Bio-Inspired Bayesian Expected Information Gain

• Bio-Inspired Bayesian Expected Information Gain is a framework combining Bayesian inference with biological sensing principles to maximize information extraction.
• It employs KL divergence and photonic hardware models to optimize sensor settings through orthogonal, complete, and sparse encoding strategies.
• The unified optimization approach demonstrates high spectral resolution and adaptive sensing, linking theoretical metrics with practical device performance.

Bio-inspired Bayesian expected information gain (EIG) refers to a unifying quantitative framework, derived from both the computational and biological sciences, for optimizing sensing, learning, and decision-making processes. Central to the framework is the principle that agents or systems—biological or artificial—can adapt their encoding strategies, device architectures, or selective actions to maximize the average information they gain about hidden variables in their environment. This concept, rigorously formalized through information-theoretic constructs such as the Kullback–Leibler divergence and operationalized in hardware and active inference, seeks to bridge the gap between theoretical optimality and bio-inspired design principles for autonomous information acquisition (Zhang et al., 18 Jan 2026, Sajid et al., 2021, Yanagisawa et al., 2023).

1. Foundations: Bayesian Expected Information Gain

The expected information gain under a Bayesian framework is formally defined as the expected Kullback–Leibler (KL) divergence between the posterior and prior beliefs about a latent variable after observing a new datum or measurement. For a spectrometer, the unknown input spectrum $f \in \mathbb{R}^{M_\lambda}$ is measured via a photonic encoder parameterized by tunable settings $\theta \in \Theta$, yielding

$h = G(\theta) f + \eta,$

where $G(\theta) \in \mathbb{R}^{N \times M_\lambda}$ is the response matrix and $\eta$ is measurement noise. The optimal setting $\theta^\ast$ maximizes

$$\mathrm{EIG}(\theta) = \mathbb{E}_{h|\theta} \big[ D_{\mathrm{KL}}(p(f|h,\theta) \parallel p(f)) \big],$$

with equivalent expressions using joint and marginal probability densities, or in terms of entropy reduction. In a linear-Gaussian regime, with $p(f) = \mathcal{N}(0, \Sigma_f)$ and $$p(h|f,\theta) = \mathcal{N}(G(\theta) f, \Sigma_\eta)$$, the expected information gain reduces to an analytic closed form,

$$\mathrm{EIG}(\theta) = \tfrac{1}{2} \log \det \left[ I + \Sigma_\eta^{-1} G(\theta) \Sigma_f G(\theta)^\top \right],$$

facilitating computational optimization and implementation (Zhang et al., 18 Jan 2026).

2. Bio-Inspired Principles: Orthogonality, Completeness, Sparsity

Drawing inspiration from sensory processes in marine mammals, bats, and nematodes, three core attributes define effective information-encoding paradigms:

• Orthogonality: Distinct, non-overlapping encoding channels suppress noise amplification and permit rapid convergence, analogous to marine mammal sonar.
• Completeness: The capacity to represent any possible input within the operational bandwidth, as observed in bat echolocation and multimodal biological sensing.
• Sparsity: Selective activation of only the most informative pathways for energy efficiency and denoising, paralleling the signal compression mechanisms of C. elegans.

These attributes are embedded at the physical hardware level, rather than being enforced by post-processing, thereby mimicking the autonomy and robustness of natural systems (Zhang et al., 18 Jan 2026).

Quantitative metrics directly measure these bio-inspired properties:

• Condition number ($\mathrm{cond}(G)$) for orthogonality: $$\mathrm{cond}(G) = \|G\|_2 \|G^{-1}\|_2 = \sigma_{\max}/\sigma_{\min}$$, with a small value indicating well-conditioned, non-redundant encoding.
• Completeness error ($\varepsilon_{\mathrm{comp}}$): The supremum over the residuals when projecting arbitrary unit-norm spectra onto the span of $G$.
• Mutual coherence ($\mu(G)$) for sparsity: $$\mu(G) = \max_{i \neq j} |\langle g_i, g_j \rangle| / (\|g_i\|_2 \|g_j\|_2)$$; smaller values imply better fidelity for sparse reconstructions (Zhang et al., 18 Jan 2026).

3. Unified Multi-Objective Optimization

Bio-inspired Bayesian EIG is operationalized by constructing a single optimization that integrates the pure information gain objective with explicit penalties or trade-offs for the three biological attributes:

$$\theta^\ast = \arg \max_{\theta \in \Theta} \left[ \mathrm{EIG}(\theta) - \lambda_{\mathrm{orth}} \mathrm{cond}(G(\theta)) - \lambda_{\mathrm{comp}} \varepsilon_{\mathrm{comp}}(\theta) - \lambda_{\mathrm{sp}} \mu(G(\theta)) \right]$$

subject to physical feasibility (e.g., microheater power limits). Here, $\lambda_{\cdot}$ denote tunable weights. In the linear-Gaussian regime, all quantities involved are analytically or numerically tractable, enabling direct hardware realization without post hoc algorithmic correction (Zhang et al., 18 Jan 2026).

