Our brain is the world’s most complex and advanced supercomputer. Ultimately, it is this bio-supercomputer that will help evolve Artificial Intelligence (AI) to the next inception. Predictions state our brains will achieve this feat using brain-computer interfaces or BCIs, a direct partnership between the central nervous system and an AI algorithm. Presently, these systems meld the human and the machine to enable the control of assistive devices in individuals with severe motor impairments.
Still, to harness the complex processes of the brain to train AI, continuous-learning must emerge from machine-learning. Continuous-learning is a subset of machine learning, where AI progressively updates in an unbroken fashion, attaining knowledge across multiple data fields. In theory, this is an infinite process, a singularity in effect. Thus, this evolved artificial intelligence will be updating continually, self-actuating its self-awareness. For this to come to fruition it is posited continuous data will have to be supplied to AI systems.
The discipline encompassing brain-computer interfaces, known as neuromorphic engineering, has gained momentum in recent years. Despite this fact, existing neuromorphic systems still have to be reset and recalibrated each day. Subsequently, they are presently unable to run alongside, feed on, and consolidate with the brain’s natural learning processes. In short, this lack of continuous data and learning has caused major setbacks for higher cybernetics.
Continuous data in cybernetics
Now, a study from researchers at UC San Francisco develops a BCI that facilitates the continuous control of a computer cursor by a paralyzed individual, without the need to recalibrate the machine daily. The team states continuous data attained from a brain implant allowed the human brain and AI to merge completely. Importantly, this was all achieved without resetting the whole system daily. The study is published in the journal Nature Biotechnology.
Previous studies show adapting AI to merge with and learn from the brain has never been achieved in a person with paralysis. The aforementioned is disappointing as full immersion of AI into the central nervous system is expected to evolve these artificial systems. Likewise, continuous data is required to achieve this. As a result, continuous data to AI will potentially lead to continuous-learning artificial systems.
The next evolution of AI
For the moment, BCIs are unable to provide AI with unceasing data as they need to be recalibrated every day. Consequently, continuous data from the brain is not received or transmitted by these systems, and continuous-learning is not possible.
The aforementioned is due in part to the type of brain implants used to connect the human brain with the machine. Usually, ‘sharp’ invasive electrodes are employed here. These are electrodes that penetrate the brain tissue deeply for more sensitive recordings, notorious for shifting or losing signal over time.
In contrast, an electrocorticography (ECoG) array comprises a square or rectangular pad of electrodes placed on the surface of the brain. Hence, they allow uniform, long-term recordings of neural activity as they cover a larger area of the brain than sharp electrodes.
In comparison, these ECoGs are not as sensitive as sharp electrodes, yet they offer more stability. It is in this way ECoG arrays provide consistent data from the brain. The current study investigates whether ECoG arrays implanted on the brains of paralyzed patients can provide continuous data.
Integrating AI & the human brain
The current study develops a machine-learning AI to amalgamate a paralyzed patient’s brain activity with a computer. The purpose of this was to allow the user to operate a computer cursor using their brainwaves alone.
The ECoG array recorded and converted the participant’s brain signals using machine-learning to control a computer cursor onscreen. These implants were already present due to a prior trial involving the control of neuroprosthetic limbs. Furthermore, the individuals taking part in both trials have established paralysis of all four limbs (tetraplegia).
To begin with, the researchers reset the brain-computer interfaces at the start of training every day. Accordingly, this process severely limited the level of control achieved. Thus, the previous day’s progress was rarely surpassed.
The algorithm was then allowed to continue updating to match the participant’s learned memory of the system without resetting it each day. It was observed this unbroken interaction between the brain and the machine resulted in continuous data. Moreover, this caused improvements in performance and patient experience over many days.
Interestingly, the AI improved even further by ensuring the algorithm wasn’t updating faster than the brain could follow. This worked out at a rate of about once every 10 seconds, with AI becoming an extension of the user.
Continuous learning in BCIs
Results show the participant’s brain was able to amplify salient patterns of neural activity used by the AI as they began to merge. These were the signals the brain used to integrate with AI via the ECoG array. Indeed, this human-machine hybridization was achieved while simultaneously eliminating less effective brain signals in a pruning process mirroring the brain when it is learning any complex task.
Data findings show the participant rapidly re-established the same patterns of neural activity for controlling the computer cursor when the BCI was reset after several weeks. In effect, converging with and retraining the machine-learning instantly, going beyond transfer-learning.
In a crucial outcome, once the participant mastered control of the computer cursor onscreen, the AI stopped updating itself. Hence, the patient could control the BCI every day without any retraining or recalibration. In effect, the AI had fully merged with the brain to enable the human to perform augmented tasks.
The lab states the performance of the cybernetic human did not decline over forty-four days in the absence of retraining. Likewise, the participant could go days without using the BCI, exhibiting little to no decline in performance.
Training the AI to do more
Additionally, the establishment of a stable learned state in the form of moving a cursor via a BCI allowed the ‘stacking’ of additional skills. For example, where a patient learns extra functions such as ‘clicking’ a virtual button. Once again, the patients used only their brain signals and continuous data with no loss of performance.
The team surmises they have successfully married continuous data with brain-computer interfaces. Moreover, their novel BCIs, a prelude to continuous-learning, do not need to be reset every day. For the future, the researchers state these data show ECoG-based BCIs could potentially herald the actual evolution of AI via continuous learning.
Don’t miss the latest discoveries from the health innovator community: