# Delay-signature coincidence branches

- ID: 3168
- Canonical URL: https://synthcore.org/idea/3168/delay-signature-coincidence-branches
- API JSON: https://synthcore.org/api/idea/3168.json
- API Markdown: https://synthcore.org/api/idea/3168.md
- Verification status: unverified
- Source: [arXiv:2609.04195](https://arxiv.org/abs/2609.04195)
- Category: architecture
- Solves: accuracy, sample-efficiency, scalability
- ML areas: rnn, ssm, attention
- Math tags: dynamical-systems, probability, approximation-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Add a bank of temporal branches whose synapses apply heterogeneous delays before coincidence pooling. A branch selectively responds to an ordered pair or short sequence when the delay difference compensates the sequence interval, while low delay dispersion makes it primarily an order-invariant event detector. The branch threshold creates sparse, interpretable activations that can replace part of a recurrent or attention-based temporal module.

## Mathematical statement

The paper defines branch calcium accumulation as $$\mathrm{Ca}_{j}=\sum_f w_{jf}\mathbf{1}_{\{|a_{jf}-t_c|<\tau_{\mathrm{w}}\}}$$, where branch $j$ receives synapses indexed by $f$, $w_{jf}$ is the synaptic weight, $a_{jf}$ is the arrival time after axonal delay, $t_c$ is the branch coincidence reference time, and $    au_{\mathrm{w}}$ is the integration-window half-width. Its output is $$I_j=\eta_j g(\max(0,\mathrm{Ca}_j-\theta))$$, where $\eta_j$ is branch-to-soma efficiency, $    heta$ is the calcium threshold, and $g$ is monotone and saturating. For an input at time $t_i$ sent through a synapse with delay $d_{ji}$, use $a_{ji}=t_i+d_{ji}$. For two inputs at times $t$ and $t+\Delta$, coincidence occurs when $|(t+d_{j2})-(t+\Delta+d_{j1})|<    au_{\mathrm{w}}$, equivalently $|d_{j2}-d_{j1}-\Delta|<    au_{\mathrm{w}}$. Therefore a branch detects interval $\Delta$ when its delay-difference signature matches that interval. Sample delays from a distribution with dispersion $\sigma_d$, such as truncated $\mathcal{N}(\mu_i,\sigma_d^2)$, to control the range of detectable intervals. During gradient training replace the indicator by $s_\beta(x)=\operatorname{sigmoid}(\beta x)$ and anneal $\beta$ upward; at inference use the hard indicator. The threshold $    heta$ controls sparse all-or-none activation, and $g(z)=\min(z,I_{\max})$ gives bounded branch amplitude.

## Key formulas

- $$\mathrm{Ca}_{j}=\sum_f w_{jf}\,\mathbf{1}_{\{|a_{jf}-t_c|<\tau_{\mathrm{w}}\}}$$
- $$I_{j}=\eta_{j}\,g\!\left(\max(0,\mathrm{Ca}_{j}-\theta)\right)$$
- $$a_{ji}=t_i+d_{ji},\qquad |(t+d_{j2})-(t+\Delta+d_{j1})|<\tau_{\mathrm{w}}\iff |d_{j2}-d_{j1}-\Delta|<\tau_{\mathrm{w}}$$
- $$\widetilde{\mathrm{Ca}}_j(t_c)=\sum_f w_{jf}\,\sigma\!\left(\beta\left[\tau_{\mathrm{w}}-|a_{jf}-t_c|\right]\right),\qquad \widetilde I_j=\eta_j\min\!\left(\max(0,\widetilde{\mathrm{Ca}}_j-\theta),I_{\max}\right)$$

## Implementation notes

Integrate the module immediately before a temporal encoder or classifier. Inputs are timestamped token or event embeddings $x_i\in\mathbb{R}^D$ with times $t_i$. Instantiate $B$ branches and $F$ delayed synapses per branch. Each synapse has a source projection, a nonnegative delay $d_{jf}$, and weight $w_{jf}$. For each candidate coincidence time $t_c$ on a discretized temporal grid, calculate arrivals $a_{jf}=t_i+d_{jf}$ and accumulate their soft coincidence contributions. Pseudocode: `for branch j: Ca=0; for synapse f: a=t[source[f]]+d[j,f]; Ca += w[j,f]*sigmoid(beta*(tau_w-abs(a-tc))); I[j]=eta[j]*clamp(relu(Ca-theta),0,Imax)`. Use the soft equation during training, anneal `beta` from 10 to 100, and use hard indicators for evaluation. Initialize delays from narrow and broad dispersions, for example $\sigma_d\in\{0,\Delta/2,\Delta,2\Delta\}$, with $\Delta$ covering the task's interval range. Keep delays fixed initially, then compare with projected nonnegative learnable delays. Present isolated events, simultaneous pairs, and ordered pairs at intervals $\Delta$; measure activation sparsity and an order-selectivity index such as $(A_{AB}-A_{BA})/(A_{AB}+A_{BA}+\epsilon)$. The first cheap test should use a small GRU or Transformer on synthetic two-event sequences plus permuted-MNIST or Speech Commands. Compare equal-parameter baselines. Success means higher ordered-sequence accuracy at equal FLOPs, sparse branch activations, and a monotonic increase in order selectivity as $\sigma_d$ grows.

## Disclaimer

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