"Feature Detection" vs. "Predictive Coding" Models of Plant Behavior

Considering that repetition suppression is a robust phenomenon that has been observed at different time scales in a number of different sensory modalities, and in response to a variety of stimulus properties (Grill-Spector et al., 2006), it is a possibility that the phenomenon also exists in plants. In support of this scenario, osmo-sensory potentiation has been reported recently. Stephan et al. (2016) have found rapid onset and reversibility of sensory potentiation via plastid-mediated calcium spikes primarily in root apices.

In line with Todorovic et al. (2011), we could test for the possibility that the response in terms of activity and synchrony of transition zone (TZ) cells is attenuated under conditions of expected stimulus repetition (TZ is a root apex region interpolated between the apical meristem and sub-apical cell elongation region that is specialized in processing and integrating sensory information, and instructing the motoric regions driving goal-directed avoidance and exploratory root tropisms—see Baluška et al., 2009, 2010; Baluška and Mancuso, 2013). If repetition is unexpected, by contrast, then we should expect that TZ cell responses would undergo less attenuation. The predictive coding explanation would be that there is less prediction error under expected experimental conditions. Likewise, unexpected omissions may result in a rise in synchronous activity (Baluška and Mancuso, 2013) over TZ cells.

If we were to draw an analogy between transition zone cell response in a study like this and the activity of the FFA in Egner et al. (2010), then we might be construing transition zone cell response as correlated with the levels of prediction error that are generated in response to the presented stimulus—i.e., the mismatch between expected states and sensory states. Alternatively, we might be construing the discrepancy in attenuation as a function both of prediction and prediction error (see Egner et al., 2010). But in either case, we need the mismatch between expectation and sensory states in order to give the explanation.

Behaviorally relevant measurements in bioacoustics may involve the angle of root bending induced by auditory cues that have different probabilistic relationships with the distribution of the sources of stimulation (low frequencies of 200 Hz corresponding to water stream, as opposed to higher frequencies—Gagliano et al., 2012). Or consider water and nutrient availability at the root level. One could use a simple binary maze system, allowing maize roots to take decisions in their growth direction that are stimulated or inhibited by well-defined chemicals, plant growth regulators, and plant nutrients (Yokawa et al., 2014). Growing maize roots make robust decisions on the basis of the available chemicals, and an experimental set-up would allow further analysis of root decisions and learning/memorization phenomena in plants. It is technically possible to increase the number of possible choices by increasing the number of available paths containing stimulants or repellents for root growth. Besides chemical stimuli, one could also load some paths using physical stimuli such as light or sound (mechanical vibrations). In our previous study, we reported that maize roots grow vigorously inside of glass capillaries even when these are inversed and root grow up against the gravity vector (Burbach et al., 2012). However, as soon as they are exposed to light, roots accomplish promptly U-turn movements via complicated thinning of their apices (maize roots are too thick for turning in these thin capilaries), and grow down the gravity vector (Burbach et al., 2012). Similarly, maize roots growing on hard substrate via their characteristic “crawling” movements in darkness stop this and try to grow down the gravity vector if exposed to light. This photophobic root behavior, similar to the root “crawling” movements, requires intact root caps.

Decapped roots grow vigorously but are unable to behave (Burbach et al., 2012). Since light acts as stressor to plant roots, we proposed that this is not just a negative phototropism but rather an active escape tropisms (Yokawa et al., 2011, 2013; Burbach et al., 2012; Xu et al., 2013). This fast root response to sudden illumination implies that maize roots are well prepared to this stress situation and respond very fast to escape back into the darkness. Similarly, placing roots horizontally induces very rapid (within a few seconds) gravistimulated root bending, again implying that roots are well prepared for such a situation (Baluška and Volkmann, 2011).

The next well-studied phenomenon, related to memory-based plant behavior, is stress memory in plants (Bruce et al., 2007). As discussed below, repetitive exposure of plants to particular stress situations is well known to be memorized by plants as these “expect” the next stress challenge and are already prepared to cope better with these stress challenges (Knight et al., 1998; Goh et al., 2003; Bruce et al., 2007; Harb et al., 2010).

