When Integration Fails

on

 

When Integration Fails

Modeling Dyslexia in Artificial Neural Networks

What if cognitive impairment isn’t about “broken parts” … but about broken coordination?

That’s the idea behind a new NeuroAI framework we’ve been developing: modeling dyslexia not as random failure, not as noisy computation, and not as a damaged representation — but as a disruption in information integration between systems that normally work together.

And it turns out that shift in perspective changes everything.

Reading Is Not a Single Skill

Reading feels effortless for fluent readers. But underneath that smooth surface, the brain is orchestrating a remarkable coordination:

  • Visual letter recognition
  • Phonological sound mapping
  • Sequential decoding
  • Rapid temporal processing
  • Semantic access

Dyslexia, affecting roughly 5–10% of people, is typically characterized by difficulty with phonological processing and decoding — especially with unfamiliar or pseudowords.

Neuroimaging consistently shows something important:

It’s not that one “reading area” is missing.

It’s that connectivity between regions — especially between orthographic and phonological systems — is atypical.

So the real question becomes:

What happens if we reduce integration in an artificial system that reads?

From Performance to Failure

Artificial neural networks are usually optimized for success. They map graphemes to phonemes with near-perfect accuracy when trained properly.

But biological systems don’t just succeed — they fail in structured ways.

So instead of asking:
“How do we make a model read better?”

We asked:
“How do we make a model fail like dyslexia — in a principled way?”

Not through random noise.
Not by deleting half the model.
Not by sabotaging learning.

But by systematically reducing the strength of integration between modules.

Integration as a Control Parameter

We built a modular reading architecture:

  • Orthographic encoder
  • Phonological encoder
  • Cross-attention integration layer
  • Sequential decoder

Then we introduced a simple but powerful variable:

Integration strength (α).

When α = 1 → full coordination.
When α decreases → reduced cross-representational binding.
When α approaches a critical threshold → decoding collapses.

And here’s the interesting part:

Performance doesn’t degrade linearly.

It drops sharply near a threshold — almost like a phase transition.

Pseudoword decoding fails first.
Substitution errors rise.
Letter-position confusion increases.
Confidence destabilizes.

This pattern mirrors empirical dyslexia profiles.

Not because we “damaged” the model —
but because we weakened integration.

Measuring Integration (Without Mysticism)

We borrowed from Integrated Information Theory (IIT), but in a computationally practical way.

Instead of calculating full Φ (which is intractable in large networks), we measured:

  • Mutual information between modules
  • Transfer entropy
  • Effective connectivity strength

We defined an Integration Index that quantifies how much coordinated information contributes to output accuracy.

The striking finding:

Integration metrics begin to drop before performance visibly collapses.

In other words:
Loss of coordination precedes observable impairment.

That has implications far beyond reading.

Remediation as Reintegration

We didn’t stop at degradation.

We simulated structured literacy intervention:

  • Curriculum-based grapheme–phoneme alignment
  • Multi-task phonological training
  • Gradual reintegration of cross-attention weights

What we observed:

Integration increased before accuracy fully recovered.

Reintegration predicted recovery trajectory.

This reframes remediation not as “fixing errors” —
but as restoring coordination.

Why This Matters for NeuroAI?

This work suggests a broader principle:

Cognitive impairment may be modeled as disruption of integration across representations.

That’s not limited to dyslexia.

It could extend to:

  • Aphasia
  • Developmental language disorder
  • Certain psychiatric conditions
  • Even broader questions of cognitive coherence

In artificial systems, we often obsess over scaling.
But maybe understanding cognition requires studying how things break — systematically.

NeuroAI shouldn’t just model intelligence.
It should model vulnerability.

A Bigger Theoretical Move

This framework reframes cognitive capacity as a dynamical property:

Integration is a control parameter governing cognitive coherence.

Too little integration → fragmentation.
Too much rigid integration → inflexibility.
Healthy cognition may sit in a critical regime between independence and collapse.

That’s not just a dyslexia story.

That’s a systems neuroscience story.

Ethical Framing

This work does not model suffering.
It does not simulate lived experience.
It does not replace diagnosis.

It models mechanisms.

And mechanism is what allows:

  • Hypothesis testing
  • Intervention simulation
  • Educational tool development
  • AI-assisted screening support

The goal is not pathology replication.

It’s understanding coordination failure.

Where This Could Go

Next directions include:

  • Developmental learning trajectories
  • Cross-linguistic orthographic depth modeling
  • Spiking neural implementations
  • Phase-transition formalization
  • Personalized AI reading tutors

But the deeper question remains:

If cognition depends on integration…

What other mental phenomena collapse when coordination weakens?

And can AI help us see those thresholds more clearly than biology alone?

This session wasn’t just about dyslexia.

It was about moving from modeling success to modeling failure —
and discovering that failure, too, has structure.

And structure is where science lives.

Comments

Post a Comment