Predicting green: really radical (plant) predictive processing

As in the animal neurobiology literature, plant neurobiology entertains the idea of a hierarchical arrangement; in terms of the distinction between bottom-up, forward and top-down, backward connections. The requirement of functional asymmetry, however, does not entail that the different types of non-neural cells are needed. Vascular cells may ‘simply’ exhibit distinct dynamics depending on whether they mediate predictions or encode error. But the way to propagate top-down and bottom-up signals can vary a lot. Plants may implement PPP principles either by means of other substrates/cell subpopulations for the same prediction and error flows, or even exploiting different types of flows of predictions and errors altogether.

On the other hand, the distinction between forward and backward connections may be based upon the specificity of vascular tissues that connect the overall vascular network of plants, and of intrinsic connections found within each vascular tissue. One possibility worth exploring is provided by phloem vascular and apical cells, respectively, exhibiting different functionalities. Distinct vascular cell populations may encode expected states of the world and prediction error, respectively. It is possible that vascular cells in deep phloem layers serve to encode the former (i.e. mediate predictions), and apical cells encode the latter (i.e. convey prediction errors). We elaborate further in §6.

It should also be pointed out, as noted in [24]: ‘even non-neuronal cells possess many of the same ion channel- and electrical synapse-based mechanisms as do neurons and use them for pattern formation and repair’ [60–62]. As to the way in which cellular dynamics encode particular states, it is worth noting that although the role of acetylcholine, glutamate, dopamine, serotonin, and other neurotransmitters found in plants still needs to be clarified [49,57], the possibility that in analogy to the animal neuron, acetylcholine or dopamine, for example, underlie modulation in neural-like tissue cannot be discarded. The functional analogues of signalling molecules, across the plant and animal kingdoms, have been disclosed recently by the study of neurotransmission in plants. γ-Aminobutyric acid, for instance, has been primarily studied in its metabolic role, although see [63]. From that perspective, cellular signalling looks pretty much the same in animal neural and plant networks. In fact, the way cellular signalling is optimized predates the evolution of central nervous systems in mammals and the vascular system of plants. Early in evolution, cellular signalling in somatic networks found a way to coordinate physiological needs, so that their behaviour could remain adaptive [64].

The simulations reported in [24] are particularly prescient for the current argument. These simulations used an ensemble of free energy minimizing ‘cells’ that employed chemotaxis to self-organize and emulate morphogenesis. Crucially, the chemotactic signals were generated by the cells themselves, providing a form of self-assembly that was mediated by predictive processing; namely, non-neuronal inference about where each cell was located in relation to others. In this example, predictions and prediction errors were encoded by intracellular macromolecules and extracellular electrochemical gradients. One way or another, our working hypothesis is that error correction is implemented by biophysical processes within vascular structures [8] that (spatially) organize an exchange with the world that complies with active inference.

Overall, different organisms may have evolved alternative strategies to appropriately combine top-down, bottom-up, and lateral flows. Non-neural patterns of connectivity may thus allow plants to predict sensory states. The capacity to anticipate may be implemented for instance in temporal patterns of synchronous oscillatory firing of specific populations of plant cells [13]. In short, if animal synaptic activity can bring about active inference, the same in principle applies to the electrosensitive cells of plants.

According to the foregoing, plant cognition can be grounded in the absence of a neural-based generative model. But by the same token, strictly speaking, plant cognition is not grounded either (exclusively) in a non-neural-based functional analogue. Cognition extends throughout the substrate of plants, their body and the environment. In this way—focusing upon the plant body—facts about morphology and the material composition of plants are crucial. Consider the ‘Yokoi hand’ [65]. This prosthetic hand was designed with flexible and soft gripping materials that, courtesy of a particular morphology, could grip a variety of objects. Importantly, both morphology and materials deliver the goods—reducing the amount of predictive control that would be otherwise required. A morphological design embodies the coming together of the finger tips as the hand itself closes. On the other hand, the type of material—that the finger tips are made of—fits hand in glove with the type of objects to be gripped.

If a robotic hand can exhibit a degree of adaptation to different shapes and contingencies through its material constituency and morphology, the same can be said of plants. Plants, we may say, compute with their bodies in the service of adaptive flexible behaviour. Consider a vine climbing up a host tree for photosynthetic purposes. Vines can attach to supports in many different ways; insofar as attachment mechanisms and stem structure and function are concerned. Their capacity to cling to host trees resembles pretty much the type of extended solutions found by morphological roboticists. In fact, depending on the type of attachment, vines are able to climb some supports and not others. Thus, tendril and twining climbers have more efficient climbing mechanisms than hook-climbers [66]. The stems of the former are more flexible; those of hook-climbers, more rigid. As a result of their very phenotypic composition, some climbers can cling to supports of different diameters. Some redvines make use of jelly-like fibres allowing the tendrils to squeeze supports. Importantly, depending on the direction of coiling, the tendrils have their gelatinous fibres distributed in different ways (e.g. those that coil always in the same direction have all the fibres on the concave side [66]). Other anchoring strategies rely on glandular secretions, widening the range of supports potentially climbable [52].

