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

In this article we consider the possibility that plants exhibit anticipatory behavior, a mark of intelligence. If plants are able to anticipate and respond accordingly to varying states of their surroundings, as opposed to merely responding online to environmental contingencies, then such capacity may be in principle testable, and subject to empirical scrutiny. Our main thesis is that adaptive behavior can only take place by way of a mechanism that predicts the environmental sources of sensory stimulation. We propose to test for anticipation in plants experimentally by contrasting two empirical hypotheses: “feature detection” and “predictive coding.” We spell out what these contrasting hypotheses consist of by way of illustration from the animal literature, and consider how to transfer the rationale involved to the plant literature.

Sculpted by evolution, plants allegedly react to environmental inputs only in an instinctual manner, with their behavioral repertoire reducing to invariant tropistic (directional) or nastic (non-directional) responses implemented in the form of sets of fixed reflexes (Silvertown and Gordon, 1989; Trewavas, 2009, 2014). This behavior is usually accounted for in hard-wired terms, and being hard-wired undermines the ascription of intelligence to plants. That is because behavioral flexibility is one marker of cognitive sophistication, and hard-wired behavior does not admit of behavioral flexibility; it is “hard-wired” to particular stimuli or cues. This mechanistic and non-cognitive view of plant behavior goes back to Julius von Sachs and Jacques Loeb (Greenspan and Baars, 2005)—incidentally, Loeb introduced this concept to animal behavior from his earlier studies on plants, proposing similar hard-wired explanations of both animal and plant behavior (Loeb, 1918).

This reductionist approach is likely to fail in animal biology (Greenspan and Baars, 2005; Greenspan, 2012) and, we contend, is equally likely to fail in plant biology (Trewavas, 2005, 2009, 2014; Karban, 2008; Baluška and Mancuso, 2009a,b, 2013; Calvo Garzón and Keijzer, 2011; Trewavas and Baluška, 2011; Marder, 2012, 2013; Gagliano et al., 2014; Cvrčková et al., 2016). For one thing, plants are motile, and their behavioral repertoire is richer than commonly acknowledged. Virtually no growing part of any single plant fails to exhibit a movement of nutation (Mugnai et al., 2007). Shoots of climbing plants guide their movements to reach a support; roots navigate belowground, guiding their movements to secure nutrients intake; young and terminal leaves display helical and rotational oscillatory movements, etc. (Darwin, 1875; Darwin and Darwin, 1880). In fact, sophisticated forms of plant behavior abound. Plants and their roots are sensitive to a variety of signals other than water, light, minerals, or gravity. Plants can sample more than 20 different biotic and abiotic parameters from their environment and integrate this complex sensory information to mount appropriate behavioral responses (Knight et al., 1998; Karban, 2008; Hodge, 2009; Trewavas, 2009; Baluška, 2012; Baluška and Mancuso, 2013). Roots grow by assessing the future acquisition of minerals and water, a process that requires the integration of gravity, moisture and mechanical perturbations, among other vectors (Takahashi et al., 2002). Likewise, roots can, for instance, sense available space and discriminate self-roots from alien roots. Less familiar examples include salt-avoidance behavior (Li and Zhang, 2008; Sun et al., 2008; Yokawa et al., 2014).

Here salinity interacts with the gravitropic response and an overall integrated signal assessment appears to be needed in order to optimize growth under abnormal saline conditions.

Unveiling why plant behavior is so flexible (Trewavas, 2009, 2014) may cast a new light on intelligence without recourse to anthropo- or zoomorphisms. In this context, our theoretical motivation in this article is to consider the possibility that plants exhibit anticipatory behavior, a mark of intelligence. If plants are able to anticipate and respond accordingly to varying states of their surroundings in a context-sensitive way, as opposed to merely responding online to environmental contingencies, then that capacity may be in principle testable, and subject to empirical scrutiny. Our main thesis is that adaptive behavior can only take place by way of a mechanism that predicts the organism's own states of sensory stimulation (Egner et al., 2010; Chennu et al., 2013).

