How Brain-Computer Interfaces offer hope to paralyzed individuals
Repeated Neural Networks offer hope in a recent Stanford study (May 2021)
Becoming paralyzed has long meant an inability to perform just about every human function. Recent advancements out of the academic and industry settings have given hope to paralyzed individuals, in addition to those with similar neurodegenerative diseases. At the time of this writing, Brain-Computer Interfaces (BCI) are extremely narrow in their use and have yet to achieve the speed or accuracy needed for general use in patients. A major current limitation in BCI’s is the slow speed at which users are able to “type”. Before the publication of Stanford's study, BCI’s had only reached a max of 40 words per minute. The subject of this study achieved a benchmark of 90 words per minute.
The Stanford study is unique in three majors ways, and in doing so addresses some of the major weaknesses of past BCI’s.
Generalized Function- The paralyzed subject is able to express any sentence, not singularly a pre-determined sentence
Eye and Pace Freedom- The subject can move their eyes freely and perform the “thought” tasks at their own pace
Accuracy- The subject achieved 94.1% raw accuracy, while 99% accuracy was achieved with a “spell-check” like assistance.
A survey of Intracortical Microelectrode Array Technology
A challenge that BCI’s encounter is the necessity to achieve “single-neuron resolution” to address massive sensory and motor deficits. A generalized approach to addressing these deficits would lead to a lack of specificity in sensory and motor movements. Essentially these interfaces must be able to account for each neuron and properly channel that message through a pathway that was previously dead. We can think of this BCI through the lens of a toll-booth. Through either paralysis or other neurodegenerative disorders, the roads that leads to this “toll-booth” have been cut off and are no longer functional. This could happen through a variety of avenues. One could be through old age, equivalent to a road taking on the wear and tear of drivers for 65+ years, rendering the road unfunctional. On the other hand, the road could be diminished at a moments notice, such as when a human is paralyzed in a car accident. The goal of these BCI’s is to find new ways to channel the “information” from those roads through itself, in hopes that the right toll booths are present within the BCI to deliver that information to the proper destination.
A generalized BCI can be represented by a toll booth that lacks specificity in destination. Say the user wants to move a pen to scribe the letter “A”. A BCI that lacks specificity would not yield the letter “A” everytime. Even if it could, another toll booth would need to be created to differentiate between a capital A and a lowercase a. This example alone exemplifies the difficulty in creating effective BCI’s.
A Fascinating Feedback Loop
It is a fascinating concept to understand how BCI’s can allow paralyzed individuals to regain normal functions. The Feedback Loop most important to BCI’s is the relationship between recording and stimulation. Recording was the focus of the aforementioned Stanford study, in which they quite literally recorded the letters of the alphabet the subject was able to perceive in his own mind. The way in which recording becoming mainstream and “Sci-Fi like” is when the human subject is able to manipulate the exterior environment through a prosthetic. The idea is based off the fact that the motor cortex is still functional, however the pathways to motor output are non-functional. This type of recording would encompass the BCI recording thousands of action potentials, and properly associating the correct action potentials to the correlative motor output in a hypothetical prosthetic. We will discuss how this proper association is accomplished via Machine learning later/
The importance of the prosthetic hand doesn’t end with motor output. The robotic hand would be equally important in perceiving external stimuli, and channeling that properly through the BCI to allow the subject to properly perceive and subsequently react to it’s external environment.
It’s imprortant to note that BCI’s don’t use traditonal neuromuscular pathways. The BCI simply receives and interprets signals from the Central Nervous System, and through training relays that output to the external device.
A Roadblock to Adoption and Functionality
While the Stanford Study occured across many days, the mainstream use of BCI’s necessitates that they function at a high level for years. At the time of this writing, that maximum lifespan of BCI’s appears to be about 1,000 days, which equates to just over a year and a half. It also remains to be seen whether the strength of the chip remains stable over this time period. While the subjects of these clinical trials are predominantly individuals over 70 years, the long term outlook and drive of neurotechnologists will necessitate an improvement in longevity of BCI’s. The likes of Neuralink and Neuropace don’t have their sites set on treating just geriatrics. The young population suffering from neurodegenerative diseases will require the long term scaling of BCI’s.