The mechanoreceptor tactile bleps of tendrils of Bryonia dioica Jacq [67] illustrate nicely one way in which sensory and effector organs act synergistically in this sort of embodied (active) inference. The high density of plasmodesmatal connections in tendril epidermal cells underwrite the electrochemical coordination required for appropriate coiling. But tendrils do not just respond to pressure. As originally observed by Pfeffer and Haberlandt [67], sliding movements on a rough surface are needed. In short, materials and their properties are put to service in the perception of gradients.

The Yokoi hand–vine plant metaphor takes us to the idea of frugality. Plants are good at deploying the least complex model, form or generative model that will serve their needs. Recall from above that the overarching imperative prescribed by the free-energy principle is a resolution of uncertainty. If we unpack this in anthropomorphic terms, it simply means that actions (at least good actions) will resolve uncertainty. This is synonymous with selectively gathering evidence for one's own existence and choosing behaviours that have the greatest epistemic value [28,46]. This brings with it the important notion of epistemic affordance. In other words, sentient plants and creatures behave in a way that resolves uncertainty about the environment through some form of epistemic foraging. This sort of affordance goes beyond minimizing surprise (e.g. by avoiding high salt concentrations); in the sense that the opportunity to resolve uncertainty now becomes attractive (e.g. deliberately sampling some unexplored part of the environment to see whether it has a high or low salt concentration). One can see immediately that this is just a way of describing exploration and the intrinsic (epistemic) value of certain behaviours that affords the opportunity to reduce uncertainty (i.e. exploit novelty).

From the free-energy principle, it follows that plants represent the causes of sensory stimulation in terms of their internal states and morphology. Anatomical and physiological constraints underlie the capacity of plants to do so. By analogy with Friston [68], an optimization scheme under a free-energy formulation allows us to account for a number of empirical aspects at the level of both plant anatomy and physiology, as summarized in table 1. These aspects range from the hierarchical arrangement of vascular tissues and the functional asymmetries between forward and backward connections to minimal cognition as discussed in plant neurobiology [69]. In table 1 we consider three levels of explanation and predictions: (i) at the level of plant anatomy and connectivity, the hierarchical deployment of plant vascular tissues and connections; (ii) at the level of plant electrophysiology, phenomena like repetition suppression; and (iii) at the psychophysiological level, the working hypothesis of plant neurobiology is that it can explain cognitive level phenomena, such as plant priming [7] and learning [70] (table 1 for details on domains and predictions).

Consider first the hierarchical vascular organization of plants. Plants are characteristically modular, with a highly decentralized architecture. This reflects the evolutionary needs of sessile organisms, for which a highly centralized system would not be adaptive. Yet, a hierarchical organization is needed if predictions and prediction errors are to flow up and down the plant body. In fact, the modular decentralized architecture of plants is compatible with a considerable degree of hierarchical organization. For example, vascular cells are arranged in layers that converge at both ends, at the root and shoot apices [71]. At a different level of description, individual modules in shoots are connected by nodes, and the same goes for every anatomical subunit of the entire plant body. This enables modules to connect to each other for the purpose of signalling [72]. Overall, vertical xylem and phloem strands, which are highly cross-linked by horizontal and tangential anastomosis [59], form a ‘channel and net’ or reticulated structure that runs throughout the plant [1,55,72,73].

With regard to the electrophysiological and psychophysiological domains (table 1), we may consider repetition suppression and learning [13], respectively. Repetition suppression, a phenomenon in which responses are attenuated as a result of the repeated presentation of a particular stimulus, may well take place in plants, with top-down expectations underlying suppression. With respect to learning, habituation [44] and associative learning [45], these may well play a functionally equivalent role in plants, in the service of generating predictions that are informed by a plant's experience.

The two principal themes of functional asymmetry (between descending predictive and ascending prediction error signals) and the frugal nature of a plant's functional architecture speak to the two fundaments of free-energy minimization outlined in this article. In particular, the frugal, complexity minimizing structure of plants—well suited to their world—follows naturally from the minimization of free energy. This follows from the mathematical equivalence between free energy and (Bayesian) evidence for a generative model [16]. In other words, if we associate the structural and dynamical form of a plant with a model of its eco-niche, then any process that minimizes free energy will necessarily maximize model evidence. Because the plant is the model—it is basically self-evidencing [46].

This can be seen at a practical level in everyday statistical analysis where it is known as Bayesian model selection. The same arguments have been applied in the context of evolution—such that selective pressure is simply the process of free-energy minimization playing out at an evolutionary time scale [74–76]. In short, natural selection is nature's way of performing Bayesian model selection. A key insight here is that evidence is the difference between accuracy and model complexity. In other words, a good model with high evidence will provide an accurate explanation for sensory exchange with the environment, while minimizing its complexity.

