APL Machine Learning in September 2024. No independent teams have yet replicated the results.
The research has continued. In January 2025, Maksymov and colleague Milan Maksimovic published follow-up work showing that quantum-tunneling neural networks can train up to 50 times faster than their classical counterparts - a significant practical advantage beyond the perception findings.
More On AI In Science
AI Model Reads DNA's Hidden Switches, One Letter at a Time
AlphaGenome helps pinpoint disease-causing mutations in the 98% of DNA that controls genes, accelerating diagnosis and drug discovery.
→The quantum framing deserves a caveat. Maksymov isn't claiming the brain uses quantum effects. He's using quantum-inspired mathematics as a modeling tool - a distinction the physicist is careful to make.
Whether this approach outperforms the latest large vision models on illusion recognition remains an open question. Some recent AI systems have shown improved performance on optical illusions without quantum-inspired methods.
Still, the core finding stands. A small network using unconventional physics can do something larger conventional networks struggle with: see the world ambiguously, the way we do.
It still can't make up its mind about the Necker cube. Neither can we.
Sources
- Primary Research: Maksymov, I.S. (2024). "Quantum-tunneling deep neural network for optical illusion recognition." APL Machine Learning. https://doi.org/10.1063/5.0225771
- Follow-up Research: Maksimovic, M. & Maksymov, I.S. (2025). "Quantum-Cognitive Neural Networks: Assessing Confidence and Uncertainty with Human Decision-Making Simulations." Big Data and Cognitive Computing. https://doi.org/10.3390/bdcc9010012
- Additional Context:
- Tech Xplore/The Conversation - Detailed explainer by Maksymov
- GitHub Repository - QT-DNN code
Fact Check: Claim-by-Claim Verification Verified
Limits and uncertainties
Core claims on QT-DNN mimicking human bistable perception via oscillation and superposition states are clearly supported by the 2024 APL Machine Learning paper, including comparisons to standard activations. No independent replications yet, as noted in article. Follow-up training speed advantage is from a peer-reviewed source but minor reliance on article's summary. Quantum framing is purely mathematical inspiration, not literal brain quantum effects. Readers should note potential hype around consciousness links, which remain speculative.
Bottom line
Article accurately reports peer-reviewed findings on quantum-inspired AI for human-like illusion perception. Claims hold up well, with code public on GitHub for verification. Promising for cognitive modeling, though real-world apps untested.
Fact-checked by Perplexity Sonar Pro on 2026-01-14
