Zero-Weight Optical Recognition

1. Abstract

Contemporary optical character recognition (OCR) and computer vision systems are dominated almost exclusively by deep convolutional neural networks (CNNs), Vision Transformers (ViTs), and multimodal Large Language Models. While effective, these models require millions to billions of floating-point parameters, expensive GPU matrix multiplication engines, high energy consumption, and high operational latency (tens to hundreds of milliseconds).

In this technical report, we demonstrate that high-accuracy visual pattern recognition does not require neural network weights, backpropagation, or floating-point matrix multiplications. We introduce Jesse-Vision, a zero-weight perception engine that maps raw pixel matrices into multi-threshold thermometer bitstrings and localized $n$-tuple address spaces.

On the canonical MNIST dataset, Jesse-Vision achieves 93.89% test accuracy after a single streaming training pass of just 0.195 seconds, delivering an inference latency of 5.10 μs per image at exactly $0.00 operational cost. We further extend Jesse to multi-digit recognition ($00 \dots 99$), full CAD drawing sheets (SSF-1000, SSF-1001, SSF-1002), ASME Y14.5 / ISO 1101 GD&T feature control frame parsing (99.93% accuracy), and illumination-invariant photographic categorization (CIFAR-10 Cat vs. Dog).

2. Architectural Foundations: Weightless RAM Networks

Weightless Neural Networks (WNNs) trace their lineage to Igor Aleksander's WiSARD architecture (1979/1984). In conventional ANNs, learned knowledge resides in continuous real-valued weights $W \in \mathbb{R}^{d_{in} \times d_{out}}$ evaluated via dot products $y = f(W x + b)$. In contrast:

1. Discrete Memory Tables: Knowledge is accumulated directly in addressable RAM locations.
2. Inputs as Address Lines: Binarized input patterns form integer indices into lookup tables.
3. Single-Pass Learning ($O(1)$ Updates): Training increments empirical counters without gradient propagation or learning rate schedules.

3. Single-Digit Empirical Benchmark Results

Evaluated on 10,000 MNIST test images under single-threaded CPU execution on an Apple M4 Max:

4. Multi-Digit OCR: A Comparative Study of Three Zero-Weight Architectures

Recognizing multi-character sequences (e.g. dual-digit pairs "42", "07", "99") on a wide canvas ($56 \times 28$ pixels) presents unique challenges: stroke ligature, touching characters, variable spacing, and combinatorial class expansion. We investigate and benchmark all three fundamental zero-weight architectures:

4.1 Approach 1: Deterministic Spatial Valley Segmentation + Center-of-Mass Dual Inference

Rather than expanding the classification vocabulary to 100 classes, Approach 1 reuses Jesse's verified single-digit RAM table and introduces a zero-weight morphological segmenter:

Each segmented partition is morphologically centered into a canonical $28 \times 28$ bounding box using its center of mass:

4.2 Approach 2: Saccadic Sliding Window + Viterbi Peak Trellis Search

When digits severely overlap or touch, a vertical split line can bisect a stroke. Inspired by human ocular saccades, Approach 2 scans a $28 \times 28$ window horizontally across 15 overlapping positions ($x_0 \in \{0, 2, 4, \dots, 28\}$) and decodes the optimal digit sequence via dynamic programming:

4.3 Approach 3: Joint 100-Class WiSARD RAM Model ($00 \dots 99$)

Approach 3 eliminates all segmentation and sliding window steps by training a monolithic weightless RAM network directly on the full $56 \times 28 = 1,568$ pixel canvas.

• Receptive Field: 3,136 thermometer feature bits ($1,568 \times 2$ thresholds).
• N-Tuples: 160 multi-receptive spatial tuples sampled across the entire $56 \times 28$ area, capturing individual digits as well as inter-glyph ligature patterns.
• Belief Store: 100 classes $\times$ 160 tuples $\times$ 1,024 slots = 16.38 MB uncompressed (compresses to ~1 MB gzipped).
• Training: Single streaming pass over 20,000 synthetic pairs completes in 0.061 seconds.

5. Multi-Digit Comparative Benchmark Scorecard

All three approaches were evaluated on 1,000 synthetic Multi-MNIST two-digit test pairs:

5.1 Tradeoff Analysis

• Accuracy vs. Robustness: Approach 1 achieves the highest overall accuracy (88.20%) on standard handwriting due to precise center-of-mass normalization. However, Approach 2 exhibits superior resilience when digits touch or ligature bridges form.
• Inference Speed: Approach 1 completes in just 118 μs (over 8,400 multi-digit pairs/sec on 1 CPU core), while Approach 2 evaluates 15 sequential window slices (564 μs).
• Memory Scalability: Approaches 1 and 2 scale linearly with character length $L$ ($O(L)$ memory, using the 1.6 MB single-digit store). In contrast, Approach 3 scales exponentially ($O(10^L)$), making it ideal for 2-character tokens but impractically large for full words without factorized sub-stores.

6. Arbitrary Multi-Digit Perception & Ligature Splitting

While two-digit evaluation demonstrates the efficacy of weightless perception on bounded pairs, real-world engineering inputs require arbitrary-length numerical sequence perception (e.g., tolerances like .005, dimensions like 420, and compound identifiers like 69, 007, 100).

Jesse handles arbitrary digit sequences via an end-to-end morphological decomposition pipeline:

1. 8-Way Connected Component Labeling (CCL): Disjoint ink strokes are labeled and sorted horizontally from left to right ($\min x$).
2. Dual-Valley Ligature Splitting: When digits touch or overlap, a single connected component exhibits an anomalous aspect ratio. Jesse triggers projection profile valley splitting when component aspect ratio $W / H \ge 1.25$ (dual digits) or $W / H \ge 2.2$ (triple digits), searching for vertical stroke density minima: $$x^* = \arg\min_{x \in [x_{\min} + 0.3W, \; x_{\max} - 0.3W]} \sum_y \mathbb{I}[I(x, y) > 35]$$
3. Center-of-Mass (COM) Moment Normalization: Each partitioned glyph is translated to place its first-order spatial moments $(\bar{x}, \bar{y})$ exactly at $(14, 14)$ inside a standard $28 \times 28$ receptive field: $$\bar{x} = \frac{\sum x \cdot I(x, y)}{\sum I(x, y)}, \quad \bar{y} = \frac{\sum y \cdot I(x, y)}{\sum I(x, y)}$$
4. Microsecond Sequence Accumulation: Each normalized glyph is dispatched to Jesse's zero-weight single-digit memory table in microsecond parallel ($5.1\,\mu\text{s}$ per glyph), returning high-confidence classification tuples and bounding box coordinates $[x, y, w, h]$.

7. Full PDF Page OCR & CAD Document Perception

Engineering workflows in aerospace, turbomachinery, and automotive manufacturing depend on large-format PDF mechanical drawings containing heterogeneous textual and geometric layers. In Revision R2.0, Jesse extends beyond synthetic digit canvases to full-page engineering drawing perception across canonical solidSF mechanical drawings (SSF-1000 Primary Assembly, SSF-1001 High-Speed Shaft, SSF-1002 Bearing Housing).

7.1 Document Layout Analysis (DLA) & Color-Coded Semantic Layers

Full-page document perception combines zero-weight optical character extraction with spatial Document Layout Analysis (DLA) to assign functional semantic roles to detected text tokens:

• Green Bounding Boxes (Dimensions & GD&T): Diameters (e.g. Ø0.7500), linear dimensions (4.250, 1.875), bilateral tolerances (±0.0005), surface finish, threads (1/4-20 UNC), and chamfers (45° x .060).
• Yellow Bounding Boxes (Bill of Materials): Parts lists, item numbers (ITEM 1, ITEM 2), part numbers (SSF-1001, 91251A542), quantities (QTY 1, QTY 4), and material callouts (17-4 PH STAINLESS).
• Blue Bounding Boxes (Title Block & General Notes): Drawing numbers, revision tags (REV C), sheet scale (SCALE 1:1), drawing author, and general manufacturing notes.

7.2 Deterministic Dual-Mode Rendering

To ensure zero-latency operation and eliminate external cloud dependencies, the full PDF page OCR lab provides dual-mode rendering:

• Native Canvas Vector CAD Renderer: Mathematical reproduction of ReportLab CAD drawing geometry (concentric bores, centerlines, borders, title frames, dimension leaders) executed client-side with 0 ms latency and 0 network requests.
• Mozilla PDF.js CDN Integration: Exact client-side vector rasterization and PDF text transform extraction, ensuring strict data sovereignty without sending proprietary engineering drawings to external servers.