Particularly instructive might be measurements of synchronous oscillations at the root apex transition TZ zone (Baluška and Mancuso, 2013). The idea is that synchrony patterns determine impacts of the top–down feedback flows. Such analysis will allow us to test whether plant nutrients and relevant stimuli differ from non-nutrients and non-relevant stimuli in their elicitation of similar TZ cell responses when roots have a high nutrient (or specific stimuli) expectations. The idea is to adjudicate between our working hypotheses by registering data from transition zone action potentials while independently varying both stimulus features (e.g., 200 Hz vs. 1000 Hz) and plant expectations regarding those features (low, medium vs. high water 200 Hz expectation). If the responses elicited are similar, then that would serve as support for the predictive coding hypothesis. In contrast, should different transition zone responses be obtained, this would support the feature detection hypothesis (based on the strict bottom-up responses to nutrient features or non-nutrient features).

Taking into account the distinction between model-based and model-free modes of predictive coding as it was defined at the end of Section Expectation and Surprise in the Ventral Visual Pathway, we should note at this point that the empirical tests that we suggest here would not distinguish sufficiently between these. That is because the “anticipation” that is assayed is of a kind analogous to the Egner et al. (2010) study. The paradigm of Egner and his collaborators is not sensitive to a distinction between model-free and model-based learning, since either of these could produce the relevant sensitivity to context (though given antecedent commitments about human vision, a model is assumed). How, then, would experimenters working with plant models be able to claim support for the weak (entirely model-free) or strong (model-based, or mixed) hypothesis?

Daw et al. (2011) have designed a paradigm for human subjects which can distinguish between model-based and model-free learning processes. The experimental participant performs a sequential two-alternative forced-choice task that has two stages. What is presented in the second stage depends on the participant's choice in the first: each option has a 70% chance of leading to one of two alternative choices in the second stage. The choice in the second task may or may not yield varying levels of reward. Model-based and model-free learning processes will assign value differently in this task, allowing one to distinguish between them. A model-free process will infer a direct linear relation between any reward from the second choice and whatever option in the first choice led to the second, while a model-based learning process will also learn the frequencies with which the first task leads to one or the other second-stage choice, and this will bear on how value is assigned in the case of reward. So if the choice made in the first task leads to that of the two subsequent choices which is less likely given the first choice, and the second choice yields reward, then a model-based learning process will assign value to the first option that was not chosen (on the basis that this is more likely to lead to the choice that has just been rewarded), while that model-free process will assign value to the first option that was actually chosen. This is because the model-based process is able to recognize the statistical relations between the first and second stages.

Although this paradigm is highly human-centric, we see no principled reason why the basic abstract principle at work in it could not be applied in the case of plant-based paradigms—especially seeing as that the Egner paradigm is just as human-centric. The abstract point is that model-based learning processes can learn probabilistic relations between items in a sequence while model-free learning cannot. We believe that an analogous model could be implemented in plant models. This requires (i) two statistically related items; (ii) a reward associated with the second item; and (iii) an appropriate measuring technique for the assignment of value. This would be a fruitful area for future research.

Can plants be astonished by the unexpected? Our working hypothesis, backed by recent studies, is that plants have the capacity to make good predictions regarding forthcoming sources of stimulation, whose perception involves the continuous match of expectations against the impinging environmental signals. We propose here to empirically test plant intelligence by contrasting two alternative hypotheses. According to the feature detection hypothesis, plants are passive, hard-wired, and stimulus-driven organisms. They simply react to environmental stimuli in an online, bottom-up fashion. In that case, there seems to be little need to call such form of reactive behavior “intelligent.” According to the second hypothesis, predictive coding, plants rely upon their own expectations in a top-down fashion, and the system stabilizes only once those expectations face the incoming signals which provide the system with a measure of error.

What are the implications for our understanding of the way plants perceive the world? If the predictive coding hypothesis finds empirical support, plant perception is an active process, where the organism/plant constructs its own plant-specific world-view (courtesy of previous ontogenetic exposures coupled with phylogenetic endowment) via the sensory vectors impinging upon the system. Whenever such predictive sensory vectors (the most likely sources of perturbation in a particular context: nutrients, water, etc.) match the incoming environmental signals (the fidelity of generated world-view is high), we may say that active plant perception takes place.

