dopetalk
Mind and Body => Neuroscience => Topic started by: Chip on May 14, 2018, 09:35:43 PM
-
source: https://www.quora.com/Is-there-a-significant-difference-between-a-human-brain-and-a-simulated-neural-network/answer/Jean-Pierre-Legros
By Jean-Pierre Legros, M.D. Author of 'Stratium' a theory of mind layered from biology to consciousness
The question with its details is actually: « Are the differences between brain and simulated network so important that simulation has no chance to reproduce all the capabilities of the brain? »
Let’s take a response from Paul King on neural networks, which deserves discussion. A neural network has input-output evaluated by a supervisor, and between the two has "hidden layers" analyzing information by successive stages of processing correcting previous (deep learning). This resembles very closely the human brain, which receives multiple sensory inputs and manipulates them to the conscious exit. So the question is legitimate: How many hidden layers has the human brain?
Paul King refutes the comparison on several arguments:
1) The brain is not organized in layers but interconnected centers.
2) The neural connections are uni-directional; no feedback.
3) The brain does not learn from provided accurate answers.
4) The neurons use nerve impulses that do not match the numerical algorithms.
5) Paul think multiple layers of deep learning are no better than some, because too much information is lost on the way.
In fact the neurons are organized in successive layers of treatment (within different interconnected nerve centers) but the current artificial networks are too coarse to simulate correctly. It takes a lot of levels to them to simulate some of the neurons, for several reasons:
1) The neurons have their own feedback control: their firing is being depleted. Part of nerve impulses is feedback propagation.
2) The neurons alter their physiology based on their activity (glial cells also). They are already self-learners.
3) The feedback control is exercised through different connections and there exist inhibitory neurons.
4) Neurons fire spontaneously. This intrinsic activity produces an 'exit' in the absence of 'input'. The treatment is not just a passive process.
5) The latencies between neural activations, by adding and delaying the feedback control, provide independence on each processing stage. Other related information can meanwhile prevent the feedback control to act.
All these characteristics of neurons are that their system is both more dynamic and more stable than artificial networks. The addition of levels in the way that they are designed in artificial network only reduces the margin of error on a conceptual treatment, while a group of neurons cares little to go wrong, lives in the illusion of doing always a great job, we might say, and it is the group of supervisors neurons, above, who assesses and refers corrections invisibly for supervised neurons. The independence of treatment levels is marked for neurons, which explains how we experience our consciousness, the highest stage: merged, independent, yet connected to a multitude of concepts presented.
Designers of artificial networks have this improvement to do: increase the independence of their processing stages while keeping them coordinated. Embed latency rather than use their electronic speed. Nothing prevents theoretically they can achieve success, and therefore, to answer your question, artificial networks are potentially capable of simulating the contents of human consciousness and experience them. What will experience will be a heap of transistors rather than a set of biological cells. This will make quilled weddings ...
Last note: yes, the brain learns also from responses provided: those of his parents. As well as those given by the environment, noted in a binary way: punishment / reward.
You have in this article the key to artificial intelligence ;-)