Software that can be taught to refine the information sent from a bionic eye to its wearer is being trialled in Germany.
Retinal implants can restore some vision to blind or partially blind people by taking over the job of turning light into signals transmitted to the brain. So far, about 10 people in Germany and 15 in the US have been fitted with such implants although expanded US trials are planned.
鈥淭hese people report seeing light and dark and maybe some limited fuzzy shapes,鈥� says Rolf Eckmiller, a computer scientist at Bonn University in Germany. 鈥淏ut they don鈥檛 have any gestalt perception.鈥�
Advertisement
Eckmiller says the secret to improving these implants is to match the signals they produce with the signals that a healthy eye sends to the brain. One team in California, US, is trying to do that by building a copy of the retina鈥檚 neurons in silicon. Eckmiller, along with colleagues Oliver Baruth and Rolf Schatten, plan to use learning software instead.
Retina encoder
In their system, a camera feeds information to a 鈥渞etina encoder鈥� 鈥� software that mimics the image processing done by a healthy retina. 鈥淚t has hundreds of different parameters [that can] be properly tuned,鈥� says Eckmiller. 鈥淏ut only one setting is appropriate to allow proper perception.鈥�
So the Bonn team is developing software that learns the correct settings from a user.
It does this through a 鈥渄ialogue module鈥� that tries different settings while a user looks at standard shapes. The user selects the three settings that most closely match the real shape and the software then presents six more settings based on these three. Over time, the system learns to produce a signal that provides a more accurate picture to the user鈥檚 brain.
Trials involving more than 50 sighted people have been promising, says Eckmiller. In tests, participants wear a visor with a built-in camera and eye displays. The tests differ slightly from the way the system would work with a real implant, but the same learning procedure takes place.
Random setting
The output of the camera feeds information into a retinal encoder that has been set to produce a random visual signal. A second module must be trained, by the user, to decode the output into a coherent image, which is displayed on the visor.
Motion sensors on the visor allow a wearer to control the system with head movements. Successive rounds of the process can transform an almost shapeless collection of dots into a recognisable replica of the shape. Eckmiller now intends to start testing the training system on people with real implants.
But James Morrison, who is also developing retinal implants with colleagues at Glasgow University, UK, says exactly reproducing the signals the brain normally uses is not currently a high priority for the field. 鈥淐urrently the biggest challenge is to make a working device that interfaces with neurons properly,鈥� he says.
Furthermore, because trials have been limited, it may turn out that the brain can learn how to interpret an implant鈥檚 signal, he adds. 鈥淚 think it鈥檚 too early to tell.鈥�
