Motion Illusions Generated Using Predictive Neural Networks Also Fool Humans

This paper introduces the Evolutionary Illusion GENerator (EIGen), a generative model based on video predictive neural networks that creates new visual motion illusions, which are confirmed to fool human participants, thereby supporting the hypothesis that such illusions arise from the brain's predictive processing rather than raw visual input and highlighting the value of studying "motivated failures" in AI research.

The Gist

Imagine your brain is a super-advanced weather forecaster. Its main job isn't just to look at the sky; it's to constantly guess what the weather will be like next. It uses past patterns to predict the future. Usually, this is great—it helps you catch a ball before it hits your hand or step off a curb without falling.

But sometimes, this "guessing machine" gets too confident. It sees a pattern that usually means "storm is coming," so it screams "STORM!" even when the sky is perfectly clear and static. That "false alarm" is exactly what this paper is about, but instead of weather, it's about optical illusions.

Here is the story of how the authors used a robot brain to prove that our brains are essentially "predicting" motion that isn't there.

1. The Big Idea: "Faking It Till You Make It"

The authors, Lana and Eiji, wanted to test a theory: Do we see motion in static pictures because our brains are trying to predict the future?

To test this, they didn't just ask humans to look at pictures. They built a robot brain (a neural network) that was trained only to predict what comes next in a video.

• The Robot's Job: Watch a video of people walking in a park and guess the next frame.
• The Trick: They showed the robot a picture that looks like it's moving (like the famous "Rotating Snakes" illusion), but is actually a still image.
• The Result: The robot, just like a human, got tricked! It predicted that the still image was spinning.

2. The "Evolutionary Illusion Generator" (EIGen)

This is the coolest part. The authors didn't just use existing illusions; they built a digital "Darwinian" machine to invent new ones.

Think of this like a digital art contest where the judges are robots:

1. The Artist: A computer program (called a CPPN) randomly draws thousands of weird, colorful, concentric circles and patterns.
2. The Judge: The "Predictive Robot Brain" looks at these drawings. If the robot thinks, "Oh, this looks like it's spinning!" it gives the drawing a high score. If the robot thinks it's boring and still, it gives a low score.
3. The Evolution: The drawings that get high scores are "bred" together to make new, slightly better drawings. The bad ones are thrown out.
4. The Winner: After many generations, the computer produces brand-new images that are so good at tricking the robot brain that the robot insists they are moving.

3. The Human Test: Did the Robot Fool Us?

The authors took these robot-generated images and showed them to 293 real humans.

• The Question: "Do you see this moving? How strong is the motion?"
• The Result: Yes! The humans saw motion in the robot's creations.
  • The black-and-white versions were the strongest (very convincing).
  • The color versions were weaker (the robot was a bit confused there).
  • Interestingly, the robot even "rediscovered" famous human-made illusions (like the "Medaka" fish illusion) on its own, proving it was learning the same rules humans use.

4. Why Does This Matter? (The "Aha!" Moment)

For a long time, scientists argued about why we see these illusions. Is it because our eyes twitch? Is it because of how light hits the retina?

This paper suggests a simpler, deeper answer: Our brains are so good at predicting motion that they sometimes hallucinate it.

• The Analogy: Imagine you are listening to a song. If the rhythm is very regular, your brain starts tapping its foot before the beat actually happens. If the song suddenly stops, your foot might still tap for a second because your brain predicted the beat.
• The Illusion: In these static pictures, the patterns are arranged in a way that screams "Motion!" to our predictive brain. The brain says, "I bet this is moving!" and suppresses the actual visual input that says, "No, I'm still." The "prediction" wins, and we see motion.

5. The "Failure" is the Success

The paper has a philosophical twist. Usually, in Artificial Intelligence, we want systems to never fail. But here, the authors argue that failure is a feature.

If a robot brain fails in the exact same way a human brain fails (by seeing motion where there is none), it proves that the robot and the human are using the same underlying logic. It's like two different cars both stalling at the exact same red light; it suggests they share the same engine design.

Summary

• The Problem: Why do static pictures look like they are moving?
• The Hypothesis: Because our brains are prediction machines that sometimes get too confident.
• The Experiment: We built a robot brain that predicts video frames. We used it to evolve new, fake motion illusions.
• The Proof: Humans saw motion in the robot's fake illusions, just like the robot did.
• The Conclusion: Illusory motion isn't a glitch; it's a side effect of our brain's amazing ability to predict the future. We are seeing our own predictions, not just the raw world.

In short: Your brain is a time traveler that occasionally sees the future where there is only the present.

Technical Summary

Here is a detailed technical summary of the paper "Motion Illusions Generated Using Predictive Neural Networks Also Fool Humans."

1. Problem Statement

The paper addresses the unresolved question of why static images are perceived as moving (illusory motion). While visual motion illusions are well-documented, there is no definitive causal explanation for their mechanism.

• The Gap: Existing theories (e.g., eye movements, asynchronous luminance sensitivity) are insufficient. The leading hypothesis is Predictive Coding, which suggests the brain predicts motion based on patterns and suppresses raw sensory input if the prediction is strong.
• The Challenge: To validate the predictive coding hypothesis, researchers need an artificial system that fails in the same way biological systems do. Current AI models often fail differently than humans (e.g., via adversarial attacks), making them poor models of biological perception.
• Goal: Create a generative system that produces novel motion illusions based on a predictive neural network, and verify that these AI-generated illusions fool human observers, thereby validating the "predictive coding" mechanism as a shared failure mode between AI and biology.

