# Hemifield sum-difference orientation channels

- ID: 1158
- Canonical URL: https://synthcore.org/idea/1158/hemifield-sum-difference-orientation-channels
- API JSON: https://synthcore.org/api/idea/1158.json
- API Markdown: https://synthcore.org/api/idea/1158.md
- Verification status: unverified
- Source: [arXiv:2607.19681](https://arxiv.org/abs/2607.19681)
- Category: architecture
- Solves: accuracy, sample-efficiency, stability
- ML areas: cnn, transformer, embedding, data-augmentation
- Math tags: harmonic-analysis, geometry, linear-algebra, representation-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Compute separate doubled-angle orientation order parameters for left and right image regions, then expose their sum and difference as symmetric and antisymmetric global features. This gives a network a low-dimensional inductive bias for global vertical structure versus left-right imbalance, while retaining magnitude channels that indicate when either readout is undefined because orientations cancel.

## Mathematical statement

Let c_L=Σ_{j∈L} A_j exp(i2γ_j) and c_R=Σ_{j∈R} A_j exp(i2γ_j), where L and R index inducer or pixel contributions in the left and right hemifields, A_j≥0 is salience, and γ_j∈[0,π) is unoriented orientation. The paper's extension reads the same order parameter through symmetric and antisymmetric combinations: c_+=c_L+c_R and c_-=c_L−c_R. The symmetric channel preserves structure common to both hemifields, while the antisymmetric channel preserves left-right imbalance. A stable neural implementation should use Cartesian coordinates (Re c_±, Im c_±), not only phases, because arg(c_±) is undefined when the corresponding resultant vanishes. Define hemifield confidence values q_±=|c_±|/(Σ_{j∈L∪R} A_j+ε); these distinguish genuine absence of asymmetry from an arbitrary angle caused by cancellation.

## Key formulas

- $$c_L=\sum_{j\in L}A_j e^{i2\gamma_j},\qquad c_R=\sum_{j\in R}A_j e^{i2\gamma_j},$$
- $$c_+=c_L+c_R,\qquad c_-=c_L-c_R,$$
- $$\phi_+=\tfrac12\arg c_+,\qquad \phi_-=\tfrac12\arg c_-,$$
- $$q_\pm=\frac{|c_\pm|}{\sum_{j\in L\cup R}A_j+\varepsilon}$$

## Implementation notes

Place the module after an oriented feature extractor and before global pooling or the final prediction head. Let e_b(x,y)≥0 be the energy in orientation bin θ_b, and define soft left and right masks m_L(x,y)=sigmoid(−βx) and m_R(x,y)=sigmoid(βx), with image-center coordinate x, steepness β, and m_L+m_R approximately equal to 1. Compute z_L=Σ_{x,y,b}m_L(x,y)e_b(x,y)exp(i2θ_b) and z_R=Σ_{x,y,b}m_R(x,y)e_b(x,y)exp(i2θ_b). Form z_+=z_L+z_R and z_-=z_L−z_R, then concatenate Re z_L, Im z_L, Re z_R, Im z_R, Re z_+, Im z_+, Re z_−, Im z_−, q_+, and q_− to the ordinary pooled embedding. Use q_±=|z_±|/(Σ_{x,y,b}e_b+ε), with ε=10^−6. A compact pseudocode block is: E=softplus(oriented_conv(x)); zL=sum(mL*E*exp(2j*theta)); zR=sum(mR*E*exp(2j*theta)); zp=zL+zR; zm=zL−zR; out=MLP([real/imag(zL,zR,zp,zm), abs(zp)/mass, abs(zm)/mass]). The sum/difference construction is fixed by the paper; β, energy extraction, and the downstream MLP are learned or tuned. Start with a synthetic two-bar dataset where labels separately encode common orientation and left-right orientation difference, then test on a small vision transformer or ResNet on rotated Shapes3D or CLEVR-style images. Compare against ordinary global average pooling and concatenated left/right scalar pools. Success is improved accuracy with fewer training examples, robustness to horizontal reflection—sum channels should be invariant while difference channels should change sign—and lower sample complexity at matched parameter count.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.