More broadly, we regard (minimal) cognition to be a biological phenomenon. What seems to be involved is having a sensorimotor organization, and free-moving with the purpose of manipulating the environment (allowing for metabolic forms of adaptation and anticipatory functioning) (Calvo Garzón and Keijzer, 2011). This holds for all forms of life. It is the ability to act upon environmental contingencies that defines biological systems. Only those biological systems can survive which perceive the world veridically via successfully predicting it, and not merely reacting to it (Clark, 2016). Of course, the falsification of a reactive model would not imply that a particular anticipatory countermodel is correct. The notion of anticipation may come in a variety of forms, with weaker and stronger readings being possible. Anticipatory behavior may rely upon the capacity of the system to model internally the environmental sources themselves. But forms of anticipation according to which predictive success is a function of actual past behavior—stronger forms of anticipation that do not depend on modeling the future internally (Stepp and Turvey, 2010; Stepp et al., 2011)—cannot be discarded beforehand1. However, for present purposes we are proposing a test for anticipation in plants experimentally by contrasting two empirical hypotheses exclusively: “feature detection” and “predictive coding (processing).”

According to the feature detection hypothesis, plants perceive their surroundings, but the detection of and adaptive responses to environmental features are consistent with a reactive interpretation of plant behavior. By contrast, the capacity to perform predictive processing would point toward a basic form of agency: plant perception may be seen as an active process of probabilistic inference akin to that found in animals (Kok et al., 2013). Plants under this interpretation are pro-active; they actively sample their environment to generate information, estimating the likelihood that one external state of affairs, and not another, is the source of energy impinging upon its sensory periphery.

Now, there is a preliminary question that might be raised as to whether anticipatory behavior entails intelligence. Our position is not that such behavior is sufficient for intelligence; the existence of basic predictive abilities in artificial neural networks would seem to rule that out. But we do contend that such behavior is necessary for the existence of minimal intelligence, which is to say that it is a general feature of it. Therefore, our proposed study contributes to the question of minimal intelligence in eukaryotes in the following way: if the feature detection model of plant behavior is confirmed by the study, then we can claim to have falsified the minimal intelligence hypothesis. But if it is not then we can provisionally retain the conjecture.

In the following sections we spell out what these contrasting hypotheses consist of by way of illustration from the animal literature, and consider how to transfer the rationale involved to the plant literature. Discussion and directions for future research will follow.

The canonical domain for feature detection is visual cognition. The origin of the feature detection hypothesis dates back to Barlow's (1953) work on the frog's retina, according to which single ganglion cells could release particular motor acts triggering specific behaviors. Barlow's insight was elaborated further by Lettvin et al. (1959), who showed how different ganglion cells responded to different patterns of excitation, and subsequently by Hubel and Wiesel (1965) who extended the paradigm to mammalian research2. According to the feature detection model, neurons are seen as specialized bottom-up feature detectors that respond selectively to, for example, angles, lines, movement, or edges. Information flows upwards into deeper visual cortical layers (V2, V4, IT), allowing for the receptive field of neurons within those layers to respond to stimuli of increasing complexity.

By contrast, information in the “predictive coding” model (Rao and Ballard, 1999; Friston, 2005; Egner et al., 2010; Chennu et al., 2013) flows simultaneously in both directions as well as laterally within individual levels of processing. The nature of these two “flows” is as follows. Firstly, predictions (conditional probabilities of particular features being the cause of stimulation) are propagated top-down. Secondly, mismatches between those predictions and the incoming input signals (prediction errors) are propagated bottom-up. At each level in the hierarchy, predicted inputs are compared with actual inputs and the latter are only propagated upwards if there is a mismatch, i.e., a prediction error. The aim is to minimize prediction error (the difference between the top-down prediction of what the sensory signal is and the actual signal traveling bottom-up). In this way, perception is interpreted as the end result of a process whereby top-down predictions match the environmental input, based on a hierarchical generative model of the causes of that input.

