FOR those who reckon that brain-computer interfaces will never catch on, there is a simple answer: they already have. Well over 300,000 people worldwide have had cochlear implants fitted in their ears. Strictly speaking, this hearing device does not interact directly with neural tissue, but the effect is not dissimilar. A processor captures sound, which is converted into electrical signals and sent to an electrode in the inner ear, stimulating the cochlear nerve so that sound is heard in the brain. Michael Merzenich, a neuroscientist who helped develop them, explains that the implants provide only a crude representation of speech, “like playing Chopin with your fist”. But given a little time, the brain works out the signals.
That offers a clue to another part of the BCI equation: what to do once you have gained access to the brain. As cochlear implants show, one option is to let the world’s most powerful learning machine do its stuff. In a famous mid-20th-century experiment, two Austrian researchers showed that the brain could quickly adapt to a pair of glasses that turned the image they projected onto the retina upside down. More recently, researchers at Colorado State University have come up with a device that converts sounds into electrical impulses. When pressed against the tongue, it produces different kinds of tingle which the brain learns to associate with specific sounds.
The brain, then, is remarkably good at working things out. Then again, so are computers. One problem with a hearing aid, for example, is that it amplifies every sound that is coming in; when you want to focus on one person in a noisy environment, such as a party, that is not much help. Nima Mesgarani of Columbia University is working on a way to separate out the specific person you want to listen to. The idea is that an algorithm will distinguish between different voices talking at the same time, creating a spectrogram, or visual representation of sound frequencies, of each person’s speech. It then looks at neural activity in the brain as the wearer of the hearing aid concentrates on a specific interlocutor. This activity can also be reconstructed into a spectrogram, and the ones that match up will get amplified (see diagram).
Algorithms have done better than brain plasticity at enabling paralysed people to send a cursor to a target using thought alone. In research published earlier this year, for example, Dr Shenoy and his collaborators at Stanford University recorded a big improvement in brain-controlled typing. This stemmed not from new signals or whizzier interfaces but from better maths.
One contribution came from Dr Shenoy’s use of data generated during the testing phase of his algorithm. In the training phase a user is repeatedly told to move a cursor to a particular target; machine-learning programs identify patterns in neural activity that correlate with this movement. In the testing phase the user is shown a grid of letters and told to move the cursor wherever he wants; that tests the algorithm’s ability to predict the user’s wishes. The user’s intention to hit a specific target also shows up in the data; by refitting the algorithm to include that information too, the cursor can be made to move to its target more quickly.
But although algorithms are getting better, there is still a lot of room for improvement, not least because data remain thin on the ground. Despite claims that smart algorithms can make up for bad signals, they can do only so much. “Machine learning does nearly magical things, but it cannot do magic,” says Dr Shenoy. Consider the use of functional near-infrared spectroscopy to identify simple yes/no answers given by locked-in patients to true-or-false statements; they were right 70% of the time, a huge advance on not being able to communicate at all, but nowhere near enough to have confidence in their responses to an end-of-life discussion, say. More and cleaner data are required to build better algorithms.
It does not help that knowledge of how the brain works is still so incomplete. Even with better interfaces, the organ’s extraordinary complexities will not be quickly unravelled. The movement of a cursor has two degrees of freedom, for example; a human hand has 27. Visual-cortex researchers often work with static images, whereas humans in real life have to cope with continuously moving images. Work on the sensory feedback that humans experience when they grip an object has barely begun.
And although computational neuroscientists can piggyback on broader advances in the field of machine learning, from facial recognition to autonomous cars, the noisiness of neural data presents a particular challenge. A neuron in the motor cortex may fire at a rate of 100 action potentials a second when someone thinks about moving his right arm on one occasion, but at a rate of 115 on another. To make matters worse, neurons’ jobs overlap. So if a neuron has an average firing rate of 100 to the right and 70 to the left, what does a rate of 85 signify?
At least the activities of the motor cortex have a visible output in the form of movement, showing up correlations with neural data from which predictions can be made. But other cognitive processes lack obvious outputs. Take the area that Facebook is interested in: silent, or imagined, speech. It is not certain that the brain’s representation of imagined speech is similar enough to actual (spoken or heard) speech to be used as a reference point. Progress is hampered by another factor: “We have a century’s worth of data on how movement is generated by neural activity,” says BrainGate’s Dr Hochberg dryly. “We know less about animal speech.”
Higher-level functions, such as decision-making, present an even greater challenge. BCI algorithms require a model that explicitly defines the relationship between neural activity and the parameter in question. “The problem begins with defining the parameter itself,” says Dr Schwartz of Pittsburgh University. “Exactly what is cognition? How do you write an equation for it?”
Such difficulties suggest two things. One is that a set of algorithms for whole-brain activity is a very long way off. Another is that the best route forward for signal processing in a brain-computer interface is likely to be some combination of machine learning and brain plasticity. The trick will be to develop a system in which the two co-operate, not just for the sake of efficiency but also for reasons of ethics.