Complexity can be thought of as the degrees of freedom used by the plant to anticipate and predict its sensory exchange. This leads naturally to a principle of minimum redundancy (well established in the neurosciences [77]), whereby a good plant will retain just those sparse, frugal structures that are necessary to anticipate the world. This can be evident in the phenotypic form (as unpacked earlier by analogy with robotic gloves) or in terms of conditional dependencies and ‘action at a distance’ mediated in plants by channels and electrochemical waves (very much like axonal connections and electrochemical synaptic transmission in the brain). In short, under the free-energy principle—and the active inference that this entails—one would anticipate that plants would come to distil the essential causal structure in their environment in terms of their physical form and biophysical function. It is this form and functional architecture that constitutes the generative model and underwrites their existence.

The dynamics within this generative model (e.g. plant) correspond to signalling of the sort associated with belief propagation in the brain. This implicit (Bayesian) message passing is simply a description of dynamics that maximize Bayesian model evidence (or minimize free energy). This can manifest in many forms. A popular example is predictive coding in the brain [17,18]. The crucial aspect of these dynamics (variously known as variational message passing, belief propagation, predictive coding etc.) is an asymmetry in the messages that are passed on the internal states generating predictions. Interestingly, one can always formulate a free energy minimizing scheme in terms of prediction errors, if the generative model calls upon probability distributions within the exponential family [78]. This is just a technical way of saying that it is more than likely that any plant or creature can be described (at least mathematically) as passing predictions and prediction errors around its body in the service of minimizing prediction error.

Finally, it is worth noting that the very idea of plant perception seems to turn upside down our very intuitions as to what perception and action entail. In this context, one can consider directions for future research in the context of plant action and the perception of affordances. If plants are sentient and behaving organisms, then plant perception and action are driven primarily by anticipatory routines triggered in the plant's body. At this point we can see how radical PPP can actually be. Plants are surprised by the unexpected, and their capacity to navigate uncertain terrains may therefore depend on a cascade of top-down predictions, constantly predicting incoming signals. As in the case of pea roots that can predict the future [7], anticipatory behaviour is crucial for adaptive, surprise eluding, success.

Vines provide the sharpest contrast with conservative PP scenarios. Grasping, say, a phone, according to Hohwy [46] can only be explained via the exploitation of an internal representation of the phone itself. But once PP is fully developed with fast and frugal, morphological, and non-neurally decentralized resources, we can see that detailed internal representations may not be necessary: a climbing plant does not grasp a support because it has an internal representation of it, but rather because of the soft-assembling of the plant body–environment in the context of the approaching manoeuvres. In other words, the tendrils of the plant do not call upon representations of the affordance to climb: they are, in and of themselves, the generative model which realizes that affordance.

The very exploitation of ecological information within a PP framework permits us to envisage really radical predictive processing solutions like those be leveraged by plants. That plants can perceive ecological information is not the big news. Turvey et al. [79] considered how Monstera gigantea, a climbing vine, could perceive a form of climb-ability, as this plant would grow skototropically towards darkness [5]. Thus, vines can perceive possibilities for action, but to fully understand this ‘perception’, we need to bear in mind that the vine and its support are functionally coupled by the predictions of ‘experienced’ plants. Vines control their approaching manoeuvre to the support by anticipating sensory states, and realizing the states through embodied action.

Plants are a good example of radical predictive processing because they are not in the business of enriching a neurocentric model of their surroundings (this is effectively outsourced to natural selection); rather, they are in the business of appropriately coupling with it. A movement of circumnutation, for instance, is triggered, maintained and modified endogenously, or at least that is our working hypothesis. Taking into account that many growth-related movements are irreversible, exploration must be accomplished efficiently. Control is thus needed, if the metabolic and computational cost of irreversible but idle movements is to be minimized.

These considerations suggest an understanding of plant cognition radically different—not only with respect to mainstream cognitive science, but also to conservative approaches to PP. When we say that plants are proactively engaged with their surroundings we mean literally that they foresee possibilities of interaction with their local environment ahead of time. But it is important to emphasize that this needs to be so. A plant that only represents sensory inputs as they flow past would be dead meat. This is so precisely because they move slowly and cannot afford to react reflexively in response to the present. Plants, despite the slowness of their responses, can exploit distal nutrient-related or reproduction-related causes in their eco-niche. In short, they need to deal with contingencies ahead of time—because they cannot afford to behave otherwise.

We would like to thank Andy Clark and three anonymous reviewers for helpful guidance in the presentation of these ideas.

P.C. conceived of the original idea for this work. K.F. and P.C. wrote and revised the manuscript. All authors gave final approval for publication.

This work was funded by the Wellcome Trust (K.F. is funded by a Wellcome Trust Principal Research Fellowship: 088130/Z/09/Z). P.C. is supported by Spanish Ministry of Education, Culture and Sport through a ‘Stays of professors and senior researchers in foreign centres' fellowship.