Plant root navigation is based on sensory signal integration allowing roots to search for water and nutrients, as well as avoiding/escaping dangerous root patches, in complex and heterogeneous soil environment. As we have discussed previously, the root apex TZ is specialized for such sensory signal integration and also allows effective sensorimotor coupling as the basal cells of this zone control root bendings during root tropisms (Baluška et al., 2009; Baluška and Mancuso, 2013). As cells in this special root apex zone are unique with respect of their high demand of oxygen, which is linked to the endocytic vesicle recycling sensitive to brefeldin A (Baluška and Mancuso, 2013), one can expect that this zone represents some sort of command center responsible for sensorimotor circuits akin to the Darwinian brain of lower animals (Baluška et al., 2009).

Now, if we were able to tell apart our two working hypotheses experimentally, then we envisage that a number of outstanding questions would be forthcoming:

First, the experiment from the animal literature reported here suggests that neurons respond to a mixture of expectation and surprise, which is consistent with predictive processing insofar as bidirectionality and functional asymmetry (top-down expectations matching bottom-up signals) are needed (Friston, 2005). If the predictive processing hypothesis is correct, then the plant anticipatory system it should be reflected in some form of functional asymmetry. Such an asymmetric functional composition, however, need not imply any anatomically marked difference in between types of cells (for instance, deeper excitable cells sending out their expectation values, with error signals being forwarded in turn). When we talk about deeper and less deep populations, we have in mind cortical layers. But the integration of top-down and bottom-up informational flows need not reside in anatomically marked distinctions as in the hierarchically layered structure of the mammalian cortex. The hypothesis is that anticipation might be embodied in the temporal structure of the synchronous oscillatory activity of the TZ cells. But first we need to characterize TZ oscillations under control and sensory-loaded situations. Next, one can apply repeated trains of well-defined coupled stimuli, e.g., cold stress following particular light stimulation, or cold stress not preceded with such “light warning.” In animals, Engel et al. (2001) argue that cognition is explanatorily grounded in the dynamic bounding of cell populations from the same and from different neural systems. Sensorimotor coordination requires the soft and functional assembling of many different subsystems, and this is achieved courtesy of the synchronous oscillatory nature of populations of neurons (for another option, see Damasio's “convergence-zone” model; Damasio, 1990).

In plants, top-down influences can be couched in dynamical terms. In the same way that “large-scale dynamics can have a predominant influence on local neuronal behavior by “enslaving” local processing elements” (Engel et al., 2001), we may interpret synced populations of neurons as entraining those out of sync in order to assemble together and enhance their firing saliency. The capacity of plants to anticipate stress situations might be implemented in the temporal structuring of transition zone activity patterns. Sensorimotor integration might therefore be accounted for by transition zone oscillations (Baluška and Mancuso, 2013).

Second, predictive processing in the mammalian brain is thought to take place hierarchically, as top-down and bottom-up sources of information interact at different spatiotemporal scales throughout the network. For the system to successfully anticipate environmental contingencies, an internal model is sculpted via learning; a model that is subsequently exploited for the purpose of generating predictions. What forms of learning might then play a functionally equivalent role in plants? Here, forms of learning such as habituation (Gagliano et al., 2014) look promising. Plants are well known to “learn” from previous stress challenges via modifying their metabolism, cell structures, anatomy and physiology known in plant sciences as hardening or acclimation (Knight et al., 1998; Goh et al., 2003; Bruce et al., 2007; Harb et al., 2010; Dietz, 2014; Ding et al., 2014; Kleinmanns and Schubert, 2014; Shi et al., 2015). Well accepted mechanisms for these plant memory phenomena are linked to epigenetics which is based on heritable chromatin modifications via histone markings (Ha, 2013; Iwasaki and Paszkowski, 2014; Kleinmanns and Schubert, 2014). Similar memory models are well characterized also for animals (Campos et al., 2014). Importantly, epigenetics is relevant not just for cellular levels, but also controls behavior in animals and humans (Impey, 2007; Graff and Mansuy, 2008; Peixoto and Abel, 2013). Besides epigenetic-based memories, however, plants might also use synaptic-like supracellular memories. This is suggested by the fact of immunological memories, which have already been reported to exist in plants since 1986 (Baldwin and Schmelz, 1986; Ruuhola et al., 2007), as well as the existence in plants of plant-specific immunological synapses (Baluška et al., 2005; Kwon et al., 2008; Sup Yun et al., 2008).