It is possible to test between “feature detection” and “predictive coding” in animal visual cognition. The two models generate different predictions and thus are subject to empirical contrast. In particular, Egner et al. (2010) designed an experiment to check for interaction between expectations and error (degree of surprise), as opposed to simple feature detection. In their study, they considered the fusiform face area (FFA), a visual area located in the ventral system known to be involved in facial recognition. Were “feature detection” to be correct, the FFA area would respond to stimulus facial features per se. On the other hand, if “predictive coding” is correct, the FFA area should respond to the weighted sum of expectation and surprise: the sum of top-down predictions (say, the expectation to see a face) and bottom-up surprise (the degree of expectation violation as a face is fed to the system, or not).

Egner et al. (2010) tested these hypotheses with fMRI measurements by varying both the stimulus features (pictures of faces vs. pictures of houses) and the expectations generated by experimental subjects as the material is being presented to them. They did so by pairing the stimulus features (faces and houses) with colored frames (green, yellow, blue), in such a way that the frames provided the low, medium or high cue (25, 50, 75%) that the incoming stimulus would be a face). The prediction entailed by the predictive processing model is that whereas faces and non-faces will elicit equivalent FFA responses when subjects have a high face expectation (since in both cases there is no association with activity related to face surprise), faces and non-faces will elicit maximally differing FFA responses when subjects have a low face expectation (only one condition, faces, is associated with face surprise)3. By contrast, the prediction entailed by the feature detection model would be different FFA responses under both conditions since subjects are exposed in a strict bottom-up fashion to face features or non-face features, regardless of the expectations involved. These results are plotted, respectively, in Figures 1A,B.

Would plants also be able to generate expectancies in line with the interpretation that obtains in Figure 1A? In the following sections we argue that this is an open empirical possibility.

There is an additional motivation for applying a seemingly zoo-centric theory to the understanding of plant behavior. That is the nature of predictive processing in cognitive neuroscience as a specific application of a more general theory of biological functioning. This more general theory is called the free energy principle (Friston, 2010, 2013; Hohwy, 2015; Seth, 2015). The free energy principle applies to all biological self-organizing systems, and it is a hypothesis about how it is that such systems manage to keep within the state-space that constitutes homeostasis for that system. On the free energy principle, it is the ability to act upon its environment to avoid damaging phase changes—to keep within this state-space—which defines specifically biological systems.

An example used to illustrate this idea is given by Friston and Stephan (2007, p. 423). Their example of a living/non-living pair is that of a bird and a snowflake. Each of these systems is comprised by a physical structure which entails a set of states within which that system must remain in order to maintain homeostasis. The snowflake, for example, will melt if it encounters temperatures above a certain point. The same applies to the bird phenotype, though the complexity and resilience of that phenotype implies a much wider range of possible states which it can occupy. The difference between the bird and the snowflake is that the bird interacts with its local environment in order to stay within its optimal state-space: it seeks nutrition, avoids predators, and moves around within its ecological niche.

Another way of putting this difference is that the biological agent must avoid states that are surprising. These states are not surprising in the phenomenological sense of the word, but are rather improbable given some implicit set of expectations. That is to say that they are surprising in an information-theoretic sense. In our example above, the implicit expectations in question are constituted by the phenotype of the organism. Surprising states for a snowflake are those in which it melts; surprising states for a fish include those in which it is too dry to survive (cf., Hohwy, 2015, pp. 3–4).

The free energy principle states that biological agents can perform this task by way of free energy minimization, which is an upper bound on surprise; more specifically, that “biological systems on average and over time act to minimize free energy” (Hohwy, 2015, p. 2). That is a process that is mathematically analogous to prediction error minimization as it is construed in predictive processing (free energy is equal to the long-term sum of prediction error). That means that the free energy principle implies that the organism maintains a generative model (it generates predictions of future states) of the causes of its sensory states, and works at minimizing prediction error with respect to this model.