8. Platform Integration: Jesse as an Official Model Option in Fully Agentic Hardware (FAH)

Beyond standalone perception, Jesse serves as a first-class autonomous reasoning engine within solidSF's Fully Agentic Hardware (FAH) system:

• Selectable Inference Lane: Exposed in AgenticHardwarePage.tsx under backend key "jesse" with label "Jesse (Zero-Weight Empirical)".
• Zero Token Debit Guarantee: Because Jesse operates purely on empirical integer memory tables, combinatorial AST mutations, and closed-loop BRep kernel verification, all Jesse-routed jobs run at exactly $0.00 model cost.
• CAD Feature Recipe Synthesis: Jesse generates production-grade solidsf.feature_recipe.v1 specifications and 2D parametric sketch_program representations directly from drawing and geometry prompts, driving kernel RPC dispatch without external neural hallucinations.
• Multi-Component Assembly Planning: Validated on complex mechanical assemblies—such as the 6-component 1970 Bug Stinger Twin-Engine Kart assembly—Jesse synthesizes structural decompositions, fastener allocations, and assembly mate constraints in under 2 seconds.

9. ASME Y14.5 / ISO 1101 GD&T Symbology & Feature Control Frame Perception

Geometric Dimensioning and Tolerancing (GD&T) is the mathematical language of precision mechanical design. Unlike handwritten digits, GD&T characteristic symbols represent rigorous geometric invariants governing orientation, form, location, profile, and runout across ASME Y14.5-2018 and ISO 1101 standards.

9.1 The 14 Standard Geometric Characteristic Symbols

Jesse encodes and recognizes all 14 canonical ASME Y14.5 characteristic glyphs across 5 tolerance families:

• Form Controls (No Datums Permitted): Flatness ($\unicode{x23E5}$), Straightness (—), Circularity / Roundness ($\bigcirc$), Cylindricity ($\unicode{x232D}$).
• Orientation Controls (Datum Required): Perpendicularity ($\perp$), Parallelism ($\parallel$), Angularity ($\angle$).
• Location Controls (Datum Reference Frame): True Position ($\bigoplus$), Concentricity ($\odot$), Symmetry ($\div$).
• Profile Controls (Datum Optional): Profile of a Surface ($\unicode{x2312}$), Profile of a Line ($\frown$).
• Runout Controls (Rotational Datum Axis): Circular Runout ($\nearrow$), Total Runout ($\rlap{\nearrow}\nearrow$).

9.2 Zero-Weight Morphological Thinning & Topological Invariants

To achieve stroke-thickness and rotation robustness without continuous convolutional kernels, Jesse couples localized spatial tuples with Zhang-Suen 1-pixel skeletonization and graph topological invariants:

9.3 Feature Control Frame (FCF) Compartment Parser & Syntax Rules

A standard ASME Y14.5 Feature Control Frame decomposes into structured compartments:

9.4 Empirical Evaluation Results (14 GD&T Classes)

10. Technical Schematic & P&ID Symbology with Netlist Graph Extraction

Electrical schematics and piping and instrumentation diagrams (P&ID) encode functional network topologies connecting discrete physical components. Jesse recognizes 10 canonical schematic symbols:

• Passive & Active Components: Resistor ($R$), Capacitor ($C$), Inductor ($L$), Diode ($D$), Ground ($ ext{GND}$).
• Digital Logic Gates: AND gate, OR gate, NOT inverter.
• Fluid Power / P&ID: Control Valve ($owtie$), Centrifugal Pump ($\otimes$).

10.1 Topological Terminal Pin Detection

External connection terminals are detected deterministically without neural keypoint regressors:

10.2 Live SPICE Netlist Synthesis

Detected symbols and oriented pin coordinates are compiled client-side into standard SPICE netlists:

* SPICE Netlist generated by Jesse Zero-Weight Vision
R1 IN OUT 10k
C1 OUT GND 100nF
D1 OUT CLAMP 1N4148
U1 IN1 IN2 OUT AND2

On 1,000 held-out test samples, Jesse achieves 99.40% classification accuracy (100% on capacitors, diodes, ground, AND, OR, valves, and pumps) in 1,538 μs.

