Differences in orthographic processing across species identified by a transparent computational model

Using a transparent predictive coding model, this study reveals that while humans, baboons, and pigeons can all learn to recognize letter strings without semantic knowledge, their underlying orthographic processing strategies differ by phylogenetic distance, with humans and baboons relying more on letter-sequence representations and pigeons depending primarily on pixel- and letter-level features.

The Gist

Imagine you are teaching three very different students how to tell the difference between a "real" word they've seen before and a "fake" word they've never seen. You aren't teaching them the sounds of the letters (phonics) or what the words mean (semantics). You are just showing them shapes made of letters and asking, "Have you seen this shape before?"

The three students are:

1. Humans (who can already read).
2. Baboons (our primate cousins).
3. Pigeons (birds).

Surprisingly, all three groups got pretty good at the task. But the big question this paper asks is: How are they actually doing it? Are they using the same mental "tools," or are they solving the puzzle in completely different ways?

To find out, the researchers built a digital detective called the "Speechless Reader" (SLR). This isn't a real brain, but a computer program designed to mimic how a brain might process these letter shapes.

The Three Mental Tools (The Detective's Toolkit)

The researchers gave their digital detective three different "lenses" or tools to look at the letters. They wanted to see which lens each species relied on most:

1. The "Pixel" Lens (The Photocopier): This looks at the raw image, like a photocopier. It sees the black and white dots that make up the letter. It doesn't know it's an "A"; it just sees a specific pattern of pixels.
2. The "Letter" Lens (The Single Block): This looks at individual letters. It knows, "Oh, that's an 'A' in the first spot and a 'B' in the second." It treats letters like individual building blocks.
3. The "Sequence" Lens (The Sentence Builder): This looks at the whole flow. It understands that "TH" is a common start to a word, or that "ING" is a common ending. It sees the order and the combination of letters as a single unit.

The Results: How Each Species Solved the Puzzle

The researchers ran thousands of simulations, mixing and matching these lenses to see which combination best predicted how the real animals and humans behaved. Here is what they found:

🧑 Humans: The "Big Picture" Readers

Humans were the best at the task. The digital detective found that humans almost exclusively used the Sequence Lens.

• The Analogy: Imagine you are reading a street sign. You don't look at the individual pixels of the paint, and you don't just look at the letter 'S' in isolation. You instantly recognize the shape of the word "STOP" as a whole unit. Humans have trained their brains to see the "flow" of letters. They see the forest, not the trees.

🐒 Baboons: The "Hybrid" Learners

Baboons did well, but they used a mix of tools. They used the Sequence Lens (like humans), but they also relied heavily on the Letter Lens.

• The Analogy: A baboon is like a student who is learning to read. They recognize the word "STOP" as a group, but they are still double-checking, "Is that an 'S'? Yes. Is that a 'T'? Yes." They are building the word from the middle ground—seeing the blocks and the flow.

🐦 Pigeons: The "Detail-Oriented" Observers

Pigeons also did better than random chance, but their strategy was totally different. They barely used the Sequence Lens. Instead, they relied mostly on the Pixel Lens and the Letter Lens.

• The Analogy: A pigeon is like a master puzzle solver who looks at the tiny details. They don't see the word "STOP" as a single concept. They see, "That's a specific curve here, a straight line there, and a specific letter 'S' in this spot." They are looking at the individual grains of sand rather than the whole beach.

Why Does This Matter? (The Evolutionary Story)

The paper suggests that these differences aren't random; they are tied to evolutionary history.

• Humans and Baboons are closely related (primates). Our brains are wired to see global patterns (the whole picture). This helps us recognize faces, tools, and complex scenes quickly.
• Pigeons are birds. Their brains are wired to look at local details. In nature, pigeons need to spot tiny grains of food hidden in grass or dirt. They are experts at finding small differences in a cluttered background.

The researchers call this "neuro-cognitive phenotyping." It's like taking a fingerprint of how a brain thinks, not just what it knows.

The Takeaway

Even though humans, baboons, and pigeons all learned to do the same task (telling real words from fake ones), they took very different mental paths to get there.

• Humans zoomed out to see the sequence.
• Baboons looked at both the sequence and the individual letters.
• Pigeons zoomed in to see the pixels and individual letters.

This study shows that "intelligence" isn't just about getting the right answer; it's about the unique, evolutionary toolkit each species uses to solve the problem. The pigeon isn't "dumb" for not seeing the word as a whole; it's just using the super-powerful detail-oriented brain that helped its ancestors survive in the wild.

Technical Summary

1. Problem Statement

Reading is traditionally viewed as a uniquely human skill requiring phonological and semantic knowledge. However, recent behavioral studies have demonstrated that non-human primates (baboons) and birds (pigeons) can learn to distinguish learned orthographic strings (pseudo-words) from novel letter combinations without explicit phonological or semantic training.

The core scientific question addressed by this study is: Do phylogenetically distant species solve orthographic decision tasks using similar underlying neuro-cognitive mechanisms, or do they rely on fundamentally different representational strategies? While behavioral performance (accuracy) can be similar across species, the internal cognitive processes (the "how") remain unknown. The authors aim to use computational neurocognitive phenotyping to uncover these hidden mechanisms and determine if phylogenetic proximity correlates with cognitive similarity.