There are two ways in which prediction error minimization can proceed. The first of these could be called perceptual inference. In perceptual inference, prediction error is minimized by updating expectations in order to bring them into line with actual states, rather than those that were falsely predicted by the expectations before updating. This is the process that is thought to explain performance in the visual perception tasks discussed above (Egner et al., 2010). The other possible way of minimizing prediction errors is known as active inference. In active inference, prediction error is minimized by the selective sampling of sensory states such that those states are brought into line with what is predicted by prior expectations. It is this process that is thought to explain the function of visual saccading, in which the eye moves rapidly and to take in different parts of the visual field. The active-inferential theory of visual saccading states that these saccades constitute the sampling of sensory states in such a way that evidence is gathered for predictions about visual input (Friston et al., 2012a)4.

We can cash out this quasi-cognitive distinction between perceptual and active inference in a biological context by referring to the example of a fresh-water fish. These fish cannot survive in salt water. That means that the cascade of sensory input that immersion in salt water elicits is highly surprising to that biological agent. Such immersion will generate a great deal of prediction error, or free energy. In order to minimize that free energy the fish can do one of two things. The first is to re-sample the environment such that future input is brought back into line with expectations. That means active inference; it means that the fish follows the salt-water gradient back into fresh water. The second is to update its expectations such that it expects to be in salt-water. This is perceptual inference; that likely happens initially and at some level in the predictive hierarchy; though given the relatively unlikely state that immersion in salt water represents, it is not to be ultimately chosen as an adaptive strategy.

This entails a link between homeostasis and anticipative behavior. That is to say, the free energy principle entails the hypothesis that the mechanism by which biological systems maintain homeostasis is a kind of anticipation; namely, by minimizing the difference over time between the implicit model of the world that its morphology constitutes and the sensory states that it finds itself in. That means the minimization of surprise, and this is done through perceptual inference (morphological change in response to sensory states, ranging from different patterns of activation across the visual cortices in mammals all the way down to hardening responses to abiotic stresses in plants) as well as active inference (selective sampling of sensory states through movement in the environment, such as visual saccading in animals and nutations in plants). For this reason the predictive processing theory is a prime candidate for the explanation of plant behaviors of the kind we label “minimally intelligent” in this article. The analogous and wider-scoped free energy principle provides one rationale for spreading wider the scope of explanatory concepts that have heretofore been confined to application in animal biology.

Lastly, there is an important distinction in the predictive coding literature between two kinds of learning architecture: model-based and model-free (Daw et al., 2005). Model-based architecture follows predictive coding as it has already been described; prediction error is minimized relative to a hierarchical generative model of the causes of sensory states. So it is representation-heavy mode of predictive coding that implies a relatively sophisticated model of the environment; the responses of the organism to the environment are mediated through the model. Model-free architecture lacks any such model; the responses of the organism to the environment are direct and unmediated. A model-free architecture will produce learning responses that proceed through trial and error, like classical reinforcement learning. Whether plant learning is model-based or model-free (or both) would have interesting consequences for our hypothesis. If plant learning includes model-based strategies, then it is much more sophisticated than formerly assumed, and capable of making deeper and more discriminative distinctions between different kinds of stimuli. We may therefore further massage our hypothesis into two distinct commitments: a weaker, on which plant cognition is entirely model-free; and a stronger, on which it also includes model-based learning. We consider how we might empirically distinguish between these later on in the paper.

