# Cross-Layer Structural Isomorphisms # The deep connections that make a single grammar span all three layers thesis: > The same structural patterns — tiling, dispatch, fusion, decomposition, and reduction — recur at every level, from silicon to synapse to subjective experience. layers: backend: project: TileGym domain: CUDA Tile kernels, GPU tiling role: Computational substrate grammar_summary: "decompose → dispatch → tile → fuse → reduce" interface: project: Chaos Hearing domain: Auditory science, nonlinear dynamics, spectral cognition role: Scientific model grammar_summary: "transduce → decompose → infer → group → predict" frontend: project: McDermott Lab Demos domain: Auditory illusions, scene analysis, generative listening role: Perceptual surface grammar_summary: "stimulate → analyze → group → bind → perceive" isomorphism_table: - concept: Primitive backend: "Tile T(M,N)" interface: "Oscillator O(μ,ω₀)" frontend: "Illusion Ψ(stim,percept)" - concept: Decomposition backend: "Tile blocking" interface: "Cochlear frequency analysis" frontend: "Auditory scene analysis" - concept: Dispatch backend: "Backend selector" interface: "Frequency-place mapping" frontend: "Perceptual frame selection" - concept: Fusion backend: "Kernel fusion F(∘)" interface: "Spectral mixture kernel" frontend: "Texture perception" - concept: Reduction backend: "Split-K merge" interface: "Scene integration" frontend: "Source segregation" - concept: Attention backend: "A(Q,K,V)" interface: "Bayesian prediction P" frontend: "Grouping cues G" - concept: Normalization backend: "RMS/LayerNorm" interface: "Cochlear compression" frontend: "Perceptual continuity" - concept: Self-Supervision backend: "Student-Teacher EMA alignment" interface: "Learnable filterbank" frontend: "Illusory restoration" - concept: Causal Uncertainty backend: "Objective function optimization (minimize target distance)" interface: "Precision-weighting (balancing prior vs. likelihood)" frontend: "Ambiguity resolution (hearing voices in noise)" - concept: Spectral Filtering backend: "Fixed Hankel eigenvector convolution (no gradient)" interface: "Cochlear tonotopic map (physically fixed resonance)" frontend: "Absolute pitch perception (fixed internal template)" - concept: Failure Mode backend: "Wrong dispatch" interface: "Hallucination" frontend: "Illusion" - concept: Uncertainty backend: "Backend fallback" interface: "Posterior variance" frontend: "Causal ambiguity" - concept: Improvisation backend: "Autoregressive decode" interface: "Jazz solo over changes" frontend: "Generative listening (BASS)" deep_analogies: - name: "The Dispatch-Tuning Correspondence" description: > TileGym's @dispatch decorator selects the right kernel implementation for an operation. The cochlea's tonotopic map selects the right oscillator for a frequency. The perceptual system selects the right interpretive frame for a stimulus. All three are instances of selective routing to specialized processors. backend_code: | @dispatch("fmha") def fmha(q, k, v, scaling=None, is_causal=True): ... interface_equation: "dz/dt = (μ + iω₀)z - (a + i/3)|z|²z + n_z(t)" frontend_demo: "Inharmonic Speech Segregation" - name: "The Fusion-Texture Correspondence" description: > Kernel fusion combines multiple operations into a single pass, eliminating intermediate materializations. Spectral mixture kernels combine multiple frequency components into a single probabilistic model. Texture perception combines multiple acoustic statistics into a single perceptual quality. All three eliminate intermediate representations in favor of holistic processing. backend_code: "F(linear ∘ silu ∘ mul) → SwiGLU_fused" interface_model: "K(ω) = Σᵢ wᵢ exp(-2π²ℓᵢ²τ²) cos(2πωᵢτ)" frontend_demo: "Model-Matched Sounds" - name: "The Split-K / Cocktail Party Correspondence" description: > Split-K reduction divides a large computation into independent parallel chunks, then merges results. The cocktail party problem requires the auditory system to process a mixture in parallel frequency bands, then merge the results into segregated sources. Both are instances of parallel decomposition followed by coherent recomposition. backend_code: "R(attn_splitk_out, lse_splitk_out) → attn_out" interface_model: "Σ(mixed) → S₁ + S₂ + … + Sₙ" frontend_demo: "Cocktail Party" - name: "The Autoregressive / Jazz Improvisation Correspondence" description: > LLM autoregressive decoding generates tokens one at a time, each conditioned on the full accumulated context (KV cache) and shaped by learned weights (the prior). Jazz improvisation generates notes one at a time, each conditioned on the harmonic context (chord changes + what's been played) and shaped by musical knowledge (the prior). Generative listening (BASS) infers scene decompositions conditioned on the acoustic mixture and shaped by learned source priors. All three are real-time sequential generation under constraints. Inspired by Stephon Alexander's "The Jazz of Physics" (2016), which identifies the group-theoretic structure of Coltrane's harmonic substitutions (Z₃ on Z₁₂) as a production rule on pitch space. backend_code: "fmha_decode(q, k, v) → next token logits" interface_model: "P(noteₜ | chord_context, melodic_prior) → play noteₜ" frontend_demo: "Listening with Generative Models (BASS)" reference: "Alexander, S. (2016). The Jazz of Physics. Basic Books." - name: "The Self-Supervision / Learned Cochlea / Predictive Restoration Correspondence" description: > In deep learning backends, self-supervised representation learning aligns a Student network's predictions with a Teacher network's representations (updated via weight EMA, centered and sharpened to prevent representation collapse). In the audition interface, the Audio Transformer learns its own adaptable time-frequency front-end (sinusoids, windowing, onsets) directly from raw waveforms, replacing fixed spectrograms. In the perceptual frontend, listeners perform part-whole inference, restoring missing fragments of sound masked by noise (illusory continuation). All three represent the system adapting its internal filters to predict global, coherent representations from local, partial observations. backend_code: "align(student, teacher_ema) ∘ prevent_collapse(centering, sharpening)" interface_model: "dense(waveform) → filterbank(sinusoids, windowing, onsets)" frontend_demo: "Illusory Continuation / Spectral Completion" reference: "Verma & Berger (2021)" - name: "The Hankel Eigendecomposition / Cochlear Resonance / Absolute Pitch Correspondence" description: > The Spectral STU uses fixed convolutional filters derived from the eigendecomposition of a Hankel matrix — filters that are never trained but capture the mathematical structure of sequences themselves. The cochlea's basilar membrane provides a physically fixed tonotopic map: its resonant modes are determined by structural mechanics (stiffness gradients, mass distribution), not by neural learning. In perception, absolute pitch — the rare ability to identify frequencies without a reference — suggests the existence of a fixed internal frequency template that is not learned but structurally given. All three represent processing through invariant, structurally determined spectral bases rather than through adaptive, learned representations. backend_code: "eig_vals, eig_vecs = eigh(Hankel(n)); conv(eig_vecs, input)" interface_model: "h(t) = e^{-γt} sin(2πf_c t), f_c determined by basilar membrane position" frontend_demo: "Absolute Pitch / Tonotopic Invariance" reference: "Agarwal et al. (2024); Eguíluz et al. (2000)" - name: "The Objective Optimization / Precision Weighting / Free Energy Correspondence" description: > In the YAMNet optimization loop, the system modifies sound parameters to minimize the distance between current prediction entropy ($H_{cu}$) and a target entropy. In computational psychiatry (Predictive Coding), the brain precision-weights sensory evidence against internal priors to minimize surprise (Free Energy). If sensory evidence is artificially made ambiguous (high entropy), a neurotypical brain remains uncertain, but a maladaptive brain with overly precise priors (e.g., schizophrenia) hallucinates a cause to resolve the uncertainty. In all three cases, the system attempts to resolve a mismatch between expected and actual distributions by either altering the internal model or the external state. backend_code: "loss = |H_cu(θ) - target_H_cu|; minimize(loss, θ)" interface_model: "μ_post = (π_lik μ_lik + π_prior μ_prior) / (π_lik + π_prior)" frontend_demo: "Auditory Hallucinations in Noise" reference: "Boger et al. (2021); Friston (2010)"