There has been some recent research showing that using human neuron based Brainoware is able to achieve a higher sample efficiency in some tasks compared to conventional ML techniques.
It is very difficult to keep these biological systems alive so it is difficult to directly compare sample efficiency (it is a lot cheaper to train a digital neural network compared to a physical one). However, it is possible that this biological hardware (and underlying algorithms) has such a significant advantage for some tasks that it will outperform digital AI with more research.
This market will resolve YES if research is published at any point within the next 10 years in which a biological neural network is trained to solve a learning problem which is not feasible to solve with the digital algorithms available at the time. To be clear I am referring to an "organoid reservoir" not a human or animal brain trained to complete a task. The organoid should be grown in a lab which ensures that it is not performing well based on previous life experience.
1,000While it might be unlikely I believe that there are still many inefficiencies with current ML algorithms (and neural networks). The research on brain organoid computing has show that in some reinforcement learning problems the biological systems can learn with 5 samples what a digital neural network takes 50 samples to learn. This means there is something fundamentally more efficient about the biological systems compared to current ML algorithms.
@mjmandl is there a way to make money off it before the tech is mature? The ml paradigm is working because it pays for itself quickly.
@singer I think research in this field could provide real boosts to research in ML. Similar to how biological systems (animals) have provided inspiration in the field of robotics; biological brains can help us improve digital brains. I think the fact that we are seeing a 10x better sample efficiency in smaller problems could imply a higher sample efficiency for larger problems too. Imagine being able to train a self driving system with 10x fewer samples (or maybe even fewer if it scales better to large problems).