A Foundational Theory for Decentralized Sensory Learning

TL;DR

The paper addresses how biological learning could originate and be sustained without global error signals, proposing that learning emerges from negative feedback-driven minimization of sensor problem signals across scales—from unicellular homeostasis to multicellular communication and neural networks. The core approach frames sensors as homeostatic problem signals and posits that neurons propagate and minimize these problems in a decentralized fashion, enabling local learning through conjectures that obviate backpropagation or global rewards. Key contributions include a scalable narrative from single cells to brain networks, a set of conjectures tying intercellular signaling to learning, and a linkage to classical control and plasticity theories, offering a unified interpretation of learning as problem-solving. The work suggests broad significance for neuroscience and AI, proposing that decentralized, problem-driven learning could inform new interpretations of brain data and guide AI systems toward local, self-organizing learning rules grounded in homeostasis.

Abstract

In both neuroscience and artificial intelligence, popular functional frameworks and neural network formulations operate by making use of extrinsic error measurements and global learning algorithms. Through a set of conjectures based on evolutionary insights on the origin of cellular adaptive mechanisms, we reinterpret the core meaning of sensory signals to allow the brain to be interpreted as a negative feedback control system, and show how this could lead to local learning algorithms without the need for global error correction metrics. Thereby, a sufficiently good minima in sensory activity can be the complete reward signal of the network, as well as being both necessary and sufficient for biological learning to arise. We show that this method of learning was likely already present in the earliest unicellular life forms on earth. We show evidence that the same principle holds and scales to multicellular organisms where it in addition can lead to division of labour between cells. Available evidence shows that the evolution of the nervous system likely was an adaptation to more effectively communicate intercellular signals to support such division of labour. We therefore propose that the same learning principle that evolved already in the earliest unicellular life forms, i.e. negative feedback control of externally and internally generated sensor signals, has simply been scaled up to become a fundament of the learning we see in biological brains today. We illustrate diverse biological settings, from the earliest unicellular organisms to humans, where this operational principle appears to be a plausible interpretation of the meaning of sensor signals in biology, and how this relates to current neuroscientific theories and findings.

Figures (4)

• Figure 1: The multifunctional cell. Already the unicellular organism can have multimodal sensing and at least two basic forms of actuation, mechanical and chemical actuation. The primary problem of the cell is its energy need. Energy is required to drive ion pumps, which in turn indirectly drive all transportation across the cell membrane and enable the cell to maintain its internal biochemical equilibrium. Molecules taken in are used in a multitude of metabolic processes, and the cell regulates its gene expression in order to provide the enzymes accelerating those molecular cascade reactions. Chemical sensing is achieved by receptor molecules, typically proteins, and can impact those cascade reactions. A cell can also be capable of transporting chemical signals out across its membrane, for example via exocytosis (‘chemoactuation’). It can also sense mechanical stress in its cytoskeleton, which can be coupled to mechanosensing channels Brunet2016-cj. The latter, when activated, can let in some ions, typically calcium, which can mobilize contraction in the cytoskeleton, but also the cascade reactions. This can lead to cell shape changes, or even propulsion, if the cell in addition has a cilium. Some cells even have light sensors, where the phototransduction into biochemical energy can end up changing for example the calcium level in the cell, thereby driving actuation. These functional components of the cell are known from basic cell biology Alberts2022-ymArendt2008-bj.
• Figure 2: Adaptation of intracellular molecular input-output reactions in unicellular organisms. This is a depiction of an experiment designed to illustrate amoebal behavioral adaptation and behavioral memory. The ‘task’ was to learn to overcome a barrier of a potentially harmful chemical to be able to find nutrients. The amoeba was shown to be able to solve the task where the cell learns to overcome a natural aversive reaction from the chemical in order to solve the greater problem consisting of cell-internal sensing of a lack of nutrients. Interestingly, once the amoeba cell had learned to solve this task, it would solve it much more rapidly the next time it was exposed to the same task and the adaptation was moreover specific to the noxious chemical used in the barrier. This illustrates that the input-output reactions of the intracellular molecular cascade reactions are adaptable and that the behavioral adaptation is stored like a memory. (i) The cell has a nutritional need. This drives activation of the cytoskeleton (Figure \ref{['fig1']}) such that the cell changes shape (mechanoactuation) and sends out filopodia that tries to find food. The filopodia hit the chemical barrier, which tends to inhibit its further travel since the cell has membrane surface receptors for this chemical and the activation of those receptors inhibits the extension of the filopodia. (ii) The cell-internal sensing of a lack of nutrients soon overcomes the blocking effect of the barrier and the filopodia instead reaches the food particle. (iii) The whole cell is migrating across the barrier to more rapidly ingest the nutrients. Note that this experiment also illustrates the principle of sensory prioritization by the cell, which we address in this paper. (Figure adapted from Tang and Marshall, 2018 Tang2018-po).
• Figure 3: Networks of chemical interactions between cells and the need for neurons. (A) In small multicellular organisms, cells can communicate with other cells via diffusion of signal molecules. If the cells are differentiated, then cell A can only signal to cell B if cell B has a membrane surface receptor for signal A, and vice versa. (B) In organisms with multiple differentiated cell types, cellular interactions can be made highly specific depending on which membrane receptors the specific cells express. Hence, a network of cell-to-cell interactions can be formed. (C) When the multicellular organisms grew larger, relying on diffusion was no longer effective. This necessitated the emergence of specific cell types which were longer than other cells so that they could resolve the problem created by the organism becoming too large for diffusion to be an effective form of communication. Neurons resolved this by developing morphological processes that gradually became axons Jekely2021-moJekely2015-rm. (D) Neurons later developed synapses as an effective means of transmitting signals, and then they could also rapidly communicate between each other, as well as exerting control of actuation in other cells based on the sensor signals received. In this way, the neural network became an entity of its own, while still having the purpose of minimizing the sensor signals arising from the various cells of the body.
• Figure 4: Principle of operation of a PID system in negative feedback control. A set point of the loop is compared to the actual current sensor signal. A deviation constitutes a ‘problem’ that needs to be fixed. The problem signal is fed through a converter that weighs in proportional (P) gain of the signal, an integration (I) of the signal and a derivation (D) of the signal to create an actuation control signal. The resulting actuation effect, i.e. depending on the actuation mechanisms that the cells have affordances for and the resulting impact on the environment, is sensed and fed back to the summator. In biology, the PID effects could be other types of transformation effects, or representations of the problem signal, that the cells and the cellular network is capable of, as described in the first section.