闭环神经机器人实验来检验神经网络的计算性能

The overall goal of the following experiment is to investigate the bidirectional interaction between a biological network of dissociated neurons and a small robot. The continuous dialogue can modify the computational properties of the neuronal system, inducing a kind of learning. This is achieved by preparing a neuronal culture over a micro electrode array in order to allow an experimental session of several hours without compromising cell activity.

As a second step, a response map of the culture is computed, which allows the selection of stimulating electrodes for the actual closed loop experiment. Next, coding and decoding parameters are defined in order to allow the exchange of information between biological and electronic systems. The results show a significant increase in navigational capabilities of the robotic agent based on the comparison of the average distance traveled between hits under different control conditions.

This method can help us answering key questioning in our engineering field, such as, can we arrive a dictionary to translate the language of the neurons to the language of the machines and vice versa? Generally, individuals new to this method will struggle because of the technological complexity of the experimental setup and software architecture. Demonstrating the procedure will be Martha Bizo PhD student of my lab.

The first step is to prepare a micro electrode array or MEA chip with 60 electrodes that can be used for stimulating or recording the neural network. After plating neuronal cultures on the ME, A chips wait about three weeks until the neural network has matured. On the day of the experiment, preheat the MEA heating system for five to 10 minutes by setting the temperature controller to 37 degrees Celsius and switching on the heating plate below the MEA itself.

Also use a heated cover to reduce evaporation using autoclave sterilized gas permeable caps cover the cultures with a cap to limit evaporation and prevent changes in osmolarity during recording. Now circulate car bear over the culture to maintain both stable oxygen and pH levels. Allow the culture to rest for 30 minutes After the rest period, record the spontaneous activity of the neuronal cells for 30 minutes.

Then save the data to a file by clicking the record button in the spikes box of the data recording form. In the raw data display form, identify the 10 channels with the highest spike count. Then select those channels in any of the MEA layouts.

The coding form is shown here by dragging the mouse cursor over the desired areas. Once the channels are selected, right click anywhere on the MEA layout and select add to left sensory area in the pop-up menu. These electrodes will be used to deliver electrical stimulation in a subsequent step.

Next, verify that the stimulator and MEA amplifier are correctly connected. All configurations require two wires per desired stimulation channel, while an extra coaxial cable is required to carry the synchrony signal. When ready, turn the stimulator on.

The next step is to define the stimulus parameters in the connection map form. All the stimulations delivered to the culture are by phasic square voltage waves. Set the half duration to 300 microseconds and set the amplitude to 1.5 peak tope voltage.

Begin recording the response to the stimulation by pressing the start button. In the connection map form a series of 30 stimuli at 0.2 hertz is automatically delivered from one of the previously selected stimulation electrodes. While this electrode is acting as the stimulation electrode responses are recorded from the remaining 59 electrodes on the MEA chip.

Then with a five second delay between series, the series of 30 stimuli are repeated in turn on each of the remaining nine identified stimulation electrodes while recordings are gathered from the remaining 59 electrodes. Next computer connection map for each stimulating channel using spy code, an application that performs computations on neural data. From the connection maps, discard from further analysis all stimulating electrodes that did not evoke responses.

Then from among the remaining electrodes, select the pair with the least overlap in responses as described in the text protocol, select one of these electrodes to code sensory information from the left side of the robot and the other to code readings from the right side. To do so, drag the mouse cursor over one electrode right click on the MEA layout, then select add to left sensory area. Then drag the cursor over the other electrode.

Right click and select add to right sensory area. Next in the coding form, select the coding scheme by setting coding type to linear. Then define the minimum and maximum stimulation rates by using the default range of 0.5 to two hertz and set the jitter parameter to zero.

Then in the decoding form, set the decoding algorithm parameters. That is the speed change and the time constant of decay parameters to one for a moderately active culture with approximately one spike per second per channel. Next, set the decoding algorithm burst parameters in the decoding form to zero.

The decay time is irrelevant if the speed change is zero. In the experiment manager form, select the data to be recorded by clicking the save spike data, robot data, and stimuli data check boxes. Now launch a pre-learning robot run by selecting the start experiment button in the experiment manager form.

When prompted, select new file names for the data files. Allow the experiment to run for 30 minutes. Then click the stop experiment button to stop the robot run.

Next, switch on the learning protocol by marking the deliver to 10 stimulation after. Hit checkbox in the experiment manager form. Then click the start experiment button again to perform the training robot run again for 30 minutes.

During this learning run, the stimulation from the robot changes from irregular to Titanic stimulus. When the robot hits an obstacle, the Titanic stimulus is used to train the neural network. After saving the data, switch off the learning protocol again by unchecking the deliver to tannic stimulation.

After hit checkbox, then click start experiment to perform a 30 minute post learning robot run. As before, remember to change file names to prevent overriding. In addition to the virtual robot shown previously, this same set of training runs can be used with a neural network and a physical robot.

Shown here is the path followed by a virtual robot during a 20 minute empty MEA experiment. There were no cells plated on the MEA for this control experiment. The light green areas are free for the robot to move in, and the dark green circles represent impassable obstacles that the robot can perceive through its distance sensors.

In each trial, the robot starts in the upper left section of the arena and travels to its final position depicted as a large pink dot. The smaller black dots represent hits against an obstacle. The color coded paths indicate the elapsed time.

In this experiment with no neural network, the robot was stopped by the first obstacle it encountered. Shown here is the path of a virtual robot during an open loop experiment in which the robot is effectively blinded. Instead of coding sensory information, the stimulation trains delivered to the neural network are just regular sequences.

Shown here is the path followed by a virtual robot during a 20 minute closed loop experiment in which the neural network received satanic feedback after the robot hit an obstacle. Note that unlike the open loop experiment, this robot successfully navigated around many obstacles. This indicates that a bidirectional interaction between the neuronal and artificial elements is necessary in order to obtain good navigation performances of the robot.

This graph shows the navigation performance of the robot expressed as pixels traveled between subsequent hits. The first two columns display the distribution of distances traveled, two control experiments, the empty MEA and the open loop configurations. The third and fourth columns display the performance without and with the delivery of Titanic stimulation.

Following each hit against an obstacle, the introduction of the Titanic stimulation significantly improves. The distance traveled between two consecutive hits, thus improving the navigation performances of the robot. This graph represents the probability for a given decoding algorithm of the robot to navigate successfully through a short track in a limited amount of time.

The decoding paradigms differ from one another because of the relative weights of bursts and isolated spikes After its development. This technique paved the way for URGs in the field of neuro robotics and neuroprosthetics to explore how to connect the brains and machines in order to improve the performances of modern neural interfaces. After watching this video, you should have a good understanding of how to plan and execute a hybrid experiment to investigate the computational properties of an embodied biological neural network.