By way of a caveat, it must nonetheless be noted that not all the details of the free energy principle are worked out here, and it is indeed not uncontroversial as a whole (e.g., Gershmann and Daw, 2012). On the other hand, and before we consider further the predictive processing/free-energetic vs. the feature-detection interpretation of plant behavior, it must be borne in mind that despite features themselves being hard-wired, there is room for learning under the feature detection paradigm. In effect, the release of specific motor acts/behavioral patterns via, for example, single retinal ganglion cells in Barlow's (1953) original model is consistent with the possibility that learning takes place: the events being coded for by detectors could well be hard-wired, and yet the mapping between the features that are being detected and subsequent actions be learned5. As a result, we would be confronted with two different models of how a system successfully anticipates environmental contingencies, both sculpted via learning. For present purposes, we shall thus restrict our attention to the Barlow (1953), Lettvin et al. (1959) and Hubel and Wiesel (1965) original learning-free formulation, for the sake of contrasting it with predictive processing; a model that connects anticipation with learning courtesy of an internal model subsequently exploited for the purpose of generating predictions. As to the free energy principle, it is sufficiently plausible, and its closeness to predictive processing gives us cause to think that the empirical tests that we describe are worth carrying out in the context of plant movements. Indeed, those prospective studies could be construed as a contribution to the wider project of putting the free energy principle to the test in an experimental setting.

If the predictive processing/free-energetic interpretation of plant behavior is on the right track (Figure 1A), then we should be able to observe plant excitable cells responding to the summation of activity related to expectation. Or, at the plant organ/root level, the actual context of the environmental situation should determine the behavioral outcome as roots explore their surroundings. By contrast, should feature detection be correct, then we should expect no interaction with expectations or context. In principle, we should observe plant cells and organs responding exclusively to the physical vector of stimulation.

In order to make these two hypotheses testable, a number of options are available with regard to (i) the type of gradient potentially being sensed (gravity, light, moisture, oxygen, touch, etc.), and (ii) the experimental method by which to generate the required expectations. Finally, (iii) different measurement techniques may be considered when testing our working hypotheses.

(i) As to the type of stimuli, we may consider either the perception of a single modality (for example, gravisensing) or the integrated response to several vectors, such as the weakening or inhibition of the response to gravity when stimulated by touch (Fasano et al., 2002), light (Liscum, 2002; Wan et al., 2012), water (Takahashi et al., 2002), or sound waves (Gagliano et al., 2012). As well, such stimulations may be continuous or transient. Continuous touch stimulation may obtain by placing a barrier that prevents root gravitropism. Or we may consider continuous bioacoustic stimulation (Gagliano et al., 2012). By contrast, in the case of transient touch stimulation, we may consider, for instance, the stimulation of root cap peripheral cells (Massa and Gilroy, 2003), or transient bioacoustic stimulation.

(ii) With respect to the manipulation of the degree of expectation, there are different options. In the aforementioned case of visual cognition and FFA, they achieved it by pairing the stimulus features (faces and houses) with colored frames (green, yellow, blue) in such a way that the frames provided the cue (low, medium or high) that the incoming stimulus would be a face (Egner et al., 2010). Likewise, both stimulus features for the experimental setting in plants, and the expectation of encountering those stimulus features, would need to be varied independently.

In order to generate expectations we can either use the same or a different network. That is, the cues that precede the presentation of the stimuli can belong to the same type of system or to a different one, such as happens in the integration of electrical network with chemical network systems when electrical activity is elicited by chemical means. Expectations may also be generated by mechanical stimuli, elicitation of electric spikes, firing with transient stimulation, and repetition (e.g., transient acoustic stimulation).

(iii) Finally, with regard to the available measurement techniques, we may use the measurement of behavioral events (for instance, root bending, and tropism), single-cell recordings, or non-invasive neuroimaging techniques.

To take a particular example for the sake of illustration, consider the phenomenon of “repetition suppression,” where it is found that the repeated presentation of a stimulus results in the attenuation of neural activity (Todorovic et al., 2011). In an MEG study, Todorovic et al. (2011) presented human subjects with auditory stimuli whose repetition were manipulated so that they were subject to different levels of expectation. They reported that a larger suppression is registered when subjects expect the repetition to take place. This is observed in auditory cortex activity and synchrony. In their view, repetition suppression is to be interpreted in predictive coding terms: top-down expectations underlie suppression. What that means is that as the stimulus in question comes to be expected, it becomes part of the model that is being used to predict the flow of sensory information. This means that presentation of that stimulus generates less prediction error, less stress penalty, and therefore less activity, since it is already integrated into the top-down signal which will not be contradicted by mismatch.