How Can Engineered Neuronal Networks Advance Neurocomputing?

Feedback Stimulation and Reservoir Computing

Question from the audience
‍Your feedback is a spatial stimulation, meaning that depending on the amplitude of the signal you choose the stimulation electrode (I hope I got this correctly). This itself is a non-linear mapping, do you not solve / create a reservoir already at this stage? Meaning you could remove the BNN and simple take the stimulation pattern directly as your reservoir and hence be way more efficient. (I guess my second question was answered by the fact that all electrodes that are stimulated get the same pulse at the same time).

Prof. Yamamoto's answer
‍In our system, the feedback is not implemented by dynamically choosing a stimulation electrode according to the instantaneous output. Rather, feedback stimulation is delivered through a predefined set of electrodes. For this particular work, we did not test direct regression from the feedback stimulation patterns without the BNN. However, in our previous work (Sumi et al., PNAS 2023), we showed that BNNs can contribute to the system by transforming input patterns into neural activity patterns that better support generalization.

Open Challenges in Biological Computing

Question from the audience
‍Have you faced any open problems about biocomputing? Specifically theoretical ones

Prof. Yamamoto's answer:
‍Quite a lot. Variability and instability are obvious ones, for instance. I think the interesting point is that biology somehow works under these limitations, and there must be a way to decode information accurately out from highly variable sources.  

Applications of Biological Neural Networks and Comparison with ANNs

Question from the audience
‍What are the possible applications of the engineered biological neural networks besides understanding the neuroscience better? Will it outperform artificial neural networks? There's a paper by Beniaguev et al. (Neuron 2021) that shows that it takes 5-8 layers of deep neural networks to approximate the input-output mapping of a single biological neuron.

Prof. Yamamoto's answer:
‍That could, in a way, be taken as a characteristic where biology is "outperforming" ANNs, but I think comparing BNNs and ANNs is like comparing apples to oranges. Scientifically, it's more important to look for its own applications.

Training Biological Neural Networks Across Multiple Days

Question from the audience
‍If you train the same network over different days, do you find that they become easier to train?

Prof. Yamamoto's answer:
We are aware that some papers, e.g., Cai et al. (Nature Electronics 2023) and Mritunjay et al. (Nature Electronics 2026), that have looked into those aspects. We have not tested them yet in our systems, but I agree that they are very interesting profiles.