SURE: Web Posters from SURE 2002

Neurally Controlled Robotic Drawing Arm - MultiElectrode Array Art (MEART)
Alexander C. Shkolnik, Guy Ben-Ary, Thomas B. DeMarse, Iain Sweetman, Phil Gamblen, Ionat Zurr, Oron Catts, Steve M. Potter
Lab for Neuroengineering, Georgia Inst. of Tech., Atlanta, GA; SymbioticA Art & Science Lab, School of Anatomy and Human Biology, University of Western Australia, Perth; Dept of Biomedical Engineering, Georgia Inst. of Tech. and Emory Univ., Atlanta, GA

Abstract

Traditional artificial intelligence and computer science approaches have had limited success in producing machines capable of matching the intelligence of even the simplest of organisms. The long-term goal of our lab is to learn how to take advantage of the synthesis between biological and artificial systems to produce a machine that can handle difficult tasks in real time. The current project explores topics such as creativity and plasticity (learning) in such a hybrid system. Dissociated embryonic cortical rat tissue is cultured on a MultiElectrode Array (MEA), which acts as an interface between the neurons and a computer. Software records and analyzes electrical data from the electrodes, and encodes stimulation patterns that are fed back to these electrodes. A simplified form of post-analysis data is sent to a computer at the Biennale of Electronic Arts Perth Conference (BEAP) in Australia once per second, and this data is used to control a robotic drawing arm. Video snap shots and audio from the audience is sent back to our lab, and this information is encoded into stimulation patterns to provide a feedback system for the neural network. Several interesting questions arise from such a system: are the drawings produced a form of truly creative art? Can patterns of dynamic stimulation (encoded in real-time based on video or audio feedback from the audience) be linked to patterns on paper? If given static stimulation (perhaps if the stimulus represents a snapshot for an extended period of time), are the drawings reproducible or will they change over time, and if so is this a form of learning?

Introduction

MEART is a bio-cybernetic research & development project exploring aspects of creativity and artistry in the age of new biological technologies. The first public outcome of this project, called Fish & Chips was displayed at the Arts Electronica Festival, Australia, in 2001. The robotic control segment consisted of fish neurons cultured directly on silicon chips. A good connection was made with only a few neurons, and thus the robotic drawing arm was controlled by a small selection of neurons from within a larger network. Stage 2, on display at BEAP from July 31 – Aug 26, 2002, is now known as MultiElectrode array ART (MEART). Brain tissue from rat cortex is cultured on a dish over a grid of 60 electrodes in a 2mm2 area (the MEA). Each electrode may record from several neurons and also stimulate these same neurons. Thus, compared to stage 1, a much greater amount of information from the living neural network is available to provide control in the robot. Providing the cultured neurons with this robotic “body” sets up a framework in which we can conduct experiments to study neural plasticity and data encoding within cultured neural networks.

Methods and Materials

Cell Culture

  • We used 60-electrode glass MEAs from MultiChannel Systems, 10-um diameter electrodes and 200um or 500um inter-electrode spacing.
  • Polyethylene imine solution was used to coat the bottom of the dish to enhance electrical signals.
  • Laminin was applied to allow the cells a substrate to adhere to.
  • Cells were prepared using papain digestion of embryonic-day-18 rat frontal lobe cortex, and applied to the dish.
  • To prevent osmolality change due to evaporation and to prevent infection, the MEA was sealed with fluorinated ethylene propylene, a membrane permeable to CO2 and O2, but impermeable to water vapor and microbes.

Software provided by MultiChannel systems for recording was modified and enhanced for this project: A module was built to integrate the activity of a given electrode over a period of 500 milliseconds. The value for each electrode decays exponentially over time, but is incremented with each spike at that electrode. At the end of the 500ms interval the value for each electrode is normalized over 0 – 256 and sent via TCP connection to a computer at BEAP.

Robotic Arm Movement: At the current stage, the 60 electrode values received in BEAP are used to calculate regions in the dish particularly active, and the robotic arm draws a line from its current position to the newly calculated position.

Feedback: Every 30 minutes a picture from a camera at BEAP is sent to the Lab for Neuroengineering via TCP. The image is converted to a stimulation pattern to be applied on 8 electrodes in the MEA. Another system available for stimulation involves a real-time approach, whereby audio is recorded at BEAP and encoded into a 1 bit signal used to turn a programmed pattern of stimulation OFF and ON.

Results

From the first drawing produced by MEART, one can see several concentrations of lines, which look like ink blots. These concentrations represent the most active channels in the dish. A line is drawn between different regions when there is a change in activity in the MEA. The drawing above does not reflect any feedback systems.

Conclusions and Future Studies

Through the month of August while MEART is operational we hope to experiment with varying feedback systems and drawing algorithms. For example, instead of moving the arm to a point on paper representing the most active region of the MEA, we can classify signals received in the MEA by various patterns using a back-propagation artificial neural network, and use detected patterns to indicate desired movement in the robotic arm. By comparing various forms of feedback stimulation we will find how static stimulation (same stimulation over a long period) affects the system compared to dynamic stimulation (in real-time) and, of course, compared to the system working with no stimulation as shown above. Through these comparisons we hope that plasticity in the network will be expressed on paper. In conclusion, MEART takes the basic components of the brain (isolated neurons), and attaches them to a mechanical body through the mediation of a digital processing engine to attempt and create an entity that will seemingly evolve, learn and become conditioned to express its growth experiences through “art activity”. The combined elements of unpredictability and “temperament” with the ability to learn and adapt, create an artistic entity that is both dependent, and independent, from its creator and its creator’s intentions.

Acknowledgements and Funding Attributions

This material is based upon work supported by the Howard Hughes Medical Institute under Grant No. 52003071 and by the CBN - STC Program of the National Science Foundation under Agreement #IBN-9876754

In Plain English

Rat brain tissue is cultured on a dish over a grid of electrodes. The electrodes are used to record signals from the brain tissue, and to stimulate the brain tissue. In this way we are able to "talk" to the brain tissue using a computer. The computer is used to send signals to a robotic arm in australia which draws pictures in real-time. We hope that these drawings may provide us with a new way of looking at the data encoded in the brain tissue.