Biological organisms learn from interactions with their environment throughout their lifetime. For artificial systems to successfully act and adapt in the real world, it is desirable to similarly be able to learn on a continual basis. This challenge is known as lifelong learning, and remains to a large extent unsolved. In this perspective article, we identify a set of key capabilities that artificial systems will need to achieve lifelong learning. We describe a number of biological mechanisms, both neuronal and non-neuronal, that help explain how organisms solve these challenges, and present examples of biologically inspired models and biologically plausible mechanisms that have been applied to artificial intelligence systems in the quest towards development of lifelong learning machines. We discuss opportunities to further our understanding and advance the state of the art in lifelong learning, aiming to bridge the gap between natural and artificial intelligence.

Objective We studied the fundamentals of muscle afferentation by building a neuro-mechano-morphic system actuating a cadaveric finger. This system is a faithful implementation of the stretch reflex circuitry. It allowed the systematic exploration of the effects of different fusimotor drives to the muscle spindle on the closed-loop stretch reflex response. Approach As in Part I of this work, sensory neurons conveyed proprioceptive information from muscle spindles (with static and dynamic fusimotor drive) to populations of α-motor neurons (with recruitment and rate coding properties). The motor commands were transformed into tendon forces by a Hill-type muscle model (with activation-contraction dynamics) via brushless DC motors. Two independent afferented muscles emulated the forces of flexor digitorum profundus and the extensor indicis proprius muscles, forming an antagonist pair at the metacarpophalangeal joint of a cadaveric index finger. We measured the physical response to repetitions of bidirectional ramp-and-hold rotational perturbations for 81 combinations of static and dynamic fusimotor drives, across four ramp velocities, and three levels of constant cortical drive to the α-motor neuron pool. Results We found that this system produced responses compatible with the physiological literature. Fusimotor and cortical drives had nonlinear effects on the reflex forces. In particular, only cortical drive affected the sensitivity of reflex forces to static fusimotor drive. In contrast, both static fusimotor and cortical drives reduced the sensitivity to dynamic fusimotor drive. Interestingly, realistic signal-dependent motor noise emerged naturally in our system without having been explicitly modeled. Significance We demonstrate that these fundamental features of spinal afferentation sufficed to produce muscle function. As such, our neuro-mechano-morphic system is a viable platform to study the spinal mechanisms for healthy muscle function — and its pathologies such as dystonia and spasticity. In addition, it is a working prototype of a robust biomorphic controller for compliant robotic limbs and exoskeletons.

Tendon transfer surgery is often used to restore hand grasp function following high median-ulnar nerve palsy. This surgery typically reroutes and sutures the tendon of the extensor carpi radialis longus (ECRL) muscle to all four flexor digitorum profundus (FDP) tendons of the hand, coupling them together. This makes it difficult to grasp irregularly shaped objects. We propose inserting a novel implantable passive device between the FDP tendons to surgically construct a differential mechanism, enabling the fingers to individually adapt to the irregular contours during grasping. These passive implants with no moving parts are fabricated from biocompatible materials. We tested the implants’ ability to create differential flexion between the index and middle fingers when actuated by a single muscle in two human cadaver hands using a computerized closed-loop control paradigm. In these cadaveric models, the implants enabled significantly more differential flexion between the index and middle fingers for a wide range of donor tendon tensions. The implants also redistributed fingertip forces between fingers. When grasping uneven objects, the difference in contact forces between fingers reduced by nearly 23% compared to the current suture-based surgery. These results suggest that self-adaptive grasp is possible in tendon transfers that drive multiple distal flexor tendons.