[Dee]

What we are doing presently with ANN are working with biology-inspired neurons that is the neurons are not much like biological neurons, only some inputs (dendrites), dot (soma), and output activation (axon terminal).

In reality of biological neurons, they are supposed to be much more complex with dendrite combinations (like xi+xj before multiplication with weight), soma functions, and multiple axon outputs of similar values, etc.

Should artificial neurons ever be bio-simulated instead of bio-inspired?

[HS]

Regarding artificial general intelligence (AGI), currently yes. It seems like we’re still at a level of knowledge/understanding where we’ll learn best by imitating nature’s fully functional systems. Once we get a grasp of the principles, it will be easier to create something familiar yet augmented, which seems to be the predominant goal. The "out there" approaches are cool and exciting, just probably a safer bet for hobbyists.

[infurl]

Should artificial neurons ever be bio-simulated instead of bio-inspired?

There have been attempts to simulate neurons at the biological level but the world's computing resources are not adequate to do this on more than a rudimentary scale. Each neuron is more than just a logic gate or a memory bit, it's an entire computer system, and the brain consists of a network of billions of them. Not only that, it's incredibly efficient. A human brain uses about 25 watts of power and yet all the computers in the world can't match its computational power.

On the other hand, Jeff Hawkins and his team at Numenta believe they have identified the essential elements of neurons and the data structures that they represent. They call this hierarchical temporal memory (HTM) and they've published quite a lot of information about it. Check out Numenta's YouTube channel and other sources for more details. They have working hardware and software that is bio-inspired as a result of this research.

https://numenta.org/

[Dee]

Regarding artificial general intelligence (AGI), currently yes. It seems like we’re still at a level of knowledge/understanding where we’ll learn best by imitating nature’s fully functional systems. Once we get a grasp of the principles, it will be easier to create something familiar yet augmented, which seems to be the predominant goal. The "out there" approaches are cool and exciting, just probably a safer bet for hobbyists.

Most of the base maths and mechanism for ML were done in the last century, with some types of neurons: basic dense, convo, rnn, lstm, etc.
Hopefully in this century some guys will invent more complex neuron types.

[Dee]

Should artificial neurons ever be bio-simulated instead of bio-inspired?

There have been attempts to simulate neurons at the biological level but the world's computing resources are not adequate to do this on more than a rudimentary scale. Each neuron is more than just a logic gate or a memory bit, it's an entire computer system, and the brain consists of a network of billions of them. Not only that, it's incredibly efficient. A human brain uses about 25 watts of power and yet all the computers in the world can't match its computational power.

On the other hand, Jeff Hawkins and his team at Numenta believe they have identified the essential elements of neurons and the data structures that they represent. They call this hierarchical temporal memory (HTM) and they've published quite a lot of information about it. Check out Numenta's YouTube channel and other sources for more details. They have working hardware and software that is bio-inspired as a result of this research.

https://numenta.org/

The 25 watts fact is interesting and the billions of neurons may ever be simulated using quantum computers as quantum computers work on particle scale, states of particles. However, it's still far ahead to think of doing something with quantum machines popularly since  IBM only released the first commercial IBM Q System One last year

The HTM seems promising that even LSTM by now only stores 1 or 2 inputs of the most recent past and future.