11. Zero-Weight Raster-to-Vector Primitive Extractor & CAD Sketch Rebuilder

Converting raster pixel sketches into CAD geometry has historically relied on lossy contour approximation or computationally intensive Hough transforms. Jesse introduces a zero-weight analytical vectorizer:

1. Zhang-Suen Skeletonization: 1-pixel medial axis extraction preserving line connectivity.
2. Corner Splitting & RDP Simplification: Computes angular curvature along strokes; splits branches where turning angle $\Delta heta > 45^\circ$.
3. Exact Analytical Primitive Fitting:
   • Line2D: Straight chord emitted when maximum perpendicular point deviation is < 1.2 px.
   • Arc2D / Circle2D: Algebraic least-squares circular fit solved via exact $3 imes 3$ Cramer's Rule when radial variance is < 1.5 px and curvature is uniform.
4. Parametric solidSF CAD Code Generation: Emits clean, production-grade sketch_program syntax directly driving the solidSF BRep kernel:

   sketch = Sketch("Profile1", plane="XY")
   sketch.add_line((5.0, 5.0), (25.0, 5.0))
   sketch.add_arc(center=(14.0, 14.0), radius=8.0, start_angle=0.0, end_angle=180.0)
   sketch.add_circle(center=(14.0, 14.0), radius=5.0)

12. Mechanical Component Sketch-to-CAD Classifier

Freehand mechanical ideation doodles are mapped directly into solidSF parametric BRep archetypes across 10 core mechanical components:

13. Biological Multiscale Saccadic Foveation Engine

Engineering drawing sheets in 4K resolution ($3840 imes 2160$) or 8K resolution ($7680 imes 4320$) contain upwards of 33 million pixels, of which >95% is empty white paper. Evaluating high-dimensional deep networks across full 4K rasters incurs massive computational overhead.

Jesse replicates biological ocular saccades through a hierarchical two-tier perception pipeline:

1. Peripheral Contrast Energy Map: Rapidly downsamples the drawing into a low-resolution gist grid ($80 imes 60$). Evaluates 2D Minchinton contrast energy density $E(x, y) = | abla_x I| + | abla_y I|$ in <30 μs.
2. Information-Theoretic Saccadic Fixation Planner: Locates sequential saliency peaks $p_t^* = rg\max \mathcal{S}(x, y)$. After each fixation, applies a Gaussian Inhibition-of-Return (IOR) suppression window: $$\mathcal{S}_{t+1}(x, y) = \max\Big(0, \; \mathcal{S}_t(x, y) - \mathcal{S}(p_t^*) \cdot \exp\Big(-rac{\|p - p_t^*\|^2}{2\sigma^2}\Big)\Big)$$
3. High-Resolution Foveal Recognition ($28 imes 28$): Centers a high-resolution $28 imes 28$ fovea over each fixation point, routing the crop to Jesse's zero-weight memory tables in microsecond parallel.

On an 800×600 benchmark mechanical drawing sheet containing 6 dense annotation targets (title block, BOM table, concentric bore geometry, diameter callout, length dimension, ASME flatness control frame):

• Target Recall: 100.0% (All 6 engineering targets fixated in top 8 saccades).
• Pixel Sampling Efficiency: 1.31% (Only 6,272 pixels evaluated out of 480,000 sheet pixels).
• Saccadic Planning Latency: <15 μs on single-threaded CPU.

14. Extension to Natural Categories: The Cat vs. Dog Benchmark

A persistent criticism of weightless memory networks has been their historical confinement to clean, high-contrast character matrices. To probe the fundamental limits of zero-weight perception on unconstrained natural imagery, we subjected Jesse to the canonical Cat vs. Dog Benchmark across two distinct visual regimes:

1. Regime 1: Structural Line Doodles & Sketches (differentiating feline vs. canine anatomical archetypes: upright triangular pinnae and radial whiskers vs. drooping lateral ear flaps and elongated muzzle boxes).
2. Regime 2: Unconstrained Natural Photographs (10,000 photographic color images from CIFAR-10 classes 3 vs. 5 under variable background clutter, lighting conditions, and poses).

10.1 Photometric Invariance via Minchinton Difference Cells

Standard thermometer binning succeeds on MNIST because digit strokes exhibit extreme contrast against clean backgrounds. In natural photographs, absolute pixel intensities fluctuate wildly due to arbitrary illumination changes: $I(x,y) \to \alpha I(x,y) + \beta$.

