Reservoir Computing with Plants

Reservoir Computing with Plants #

Olivier Pieters

Supervisors Prof. Francis wyffels, dr. Tom De Swaef and dr. Michiel Stock

Doctoral dissertation submitted to obtain the academic degree of Doctor of Computer Science Engineering

Summary #

Plants are ubiquitous on Earth. They are often regarded as organisms that undergo the environmental changes they experience. Instead, we advocate for a more integrated view: a plant as a computing entity. Plants are complex organisms composed of many interconnected nodes and modules. These enable a plant to deal with highly variable environmental conditions due to weather fluctuations, predation and diseases. Despite the absence of a brain-like organ and their inability to move, plants can react effectively to cues from their environment. A plant continually gathers and updates diverse information about its environment and integrates this with its present internal state. From this integrated information, it makes decisions that reconcile its well-being with its environment. We propose to consider the plants as a computing unit in the context of physical reservoir computing (PRC).

PRC is an unconventional computing paradigm that utilises physical substrates for computation. This paradigm entails using a high-dimensional, nonlinear dynamical (physical) system as a computational resource to solve a task. Examples encompass the control of mechanical systems by using a compliant robot body or the processing of optical and electrical signals. From the biological realm, a cat’s primary visual cortex and bacterial cultures have also been demonstrated as a reservoir for classification tasks.

Initially, it might appear odd that (physical) reservoir computing can work. However, it shares a lot of similarities with conventional computing. Computation is the process of transforming information to achieve a specific goal. Conventional systems perform this goal by using algorithms. The human-designed algorithm processes inputs to accomplish a goal or obtain an output. In PRC, this algorithm is replaced by a physical substrate that performs the computation. This substrate or reservoir is observed using sensors whose readout values are combined to obtain the output. The general idea is that due to the high dimensionality and memory of the reservoir, the output can be observed using a simple linear combination of the state observations. As such, reservoir computing systems can be trained using well-understood linear regression and are fast to train.

Substrates generally have to fulfil two main requirements for PRC: nonlinear characteristics and fading memory. Plants are indeed nonlinear organisms. Increasing the light intensity does not result in the same increase in the photosynthesis rate (nonlinear behaviour). There is also evidence that plants contain memory because they have the ability to learn from experience, which is used to optimise future light acclimation. However, this is not the same as fading memory. Fading memory implies that past events have decreased importance as time advances. There is not yet formal proof of this ability in plants.

Classic PRC relies on a stationary reservoir; its dynamics, nonlinearity and memory are fixed in time. In general, this is not the case for plants. Plants have continued development, even in adulthood. To alleviate this problem, we study plants over a relatively short period: eight days. During this timeframe, we consider the plant stationary. In this work, we hope to bridge reservoir computing to an – as far as we know – unexplored field: plant ecophysiology. Despite the aforementioned limitation, it can form the basis of a paradigm shift in phenotyping. Instead of focussing on specific traits and their effect on physiology, a more holistic approach can yield interesting new insights into plant behaviour.

In a first study, we investigate the applicability of the PRC framework to plants using a hyperspectral camera. A hyperspectral camera is a generalised version of a conventional camera. Instead of capturing only three spectral bands (red, green and blue), it captures many more bands of light with higher accuracy (narrower bands) and a broader range (wavelengths outside the visible spectrum). Despite extensive analysis, this study was unsuccessful. Background materials were equally good at predicting the considered regression tasks as the plant data. We suspect that the root causes are insufficient accuracy of the camera and small spectral changes. Improved sensor technology might solve some of the issues, yet it remains unsure what the required accuracy is. By design, plants were not severely stressed in this study. As a result, the spectral changes were limited. Alternative sensory equipment thus appears more promising because of several issues with hyperspectral cameras. Firstly, these cameras are expensive sensors compared to conventional ones. Secondly, they produce vast amounts of data. Storage and processing of this data can be a challenge. Thirdly, hyperspectral data is complex. There are many bands available, and the spectral resolution might not be uniform. Moreover, extracting the relevant features from the data is an ongoing research topic.

Although the results from the hyperspectral experiments were inconclusive, they are highly relevant for the phenotyping community. Using our findings, we illustrated some of the limitations of current hyperspectral technologies. Moreover, we also elucidated that the plant was not the root cause of the system’s performance despite some tasks being performant. This observation highlights that it is always essential to study the environment and sensory system’s effect on the task at hand.

So, for a second experimental study, we shifted the focus towards contact sensors. More specifically, we employed leaf thickness sensors. However, to characterise the state of the plant for PRC, we needed to measure this state sufficiently fast. Established sensory systems did not meet the required specifications in terms of accuracy, flexibility and cost. Therefore, we designed a custom system: Gloxinia.

The Gloxinia sensor platform aims to advance monitoring in fundamental and applied plant research. Four key needs were addressed: sensor scalability, accuracy, cost and versatility using an open hard- and software design. The platform is comprised of individual sensor nodes that communicate with each other. Each node has a control board to which sensor nodes are connected. These sensor boards are equipped with the necessary electronics for interfacing with most analogue sensors used for contact measurements. Digital sensors can also be connected to the control boards. To validate the accuracy of the system, we set up an experimental trial in a growth chamber. Environmental conditions, leaf thickness, and leaf elongation were successfully measured on one tomato and two strawberry plants at high resolution.

Using the Gloxinia platform, we demonstrated PRC with plants. While we did not quantify the different memory aspects and nonlinear properties of the plant separately and the processes where these originate, it is a pivotal step towards PRC-inspired computing with plants. We showed how leaf thickness measurements of strawberry plants were used to assess ecophysiological, environmental and benchmark regression targets. Our results indicate that plants are unsuited for general-purpose computation yet are highly relevant for plant-related tasks. Photosynthetic rate and transpiration rate are the two main ecophysiological tasks investigated.

Currently, PRC with plants is in the exploratory phase. We demonstrated the potential of PRC with plants for ecophysiological tasks using leaf thickness sensors. Advancements in sensor technology such as more accurate sensors and alternative sensing technologies can further improve the results. Moreover, plants are non-stationary. The PRC framework should thus be extended to deal with this behaviour. However, the most drastic implication of PRC might be a new perspective on plants and their behaviour. Treating a plant as a computing entity can help generalise plant behaviour and provide essential context to physiological studies. Each trait exhibited by a plant can be viewed as the result of the complex interaction between environmental cues and plant behaviour. Essentially, a plant can be considered a computational unit that analyses the incoming environmental signals and optimises its physiology accordingly. This more holistic approach can help breeding, phenotyping and precision agriculture advance beyond current methods.

Examination Board #

  • Prof. Filip De Turck, PhD, Ghent University (chair)
  • Prof. Peter Bienstman, PhD, Ghent University
  • Prof. Joni Dambre, PhD, Ghent University
  • Prof. Sarah Garré, PhD, Instituut voor Landbouw-, Visserij- en Voedingsonderzoek & Université de Liège
  • Xu Zhang, PhD, imec, the Netherlands
  • Prof. Francis wyffels, PhD, Ghent University
  • Tom De Swaef, PhD, Instituut voor Landbouw-, Visserij- en Voedingsonderzoek
  • Michiel Stock, PhD, Ghent University