Neuro-AI Parallels Database

An Oracle SQL database mapping structural, functional, and computational parallels between biological brain structures and artificial neural-network components — built to identify where neuroscience is mature enough to directly inform AI design.

Stack: Oracle SQL (23ai) · 14-table 3NF schema · Status: Research-grade knowledge base

Problem

Most neuro-AI research runs in one direction: AI serving neuroscience — imaging, drug discovery, diagnostics. This project explores the reverse.

The human brain is the most energy-efficient, adaptive, general intelligence system known, and decades of neuroscience describe precisely how it is built. The argument behind this database is that biological knowledge should feed back into AI architecture design, not only be analysed by AI.

The database answers one precise question:

Where are the biological solutions mature enough, and the artificial equivalents weak enough, that the brain could serve as a design template for next-generation AI?

Overview

A normalized relational knowledge base linking documented brain-to-AI parallels to the real published papers that establish them, the researchers and institutions behind them, and the funding sources that may bias them.

Entity	Count
Research papers (real, published)	57
Documented parallels	20
Biological structures	16
Artificial structures	15
Authors	20
Institutions	15
Funding sources	8
Tables	14

Every parallel is organized under six functional categories spanning the cognitive-computational stack.

Category	Biological side	Artificial side
Perception & Encoding	Visual cortex hierarchy (V1→IT)	Convolutional neural networks
Attention & Selectivity	Prefrontal & parietal cortex	Transformer attention heads
Memory & Storage	Hippocampus, sleep replay	LSTM, continual learning with replay
Learning & Adaptation	Synaptic plasticity (STDP), Hebbian	Backpropagation, plastic weights
Decision & Prediction	Dopaminergic system, predictive coding	TD learning, variational autoencoders
Output & Generation	Primary motor cortex	Sequence-to-sequence RNNs

Method

14 tables, normalized to 3NF, with PARALLEL as the central entity through which every paper, biological structure, and artificial structure connects.

FUNCTIONAL_CATEGORY
       ├──< BIO_STRUCTURE_CATEGORY >── BIOLOGICAL_STRUCTURE
       ├──< ART_STRUCTURE_CATEGORY >── ARTIFICIAL_STRUCTURE
       └──────── PARALLEL ────────┐
                     │            │
              PAPER_PARALLEL      │
                     │            │
                  PAPER ──< PAPER_AUTHOR >── AUTHOR
                     │                         │
              PAPER_FUNDING            AUTHOR_INSTITUTION
                     │                         │
              FUNDING_SOURCE             INSTITUTION

Design decisions:

• parallel_type is constrained to Structural / Functional / Computational — a deliberate taxonomy separating same-architecture parallels from same-role and same-math ones.
• confidence_level on PARALLEL reflects evidence quality, not editorial opinion.
• potential_bias on FUNDING_SOURCE makes funding analysis a first-class feature, not an afterthought.
• Junction tables use composite primary keys — no surrogate IDs where they are not needed.
• All inserts resolve foreign keys via subquery name lookups rather than hardcoded IDs, keeping the script portable across re-runs.

Schema surface: 14 primary keys, 18 foreign keys, 9 unique and 13 check constraints, ~370 rows. The SELECT layer exercises multi-table joins, correlated and inline-view subqueries, group functions with HAVING, set operators, and conditional expressions.

Results

Three views carry the analytical conclusions.

• ACTIONABLE_PARALLELS — for each parallel: supporting-paper count, methodology diversity, industry-funding count, and a design-insight preview. Filtered to High and Medium confidence. This is the intellectual endpoint of the database.
• FUNDING_BIAS_REPORT — per parallel, counts supporting papers by industry / government / non-profit source and assigns a bias-risk label. Answers: can this parallel be trusted, or is the evidence commercially biased?
• RESEARCHER_PROFILE — every author with institution type, primary domain (Neuroscience / AI / Both), h-index, and contribution count.

A selection of the most precisely established parallels:

• Dopamine → Temporal Difference Learning (Computational, High) — dopamine neurons fire as the TD error term predicts: phasic activation for unexpected reward, silence for predicted, pause for omitted. Among the most precisely quantified neuro-AI parallels. Schultz, Dayan & Montague (1997)
• V1 Simple Cells → Convolutional Filters (Structural, High) — CNNs trained on object recognition converge on oriented edge detectors identical to V1 receptive fields, discovered empirically rather than by design. Yamins et al. (2014); Olshausen & Field (1996)
• Sleep Replay → Continual Learning with Replay (Functional, High) — biological consolidation via compressed generative replay; the design insight is that AI replay should be generative and compressed, not raw stored samples. Tadros et al. (2022)
• Synaptic Plasticity (STDP) → Backpropagation (Computational, High) — STDP updates weights via local spike timing; backprop needs a biologically implausible global error signal. The open problem: local rules that approximate backprop without global error propagation. Schiess et al. (2016); Miconi et al. (2018)

Reproduce

git clone https://github.com/ayadilara10/NeuroAIParallels.git
cd NeuroAIParallels/sql

Run against Oracle (tested on LiveSQL / FreeSQL 23ai), in order:

01_create_tables.sql        # DDL — 14 tables with constraints
02_alter_drop.sql           # DDL — ALTER / DROP examples
03_insert_data.sql          # DML — full dataset
03b_insert_continuation.sql
04_update_delete.sql        # DML — UPDATE / DELETE with verification
05_select_queries.sql       # 36 queries across 7 categories
06_views.sql                # the three analytical views

Expected output: a populated 14-table schema; SELECT * FROM actionable_parallels returns High/Medium-confidence parallels ranked by supporting-paper count.

Roadmap

Each phase states the outcome it unlocks.

Now — A research-grade structured knowledge base exists. Parallels, evidence, and funding bias are queryable in SQL.

Next — Researchers query it without writing SQL. The knowledge base is usable by its actual audience.

• Export to PostgreSQL and expose a REST API (parallel map, paper evidence, bias analysis).
• Searchable web interface: e.g. "High-confidence Structural parallels in Attention with no industry-funded papers."

Later — The map stays current without a single maintainer. The resource is living, not a snapshot.

• Community contribution layer: researchers submit new parallels with citations, reviewed before inclusion — a peer-reviewed map of the neuro-AI intersection.

Structure

NeuroAIParallels/
├── sql/
│   ├── 01_create_tables.sql       DDL — 14 tables, constraints
│   ├── 02_alter_drop.sql          DDL — ALTER / DROP
│   ├── 03_insert_data.sql         DML — dataset
│   ├── 03b_insert_continuation.sql
│   ├── 04_update_delete.sql       DML — UPDATE / DELETE
│   ├── 05_select_queries.sql      36 SELECT queries
│   └── 06_views.sql               analytical views
└── README.md

Conventions

• Comments: -- intent-comments on non-obvious query logic and each constraint's purpose.
• Foreign keys resolved by subquery name lookup, never hardcoded IDs — portable across re-runs.
• Confidence and bias fields reflect evidence, not opinion.
• All 57 papers, authors, institutions, and DOIs are real published research.