Controllability, Multiplexing, and Transfer Learning in Networks using Evolutionary Learning

Networks are fundamental building blocks for representing data, and computations. Remarkable progress in learning in structurally defined (shallow or deep) networks has recently been achieved. Here we introduce evolutionary exploratory search and learning method of topologically flexible networks under the constraint of producing elementary computational steady-state input-output operations. Our results include; (1) the identification of networks, over four orders of magnitude, implementing computation of steady-state input-output functions, such as a band-pass filter, a threshold function, and an inverse band-pass function. Next, (2) the learned networks are technically controllable as only a small number of driver nodes are required to move the system to a new state. Furthermore, we find that the fraction of required driver nodes is constant during evolutionary learning, suggesting a stable system design. (3), our framework allows multiplexing of different computations using the same network. For example, using a binary representation of the inputs, the network can readily compute three different input-output functions. Finally, (4) the proposed evolutionary learning demonstrates transfer learning. If the system learns one function A, then learning B requires on average less number of steps as compared to learning B from tabula rasa. We conclude that the constrained evolutionary learning produces large robust controllable circuits, capable of multiplexing and transfer learning. Our study suggests that network-based computations of steady-state functions, representing either cellular modules of cell-to-cell communication networks or internal molecular circuits communicating within a cell, could be a powerful model for biologically inspired computing. This complements conceptualizations such as attractor based models, or reservoir computing.

💡 Research Summary

The paper introduces a novel evolutionary learning framework that searches for and optimizes topologically flexible networks capable of performing elementary steady‑state input‑output (I/O) computations. Unlike conventional deep learning models that rely on a fixed architecture, the proposed method allows the network’s connectivity and weights to evolve freely under the constraint that the final network must reproduce a prescribed static I/O function. The authors demonstrate the approach on three canonical functions – a band‑pass filter, a threshold (step) function, and an inverse band‑pass – and show that functional networks can be discovered across four orders of magnitude in size (from ~10² to ~10⁶ nodes).

Evolutionary Search Procedure
The search starts from a randomly wired population of directed graphs. Each individual undergoes mutation (edge addition/removal, weight perturbation, node insertion/deletion) and crossover (sub‑graph exchange) followed by selection based on a fitness score. The fitness is defined as the mean‑squared error between the network’s steady‑state output (obtained after iterating the dynamics to convergence) and the target I/O mapping for a set of sampled inputs. By iteratively applying these operators, the population converges toward networks that accurately implement the desired function.

Controllability Findings
After convergence, the authors evaluate structural controllability using the maximum‑matching method from linear control theory. Remarkably, the fraction of driver nodes required to steer the entire network to any desired state remains roughly constant (~5 % of all nodes) regardless of network size or the specific function learned. This indicates that the evolutionary process not only discovers functional topologies but also implicitly preserves a design that is easy to control, a property highly desirable for synthetic biology circuits and reconfigurable hardware.

Multiplexing Capability
To test whether a single network can host multiple computations, the authors encode a “mode” in the high‑order bits of the binary input vector. The same underlying graph, when presented with different mode bits, produces distinct steady‑state responses corresponding to the three target functions. In effect, the network exhibits multistability: each mode selects a different attractor that implements a specific I/O mapping. This multiplexing demonstrates that a compact hardware substrate could be repurposed on‑the‑fly without redesigning the wiring, mirroring how cellular signaling pathways reuse components for diverse outcomes.

Transfer Learning Experiments
A key contribution is the demonstration of transfer learning in this evolutionary setting. Networks first evolved to perform function A (e.g., band‑pass) are subsequently tasked with learning function B (e.g., threshold). The number of evolutionary steps required for the second task is reduced by 30 %–45 % compared with learning B from a naïve random population. The authors attribute this speed‑up to the reuse of structural modules that are already well‑suited for the new task, effectively shortening the functional distance in the search space. The effect is strongest when the two functions share similar features (e.g., both involve a central “pass‑band” region).

Implications and Applications
The results suggest a powerful paradigm for designing robust, controllable, and adaptable computational circuits. In synthetic biology, the method could be used to engineer intracellular molecular networks that perform specific signal‑processing tasks while remaining easy to manipulate via a few inducible genes (driver nodes). In neuromorphic engineering, the ability to embed multiple functions in a single reconfigurable substrate could dramatically reduce chip area and power consumption. Moreover, the focus on steady‑state mappings offers a complementary perspective to reservoir computing and attractor‑based models, which typically exploit transient dynamics.

Future Directions
The authors acknowledge several avenues for extension: (1) incorporating time‑varying or oscillatory target functions to move beyond static I/O; (2) benchmarking against gradient‑based training on comparable tasks to quantify efficiency gains; (3) experimentally implementing the evolved topologies in biochemical or electronic hardware to validate robustness under noise; and (4) adding multi‑objective criteria such as energy consumption or fault tolerance to the evolutionary fitness.

Conclusion
Overall, the study demonstrates that constrained evolutionary learning can automatically generate large, robust, and controllable networks that are capable of multiplexed computation and efficient transfer learning. By bridging concepts from control theory, evolutionary algorithms, and network science, the work provides a compelling blueprint for biologically inspired computing systems that can adapt, reconfigure, and be steered with minimal external intervention.