Memories of gravity- and photo-stimulations (Nick et al., 1990; Nick and Schafer, 1998) should also be included in future studies.

Plants live in complex environments and their survival is dependent on the reliable sampling of critical biotic and abiotic parameters. To this end, plants and their roots use their sensory systems to sample and integrate the complex sensory information from their environment for the sake of responding adaptively. Similarly, as in higher animals and humans (Laughlin, 2014; Picard and Friston, 2014), faithful representation of their outside world into their metabolic, physiological and behavioral adaptations are essential for their survival. To achieve this goal via optimization of their Darwinian fitness, plants use their own plant-specific intelligence, cognition, and behavior (Trewavas, 2005, 2014; Karban, 2008; Baluška and Mancuso, 2009a,b; Baluška et al., 2009; Karpiński and Szechyńska-Hebda, 2010; Calvo Garzón and Keijzer, 2011; Trewavas and Baluška, 2011; Marder, 2012, 2013; Gagliano et al., 2014; Cvrčková et al., 2016).

Although it might be tempting, on first glance, to relate plant cognition (Trewavas, 2002) to fitness, or as the slogan goes, to the “the survival of the fittest,” we need something other than natural selection per se to grasp plant intelligence. Evolving an adaptation and, say, learning, are different things, but the inheritance-variation-selection view of the evolution of intelligence leaves out of the picture the very learning that takes place in ontogeny. Different individual plants, despite sharing their genotype (those that belong to the same species), can still develop markedly different phenotypes as a function of diverging environmental demands and organismal epigenetic control. Signal integration encompasses phylogenetic endowment coupled with ontogenetic exposures. Individual plants do learn particularities of their local environment (Silvertown and Gordon, 1989; Gagliano et al., 2014; Gagliano, 2015) that could not possibly be predicted on an evolutionary timescale.

The predictive coding model hypothesizes that cascades of sensory input can be surprising to both animals and plants, generating free energy (prediction error). As a result, plants like animals may perform active inference. Recall the example of the fresh-water fish that might minimize free energy by re-sampling the environment in order to bring future input into line with expectations. Our hypothesis is that this may well be the case with salt-avoidance behavior in plants. Integrated environmental assessment (salinity, gravitropic, etc.) appears to be needed at the root level in order to optimize growth under abnormal saline conditions. This means that the plant root, like the fish, would follow the salt-water gradient back into fresh water. If roots enjoy salt-avoidance behavior then this implies an intelligent adaptive strategy that calls for the active, salt-induced, modification of root growth (Li and Zhang, 2008; Sun et al., 2008; Yokawa et al., 2014). Future studies will illuminate how the plant-specific sensory systems feed first into the bioelectrical phenomena at the plasma membrane, and into cellular processes underlying adaptive responses, and behavior at the subcellular, cellular and supracellular organ levels. Plant roots will serve well in these studies as the root apex is an exploratory plant organ, endowed with the sensory root cap, optimized for the search of plant nutrients and the avoidance of toxic and dangerous soil patches.

Once anthropocentric preconceptions are superseded, we may label plant behavior as cognitive insofar as it is flexible and adaptive. With that being said, it goes without saying that predictive coding is not the one and only alternative to feature detection. Unlike predictive coding-based strategies, an ecological, non-model-based approach (Calvo et al., 2014) may exploit information available at the level of the coupling itself between organism and environment, telling against top-down flows of representational predictions that operate in predictive-coding. For the sake of direct comparison, we have only considered “feature detection” vs. “predictive coding” models of plant behavior. Whether a continuous coupling of perception-action that may account for forms of anticipation in non-representational terms is plausible remains an open question for future research6.

All authors listed, have made substantial, direct and intellectual contribution to the work, and approved it for publication.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.