2. Methodology

A. The Speechless Reader (SLR) Model

The authors developed a transparent computational model based on predictive coding principles. The model simulates orthographic decisions (classifying a string as "learned" or "novel") without relying on phonology or semantics.

• Core Mechanism: The model minimizes prediction errors by comparing sensory input against a stored lexicon of learned items.
• Three Representational Levels: The model calculates prediction errors at three distinct hierarchical levels:
  1. Pixel-Level (oPE): Visual-orthographic representation based on pixel-by-pixel differences between the input and the mean of learned words.
  2. Letter-Level (LPE): Representation based on the probability of specific letters appearing at specific positions.
  3. Sequence-Level (sPE): Representation based on the probability of letter sequences (e.g., "Ha", "Hau") given the lexicon.
• Decision Process: The model sums these prediction errors to generate a "word-likeness" estimate. A binary decision is made based on a fitted categorization threshold.

B. Experimental Data & Design

The study utilized three existing datasets where subjects learned to categorize letter strings without meaning:

• Humans ($N=37$): Learned 5-letter pseudo-words.
• Baboons ($N=6$): Learned 4-letter pseudo-words.
• Pigeons ($N=4$): Learned 4-letter pseudo-words.

C. Computational Phenotyping Approach

Instead of fitting a single model to all subjects, the authors created digital twins for each individual:

1. Individual Lexicons: Constructed based on the specific items each subject successfully learned.
2. Model Variants: Seven different model configurations were simulated for each participant, representing all possible combinations of the three prediction error types (e.g., oPE only, LPE + sPE, oPE + LPE + sPE).
3. Best-Fit Selection: For each participant, the model variant that minimized the Mean Squared Error (MSE) between the simulation and the actual behavioral data was selected as that individual's neuro-cognitive phenotype.
4. Stability Analysis: Split-half reliability tests were conducted to ensure the identified phenotypes were robust.

3. Key Results

A. Behavioral Performance

• Humans: Achieved the highest accuracy (>90%).
• Baboons: Performed well (~75% accuracy).
• Pigeons: Performed above chance (~70% accuracy).
• Prediction Error Effect: Across all species, learned strings elicited significantly lower prediction errors than novel strings, confirming the model's ability to capture the learning effect.

B. Neuro-Cognitive Phenotypes (Representational Strategies)

The most significant finding was the divergence in the representational strategies used by each species:

• Humans: Predominantly relied on Sequence-Level (sPE) representations.
  • 86% of humans used only sPE.
  • This suggests humans process letter strings as holistic units or large chunks (bigrams/trigrams) rather than isolated letters or pixels.
• Baboons: Used a hybrid strategy, combining Sequence-Level (sPE) and Letter-Level (LPE) representations.
  • 50% used LPE + sPE.
  • Only one baboon matched the human phenotype (sPE only).
• Pigeons: Relied heavily on Letter-Level (LPE) and Pixel-Level (oPE) representations.
  • No pigeon relied exclusively on sequence-level representations.
  • The best-fitting models for pigeons included LPE, oPE, or combinations thereof, but rarely sPE as the sole driver.

C. Model Fit and Stability

• The SLR model successfully predicted behavior across all species (high correlation between simulated and actual accuracy).
• The identified phenotypes showed high split-half reliability (agreement >65% across all species), confirming that the representational strategies are stable traits of the individuals.

4. Key Contributions

1. Methodological Innovation: The study introduces computational neurocognitive phenotyping as a tool to compare cognitive mechanisms across species. By using a transparent model with modular components, the authors could isolate exactly which cognitive operations drive behavior, rather than just measuring output accuracy.
2. Evolutionary Insight: The findings challenge the assumption that similar behavioral performance implies similar cognitive processes.
   • Phylogenetic Proximity: Humans and baboons (primates) share more similar processing strategies (heavy reliance on sequence/structure) than either does with pigeons.
   • Ecological Adaptation: The authors link the pigeon's reliance on pixel/letter-level processing to their ecological need for local processing (distinguishing small grains in clutter), whereas humans and primates exhibit a propensity for global processing (grouping elements into configurations).
3. Decoupling Reading from Phonology: The study demonstrates that robust orthographic decision-making can be achieved using purely visual-orthographic predictive coding, even in species without phonological systems.

5. Significance

• Redefining Reading Evolution: The paper suggests that the evolutionary precursor to human reading is not a specific "reading module" but a general capacity for global visual integration and sequence prediction. Humans have simply specialized this capacity to the highest degree (sequence-level processing).
• Cognitive Continuum: It proposes a "global-to-local" processing continuum. Humans sit at the global/sequence end, baboons in the middle (hybrid), and pigeons at the local/pixel end. This provides a quantitative framework for understanding how different brains solve the same visual categorization problem.
• Future Directions: The study highlights that while visual-orthographic processing explains a large variance in behavior, human reading involves additional layers (phonology, semantics) not captured by the "Speechless" model. This sets a baseline for future models that integrate these higher-level linguistic components.

In conclusion, the paper successfully uses a transparent computational model to reveal that while humans, baboons, and pigeons can all learn to recognize letter strings, they do so using distinct neuro-cognitive phenotypes that reflect their evolutionary history and ecological niches.