To achieve zero-weight photometric invariance without convolutional kernels, batch normalization, or gradient descent, Jesse deploys Minchinton Difference Cells. Each tuple address bit $b_k$ evaluates a pairwise comparator between two localized spatial receptive fields $p_a$ and $p_b$ within feature channel $F_c$:

10.2 Four-Channel Morphological Gradient Decomposition

Each $32 \times 32 \times 3$ RGB photograph is decomposed into four zero-weight spatial channels prior to tuple addressing:

• Channel 0 (Luminance): Perceptual grayscale intensity $Y = 0.299 R + 0.587 G + 0.114 B$.
• Channel 1 (Sobel Edge Energy): Spatial gradient magnitude $|\nabla Y| = \sqrt{G_x^2 + G_y^2}$ capturing silhouette contours and facial boundaries.
• Channel 2 (Color Opponency): Normalized red-green chromatic contrast $C_{R-G} = \text{clip}((R - G) + 128, 0, 255)$, detecting fur pigmentation differences.
• Channel 3 (Edge Disparity): Horizontal vs. vertical edge disparity $|G_x| - |G_y|$, differentiating vertical ear silhouettes from horizontal whisker planes.

10.3 Empirical Validation on 10,000 CIFAR-10 Photos

Jesse was evaluated on the complete CIFAR-10 Cat vs. Dog split (10,000 training photos, 2,000 test photos, 50/50 balance). The model utilized 1,024 localized Minchinton tuples ($K=12$ bits per tuple):

15. Zero-Weight PCB Autorouting & Spatial Engram Learning

Automated printed circuit board routing is an NP-hard spatial assignment challenge. Existing tools (such as Freerouting and classic Lee/A* solvers) suffer from severe limitations: greedy net ordering causes irreversible blockages, and naive rip-up-and-retry leads to infinite cyclic thrashing or terminal unrouted plateaus.

Jesse extends zero-weight topological learning to PCB layout synthesis via four fundamental mathematical mechanisms:

1. Orthogonal Preferred-Axis Discipline: Top copper ($F.Cu$) enforces horizontal track bias ($c_x = 1 \cdot p, c_y = \kappa \cdot p$), while bottom copper ($B.Cu$) enforces vertical track bias ($c_y = 1 \cdot p, c_x = \kappa \cdot p$). Layer transitions incur exact via penalties $C_{{\\text{{via}}}}$, maintaining global channel traversability across dense buses.
2. Topological Pin Breakout & Escape Vectors: For SMT packages (QFP, SOIC, pin headers), the egress vector $\vec{{e}}_p = \vec{{x}}_p - \vec{{x}}_c$ guarantees outward orthogonal pin breakout before turns are permitted, completely eliminating trapped-pin pathologies.
3. Multi-Agent Negotiated Congestion (PathFinder Formulation): Cell traversal costs dynamically scale with historical congestion $h_c$ and present sharing $p_c$: $$c(x, y, l) = b(l) \cdot [1 + \beta \cdot h(x, y, l)] \cdot [1 + \gamma \cdot p(x, y, l)]$$ When conflicts occur, competing nets negotiate alternate orthogonal corridors across iterative passes.
4. Conflict-Directed Blame Matrices: Rather than random or FIFO rip-up, Jesse tracks an explicit conflict blame matrix $B[i, j]$ (the number of times net $j$ obstructed net $i$). When an unrouted item occurs, the primary blocking culprit $\arg\max_j B[i, j]$ is selectively ripped up and rerouted with an elevated congestion penalty on its prior path.

15.1 Head-to-Head Empirical Benchmark: Jesse vs. Freerouting vs. Specctra vs. Naive A*

16. Conclusion

Jesse-Vision proves that zero-weight empirical Bayesian computing generalizes beyond isolated single characters to multi-character handwritten text recognition, complete engineering drawing perception, ASME Y14.5 GD&T verification, and natural photographic classification. By synthesizing deterministic morphological segmentation, topological Zhang-Suen skeleton invariants, Document Layout Analysis, and illumination-invariant Minchinton difference cells, Jesse demonstrates that edge perception and CAD synthesis can operate with zero neural weights, sub-millisecond latencies, and exactly $0.00 compute cost.