Training that Chip in your Head
Above is helpful diagram from a review paper, which was a truly unblemished paper of BCI’s as they relate to students with BCI’s.
Common BCIs contain EEGs which have the ability to pickup on brain activity through receipt of signals from neurons. How these signals are interpeted and translated into motor actions involves machine learning. It is also important to note some basic structural principles of neurons. Neurons have both inputs and outputs. The inputs of neurons are their dendrites while the outputs are axons. The inner-workings of these complex structures reveal just the tip of the iceberg of neurodegenerative diseases. For instance, myelin sheath, which essentially coat the axon to provide efficiency for the transfer of the signal down the neuron are implicated in Multiple Sclerosis (MS). Your body attacks this myelin in MS, leading to a downstream effect that essentially disables the bodies ability to communicate and move.
So what is machine learning and why is it important in BCI’s?
Without machine learning and neural networks, the BCI and its EEG are just sensors. Machine learning allows for the BCI to be “trained” properly such that the proper inputs lead to the appropriate outputs. Neural networks sounds like a super complicated system. In reality, neural networks are simply a group of neurons, whether organic or inorganic, that work together. The training implicated in BCI’s is how scientists, such as those at Stanford, are able to get these groups of neurons to effectively mimic the behavior of a fully functional brain. The neural network that these groups use are algorithims that learn patterns. A simple example is this:
The network must learn that a specific neuron firing (input) is responsible for moving the right thumb toward the index finger (output).
Two main functions of neural networks is CLASSIFYING and CLUSTERING inputs.
Classifying data is how we “supervise” the machine to associate an input with our desired output. An essential aspect of classifying is putting a direct LABEL on the data. This is your traditional computer. The computer is fast at accomplishing this task, and rarely, if ever, messes it up.
Clustering is how machines associate similarities between outputs. There are NO labels associated with clustering. If classified learning is supervised by human output, clustering is unsupervised learning. This is difficult, but represents the greatest ability of the BCI to learn at a higher level. Clustering also allows the neural network to “learn” a higher quantity of data.
Supervised vs Unsupervised Learning
As aforementioned, unsupervised learning opens up a world of opportunity to increase the quantity and accuracy of data taken in by the BCI. We’ll now look at an example from the Stanford Study and how the BCI can be trained to properly represent what the subject is theoretically thinking. In the study, the subject is attempting to write sentences. Thus the neural network must convert the subjects neural activity “into probabilities describing the likelihood of each character being written at each moment in time”.
Above is the most fascinating image from the Stanford Study, and does a great job creating a visual for the reader. We can read the diagram from left to right. The experiment begins when the subject is thinking of the word paper. In order to come up with those 5 letters, thresholds must be met which then correlate with the algorithm that the RNN (repeated neural network) has learned to associate with an output, such as the letter “p”. We see in the top arrow that the subject mispelled the word “the”. The h is replaced by the letter n. As aforementioned, the autocorrect feature, which represents the unsupervised learning of the BCI was able to come up with the correct offline output, ie spelling “the” correctly, leading to a much higher accuracy level. This offline feature was trained on a large-vocabulary language model, which is highly capable of successfully predicting likely words.
As Ardi Tarca brilliantly wrote, “The goal in supervised learning is to design a system able to accurately predict the class membership of new objects based on the available features.” The available features in the Stanford Study is the threshold crossing feature. Downstream, the RNN can accurately predcict which letter is “being thought of” based on that available feature.
Closing Thoughts
Brain Computer Interfaces offer a tremendous opportunity to use the marvels of the engineering and mathematics world to alleviate terminal biological conditions. Proof of Concepts have properly demonstrated that this novel technology has the potential to address problems and diseases that appeared to be the end of road for many. It will be exciting to see and understand how Machine Learning and the engineering side of this world help further address the weaknesses of BCIs and lead to even better outcomes.
Sources:
https://www.sciencedirect.com/science/article/pii/S0896627320308114
https://www.nature.com/articles/s41586-021-03506-2
https://wiki.pathmind.com/neural-network
https://www.emotiv.com/bci-guide/
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3497935/
I wonder what the possible implications of BCI and machine learning are in an educational setting to support individuals with profound learning disabilities. The advances you describe can translate into an avenue of access for many additional populations.