Scandots Representation & Scagnostic Measures

• Scandots representation is a computational method that encodes spatial structures using scan-transcription for 2D shapes and graph-based scagnostics for point clouds.
• The algorithm converts pixel-level boundary data into normalized feature vectors, facilitating similarity evaluation through techniques like SSD.
• This approach achieves translation and size invariance with minimal computational overhead, supporting neuromorphic applications and synthetic dataset generation.

Scandots Representation refers to two distinct yet related computational methodologies for encoding spatial structures: first, as the scan-transcription algorithm for 2D shape boundaries in neuromorphic vision (Greene et al., 2018), and second, as the set of graph-theoretic scagnostic measures for bivariate point clouds (Wanniarachchi et al., 2022). Both approaches exploit the reduction of high-dimensional geometric or statistical structures into concise feature vectors suited for similarity evaluation, classification, or dataset synthesis.

1. Latency Coding and Scan-Transcription of 2D Shapes

The original "scandots" encoding, introduced by Greene & Patel (Greene et al., 2018), encodes a 2D shape as a sparse sequence of pixel-level boundary markers on an $N \times N$ grid. A "scan wave"—mimicking biological retinal wave propagation or achieved via eye drift, stimulus translation, or engineered circuit—progresses across the image, and the temporal order of spike generation at each encountered boundary marker reflects its spatial arrangement.

Latencies in spike generation encode local boundary density along the direction of scanning; counts per scanline (row/column) form the core representational features. This population-based latency coding emphasizes an alternative to orientation-selective filtering, supporting translation and size invariance through simple accumulation and normalization steps. The implementation is motivated by the biological plausibility suggested by the propagation of slow waves in polyaxonal amacrine (PA1) cells, which temporally modulate retinal ganglion cell output.

2. Algorithmic Pipeline for Shape Encoding

The scandots algorithm for 2D shapes operates as follows (Greene et al., 2018):

1. Boundary Discretization: Represent the shape boundary as a set $B = \{ (x_k, y_k)\}$, where each $(x,y) \in \{1…N\}^2$ denotes a pixel on the grid.
2. Scan Accumulation:
   • For each row $i$, $R_i = |\{ (x,y) \in B : y = i \}|$
   • For each column $j$, $C_j = |\{ (x,y) \in B : x = j \}|$
3. Concatenation: Form the raw scan vector $H_\text{raw} = [C_1, …, C_N, R_1, …, R_N]$.
4. Zero Trimming: Remove leading/trailing zeros to focus on the active region, yielding $H_\text{trim}$.
5. Rebinning: Partition $H_\text{trim}$ into $M$ equal-width bins, computing $H_b = \sum_{k=(b−1)L+1}^{bL} H_\text{trim}[k]$, with $L = \text{length}(H_\text{trim})/M$.
6. Normalization: Normalize the feature vector so that $\sum_{b=1}^M H_b = 1$.

Shape similarity is then quantified by the sum-of-squared-differences (SSD) between scandots vectors.

3. Mathematical Characterization

The boundary marker set $B$ is captured by the indicator function $\delta_{xy} = 1$ if $(x,y) \in B$, else 0. The entire pipeline can be formalized as:

$$\begin{align*} R_i &= \sum_{x=1}^N \delta_{x,i}, \quad i=1...N \ C_j &= \sum_{y=1}^N \delta_{j,y}, \quad j=1...N \ H_\text{raw} &= [C_1, ..., C_N, R_1, ..., R_N]^T \ H_b &= \sum_{k=(b-1)L+1}^{bL} H_\text{trim}[k], \quad b=1...M\ \hat{H}_b &= H_b / \sum_{m=1}^M H_m \end{align*}$$

$$SSD(A, B) = \sum_{b=1}^M (\hat{H}_b^A - \hat{H}_b^B )^2$$

Rotation-invariance can be approximated by scanning at multiple angles; in the original work, only $\theta=0^\circ, 90^\circ$ were used (Greene et al., 2018).

4. Implementation, Invariance, and Trade-offs

The method is typically instantiated on a $64 \times 64$ cell grid, using both row and column scans. Nonzero bins are rebinned to $M=20$ feature bins to form the final shape descriptor. The computational overhead is minimal, requiring $O(N)$ operations for scanning and $O(M)$ for rebinning.

Feature vector length $M$ can be tuned, controlling the level of shape detail preserved versus resource constraints; $M=20$ has experimentally sufficed to reproduce human similarity scaling. Size and translation invariance arise directly from the normalization and trimming operations. Rotation-invariance is not inherent but may be implemented by aggregating over several scan orientations.

In neuromorphic hardware contexts, scandots encoding maps naturally to silicon-retina architectures using shift-register or accumulation circuits.

5. Evaluation: Human Similarity Scaling and Comparison to Filter-Based Models

Empirical evaluation involved large-scale similarity scaling (evaluating SSD across all pairs in a 480-shape set) and rapid matching tasks with human observers (Greene et al., 2018). SSD values directly predict the subjective probability of matching ("same") responses even in strongly degraded stimuli presentations. Regression of "same" probabilities against SSD displays a robust negative trend ($p < .0001$). Notably, a significant portion (77%) of low-SSD pairs involved distinct shapes but were still judged similar by humans, highlighting the psychological plausibility of the encoding.

The scandots approach achieves translation and scale invariance without weighted neural pipelines, in contrast to the complexity and data requirements of orientation-selective filtering networks. Biological plausibility is further enhanced by the absence of extensive offline training.

6. Scagnostics: Feature-Fingerprints for Point Clouds

A related "scandots" representation arises in the context of scagnostics for bivariate point clouds (Wanniarachchi et al., 2022). Here, the spatial arrangement of $D = \{p_1,...,p_n\} \subset \mathbb{R}^2$ is summarized by nine standardized graph-theoretic measures:

1. Outlyingness: Longest MST edge length as a fraction of total MST length.
2. Skewness: Absolute mean-median deviation of nearest-neighbor distances, scaled by their standard deviation.
3. Clumpiness: Number of MST-connected components after thresholding, normalized by point count.
4. Sparsity: Proportion of point pairs exceeding a distance threshold.
5. Striation: Maximal correlation between MST edge lengths and edge projection in any direction.
6. Convexity: Ratio of $\alpha$-shape area to convex hull area.
7. Skinny: Deviation from circularity based on perimeter and area.
8. Stringiness: Ratio of MST diameter to number of edges.
9. Monotonicity: Absolute Pearson correlation between $x$ and $y$ coordinates.

The vector $s(D) = [s_1(D),...,s_9(D)]^T$ serves as a shape fingerprint. The scatteR tool solves the inverse problem of generating a point cloud with a prescribed scagnostics profile using Generalized Simulated Annealing (Wanniarachchi et al., 2022).

7. Applications and Broader Significance

Scandots representations afford ultra-fast, translation- and size-invariant shape or point-cloud encoding for tasks in vision, data analysis, and synthetic dataset generation. In neuromorphic settings, scandots can underlie rapid object classification and similarity search with fixed-size, hardware-friendly feature vectors. In statistical graphics and data pedagogy, scagnostics-based scandots are valuable for constructing synthetic datasets with precisely controlled graphical characteristics, supporting benchmarking, teaching, privacy-preserving data synthesis, and data-augmentation.

A plausible implication is that scandots-type representations may bridge the gap between informal visual pattern recognition and formal, algorithmically computable similarity spaces, particularly in domains where shape and topology, not pixelwise detail, are the organizing principle (Greene et al., 2018, Wanniarachchi et al., 2022).