Summary: Researchers have introduced a new type of artificial neuron called infomorphic neurons. These units independently learn and self-organize with nearby neurons, mimicking the decentralized, local learning that occurs in biological brains. Drawing inspiration from pyramidal cells in the cerebral cortex, infomorphic neurons process local signals to adapt, specialize and cooperate without centralized control.
Each infomorphic neuron uses a novel information-theoretic measure to decide whether to seek redundancy with neighbors, coordinate synergistically, or specialize on unique aspects of input. This local, interpretable learning strategy promises improvements in training efficiency, model transparency, and offers fresh perspectives on how learning may emerge in biological neural tissue.
Key Facts:
- Local learning: Infomorphic neurons learn primarily through interactions with neighboring units, removing the need for external centralized coordination during training.
- Brain-inspired design: The architecture is modeled on pyramidal cells and their compartmentalized processing, capturing aspects of how biological neurons integrate multiple input sources.
- Flexible and interpretable: An information-theoretic framework gives each neuron an explicit, understandable objective—whether to specialize, collaborate, or share information—improving performance and interpretability.
Source: Max Planck Institute
Introduction
Both the human brain and contemporary artificial neural networks are extremely capable, yet they differ fundamentally in how they learn. Artificial networks typically rely on global coordination—algorithms and loss functions applied across the whole network—while biological neurons learn through local interactions with nearby cells. That local, decentralized nature of learning in biological systems contributes to flexibility and energy efficiency that current artificial systems often cannot match.
Typical artificial neural networks are constructed from layers of homogeneous units. Input signals propagate through these layers, and learning is driven by global objectives and backpropagation. In contrast, biological neurons only receive and process signals from a limited local neighborhood. This distinction motivated the development of infomorphic neurons: units that determine their own learning behavior from the local signals they receive.
Infomorphic neurons are designed so that each unit sets its own objectives and adapts its processing accordingly. Instead of being governed entirely by external training protocols, each neuron evaluates information flowing through its compartments and decides whether to emphasize redundancy with neighbors, focus on unique signals, or coordinate to form synergistic representations. That decision is guided by an information-theoretic goal derived from recent advances in Partial Information Decomposition (PID).
By formalizing these local goals, the researchers give neurons the means to discover their own learning rules. “We now directly understand what is happening inside the network and how the individual artificial neurons learn independently,” says Marcel Graetz from CIDBN. This transparency makes it easier to interpret how local dynamics contribute to overall task performance.
Valentin Neuhaus from MPI-DS explains that through specialization and local coordination, infomorphic neurons learn to contribute effectively to network-level tasks. The authors demonstrate that networks built from these units can address problems across supervised, unsupervised and memory learning domains, illustrating the approach’s versatility.
About this AI and learning research news
Author: Manuel Maidorn
Source: Max Planck Institute
Contact: Manuel Maidorn – Max Planck Institute
Image: Image credited to Neuroscience News
Original Research: Open access.
“A general framework for interpretable neural learning based on local information-theoretic goal functions” by Marcel Graetz et al. PNAS
Abstract
A general framework for interpretable neural learning based on local information-theoretic goal functions
Despite the impressive performance of both biological and artificial networks, it remains challenging to form an intuitive account of how local learning dynamics contribute to network-level solutions. Moving learning to a more local scale can yield valuable insights, but a general, constructive approach that produces interpretable local learning objectives adaptable across tasks has been missing.
The authors previously proposed a local information-processing goal suited for a compartmentalized model neuron. Building on recent advances in Partial Information Decomposition (PID), they derive a parametric local learning rule that gives rise to infomorphic neural networks. These networks embody interpretable local objectives and can solve tasks spanning supervised, unsupervised, and memory-based learning. By leveraging the interpretability of the PID framework, infomorphic networks offer a promising tool for probing the structure of local learning and for designing more transparent, flexible artificial systems.