decoding-encoding-alignment-critique

name: decoding-encoding-alignment-critique description: > Critical analysis framework for brain-model alignment methodology. Demonstrates that representational similarity analysis (RSA) and decoding-based alignment metrics are fundamentally insensitive to encoding manifold topology. Similar decoding behavior and high representational alignment can arise from small, non-representative neuron subpopulations. Use when: evaluating brain-DNN alignment, RSA/DSA methodology, encoding vs decoding analysis, neural representation comparison, brain-model similarity metrics, neuroscience interpretability. Activation: decoding alignment, encoding alignment, RSA critique, representational similarity analysis, brain model alignment critique, encoding manifold, decoding manifold, neural representation comparison, alignment methodology.

Decoding-Encoding Alignment Critique

Fundamental critique of similarity analysis methods in neuroscience. Shows that decoding-based alignment metrics (RSA, DSA) are misleading because they can be driven by small, non-representative neuron subpopulations.

Core Argument

Popular methods (RSA, DSA, perceptual manifolds) assume that similarity in decoding representations implies similar computation. This is not necessarily true:

1. Subpopulation dominance: High alignment can be driven by a tiny subset of neurons
2. Encoding insensitivity: Alignment metrics are blind to encoding manifold topology
3. Causal evidence: Decoding metrics unchanged when encoding topology is manipulated

Key Findings

1. Subpopulation Effect

• Similar decoding behavior and high representational alignment can arise from small, non-representative subpopulations of neurons
• The representational geometry of a population may be shaped by very few neurons
• Alignment to a few neurons ≠ alignment to the whole population

2. Encoding vs Decoding Manifolds

• Decoding manifold: How well stimuli can be read out (what most metrics measure)
• Encoding manifold: How neurons are globally organized in their responses (what alignment metrics should also consider)
• The complementary encoding paradigm characterizes global neuron organization and reveals differentiation that decoding metrics miss

3. Causal Evidence (MNIST)

• Decoding metrics remain unchanged when encoding topology is causally manipulated via training loss
• This proves decoding similarity ≠ computational similarity

Methodology: Complementary Encoding Analysis

When evaluating brain-model alignment, go beyond RSA/DSA:

1. Check subpopulation contribution: Does alignment hold when excluding top-K neurons?
2. Analyze encoding topology: How is function distributed across the population?
3. Use encoding manifolds: Characterize neuron response organization globally
4. Causal intervention: Manipulate encoding and verify decoding metrics respond

Practical Guidelines

Gradient-Level Alignment Analysis (2026-06-01 Addition)

Standard brain-model alignment only tests forward activations. arXiv:2605.28693 extends encoding analysis to backpropagated gradients:

• Traditional encoding: neural_response = W * forward_activation + b
• Gradient encoding: neural_response = W * backprop_gradient + b

Key finding: DINOv3 gradients CAN predict fMRI/MEG signals (higher visual cortex, later latencies), but their spatial/temporal organization diverges from biologically plausible backpropagation. Forward activations show strong hierarchical alignment; gradients do not.

Implication: Alignment studies should test multiple computational levels (activations, gradients, optimization dynamics) — representation similarity alone is insufficient to claim mechanistic alignment. This extends the subpopulation critique: even when decoding metrics agree, the learning mechanism may be fundamentally different.

Reusable Pattern: Gradient Encoding Pipeline

def gradient_encoding_analysis(model, images, neural_data):
    activations, gradients = {}, {}
    for layer in model.layers:
        layer.register_forward_hook(capture(activations, layer.name))
        layer.register_full_backward_hook(capture(gradients, layer.name))
    output = model(images)
    loss = some_objective(output)
    loss.backward()
    return {name: ridge_regression_predict(grad.flatten(), neural_data)
            for name, grad in gradients.items()}

When This Matters

• Brain-DNN comparison studies
• NeuroAI model validation
• Cross-species representation comparison
• Model interpretability in neuroscience
• Evaluating whether AI models "think like brains"

References

• Paper: arXiv:2605.05907 (40 pages, 27 figures)
• Authors: Bertram, Dyballa, Keller, Kinger, Zucker