2. Methodology: The Artificial Perception Approach

The authors employ a four-step "Artificial Perception" framework to demonstrate that the AI and human systems share fundamental working principles:

1. Document a biological failure: Humans perceive motion in static images (e.g., Rotating Snakes).
2. Replicate in AI: A predictive neural network (PredNet) is shown to predict motion in the same static images.
3. Generate novel failures: Use an evolutionary algorithm to create new images that trick the AI into predicting motion.
4. Verify biological replication: Test if humans also perceive motion in these AI-generated images.

The System: Evolutionary Illusion GENerator (EIGen)

EIGen is a closed-loop system consisting of two main modules:

A. The Evaluator (Predictive Model)

• Architecture: Based on PredNet, a deep neural network trained on the First Person Interaction (FPSI) dataset to predict future video frames.
• Mechanism: The network takes a sequence of 20 identical static frames as input.
  • Ideal Output: A perfect copy of the input (no motion).
  • Failure Mode: If the network outputs a rotated or shifted frame, it interprets the static input as moving.
• Metric: The Lucas–Kanade optical flow algorithm calculates the magnitude and direction of the predicted motion vectors. Any non-zero motion prediction is classified as an "illusion."

B. The Generator (Evolutionary Module)

• Architecture: Uses Compositional Pattern-Producing Networks (CPPNs) to generate image patterns.
• Optimization: The NEAT (NeuroEvolution of Augmenting Topologies) algorithm evolves the CPPN parameters.
• Constraints: Images are constrained to concentric circles with alternating clockwise/counterclockwise patterns (mimicking the structure of the Rotating Snakes illusion) to facilitate the generation of specific vector fields.

C. Fitness Function
The evolutionary algorithm optimizes images based on a heuristic fitness score designed to mimic the conditions of strong human illusions:

1. Quantity: Many motion vectors (sources of illusion).
2. Magnitude: Large motion vectors (strong illusion), excluding implausibly large vectors caused by noise.
3. Alignment: Vectors should align with neighbors (agreement).
4. Contrast: Some vectors should oppose neighbors (simulating the alternating patterns found in known illusions).

• Note: The function is an ad-hoc combination of these factors, weighted to prioritize alignment (0.7) over strength (0.3).

3. Key Contributions

1. EIGen Framework: The first system capable of autonomously generating novel visual motion illusions using a predictive neural network as a "critic" and evolutionary algorithms as a "generator."
2. Validation of Predictive Coding: By successfully generating images that fool both the AI and humans, the paper provides strong empirical evidence that illusory motion is a side effect of predictive coding mechanisms shared by biological and artificial systems.
3. Rediscovery of Known Illusions: EIGen independently "rediscovered" known human-designed illusions (e.g., Fraser-Wilcox, Medaka), suggesting the evolutionary process converges on the same structural solutions that human designers found through trial and error.
4. Philosophical Shift: The paper advocates for "motivated failures"—designing AI to fail like humans—as a superior method for understanding complex biological systems compared to focusing solely on success metrics.

4. Results

A. Generated Images

• EIGen successfully generated novel black-and-white and color illusions.
• The system converged on solutions resembling the Rotating Snakes and Fraser-Wilcox illusions, confirming that the predictive network identifies similar "traps" as the human visual system.

B. Human Psychometric Survey

• Participants: $N=293$ participants evaluated 10 images (2 human-made, 8 AI-generated, including controls).
• Quantitative Findings:
  • Black-and-White vs. Color: AI-generated black-and-white illusions were perceived as moving significantly more than color illusions. The color model produced weak illusions.
  • Strength: The AI-generated illusions were generally weaker than the famous "Rotating Snakes" (human-made) but stronger than the "Fraser-Wilcox" illusion.
  • Control: Images with no predicted motion vectors were correctly rated as static by humans.
• Qualitative Findings:
  • Participants predominantly perceived rotational motion, even when the AI predicted radial (expanding/contracting) motion.
  • However, for specific AI-generated radial images, a significant portion of participants did perceive shrinking/expanding motion, aligning with the AI's prediction.
  • Free comments revealed complex perceptions (vibrating, ghosting, flickering), indicating that the illusion is not just simple rotation.

C. Correlation
There is a strong correlation between the AI's "failure" (predicting motion) and the human's "failure" (perceiving motion). This confirms that the AI and human visual systems share a fundamental mechanism for processing static patterns.

5. Significance and Implications

• Scientific Validation: The study moves beyond theoretical speculation to provide experimental proof that predictive coding is a primary driver of motion illusions. The fact that an AI trained only on video prediction (without explicit knowledge of illusions) generates illusions that fool humans suggests the mechanism is intrinsic to the architecture of prediction.
• AI as a Biological Model: The paper argues that AI models which share "failure boundaries" with humans are likely cognito-mimetic (sharing cognitive principles). This validates PredNet as a robust model of human visual perception.
• Future Research Directions:
  • The current fitness function is ad-hoc; future work should derive a "true" fitness function that perfectly matches human perception distributions.
  • The discrepancy between the AI's prediction of radial motion and human perception of rotation suggests that while the mechanism (prediction) is shared, the interpretation layers may differ.
  • This approach opens a new avenue for AI research: using "motivated failures" to reverse-engineer biological intelligence.

In conclusion, the paper demonstrates that by forcing an artificial system to fail in the same way a biological system does, researchers can uncover the underlying causal mechanisms of perception, specifically confirming that illusory motion arises from the brain (and brain-like networks) perceiving its own predictions rather than raw sensory input.