4. Physical Realization in Photonic Hardware

The hardware instantiation of this framework leverages a reconfigurable silicon photonic device consisting of a cascaded high-Q Mach-Zehnder-interferometer–assisted microring resonator (MRR I) and a Vernier-tuned secondary ring (MRR II). Three TiN microheaters $(P_1, P_2, P_3)$ serve as control parameters $\theta$:

• $P_1, P_2$ adjust resonance linewidths for orthogonality control;
• $P_3$ induces resonance shifts (Vernier effect), manipulating completeness and sparsity attributes.

By discretizing $P_3$ and sampling across wavelength, the system generates a diverse library of response matrices $G$. Constraints from thermal coefficients and device physics are captured via calibrated transfer functions, derived from both electromagnetic (Lumerical FDTD/INTERCONNECT) simulations and experimental calibration (Zhang et al., 18 Jan 2026).

A table summarizing the mapping from hardware parameters to bio-inspired metrics:

Attribute	Physical Parameter(s)	Metric in G
Orthogonality	$P_1$, $P_2$	$\mathrm{cond}(G)$
Completeness	$P_3$ (Vernier shift)	$\varepsilon_{\mathrm{comp}}(G)$
Sparsity	$P_3$ (Vernier shift)	$\mu(G)$

5. Experiments, Performance, and Adaptivity

The photonic device's performance was validated under benchmark conditions spanning linewidths and encoding regimes. Key experimental results include:

• Ultra-high spectral resolution: Resolving two continuous-wave laser lines with separations down to $6$ pm.
• Continuous spectra: High-fidelity (R²≥0.994, RMS error ≤0.08) reconstruction of both Gaussian and complex spectra using convex regularized least-squares solvers.
• Broad bandwidth: Measurable span surpasses 30 nm, employing the Vernier effect and randomized encoding masks with low cross-correlation (average autocorrelation ≈0.22).
• Response function diversity: Randomization of $P_1, P_2$ yields a heterogeneous library of 200 encoding masks, covering a broad spectrum of orthogonality, completeness, and sparsity traits. The device can adapt on-the-fly to unknown inputs at the hardware level—no software re-tuning required (Zhang et al., 18 Jan 2026).

6. Theoretical Context: EIG, Free Energy Principle, and Active Inference

Within computational neuroscience, Bayesian EIG emerges as a limiting case of the expected free energy (EFE) in the active inference framework. Removal of explicit outcome preferences from the EFE reduces the objective to pure EIG maximization, which formalizes “curiosity-driven” or intrinsically motivated behavior. Conversely, removal of ambiguity and risk yields classic expected utility maximization. Only the full EFE objective, which combines intrinsic and extrinsic drives, replicates the balanced exploration-exploitation trade-off characteristic of biological agents (Sajid et al., 2021).

The decomposition:

• Risk/ambiguity: Intrinsic information-seeking pressure (EIG).
• Extrinsic value: Goal-directed, utility-maximizing behavior.

In decision-making and design, bio-inspired Bayesian EIG thus connects optimization for information gain (with biologically motivated constraints) to broader theories of adaptive control and inference in living systems.

7. Bio-Inspired EIG in Inquiry and Epistemic Emotion

Recent mathematical models formalize the role of Bayesian information gain in driving epistemic emotions such as curiosity and interest. Two core gains—free-energy reduction (KLD) and Bayesian surprise (BS)—together form a convex (inverted-U or “Wundt curve”) function of prediction error (surprise). Cyclic alternation between maximizing KLD (confirmation/consolidation) and BS (novelty exploration) yields an “ideal inquiry cycle” that dynamically maintains optimal arousal and exploration. Sensitivity to prediction and observation uncertainty offers a prescription for maximizing information gain consistent with both bio-inspired and active-inference perspectives (Yanagisawa et al., 2023).

The inquiry cycle regulates the agent’s attentiveness and open-mindedness by tuning prior and likelihood variances, mapping directly onto the policies of information acquisition in physical sensors or artificial agents.

---

Collectively, bio-inspired Bayesian expected information gain provides a comprehensive framework linking information-theoretic optimality, biomimetic engineering, active perception, and epistemic emotion into a mathematically grounded approach for adaptive sensing, learning, and inference. Its practical realization in dynamic spectrometers demonstrates the possibility of embedding such principles directly in physical hardware, setting a precedent for future compact, autonomous, and high-fidelity sensory systems.