Is AI == Brain?

Research has shown that when our mind learns some new things, it changes its structure, and those progressions happen at those connections of synapses of neurons, which is truly stunning. So in the event that we take Neural Network in Deep Learning as it's simple, one can envision that after each propagation of training, on the off chance that it changes its connection of neurons it can have a greater capacity of learning as a human brain and assuming those chemical signals passing through the neurons of the brain as weights evaluated at each neuron of the layer. So in this article, we will examine on this thought and what components and method makes neural network to have the ability to work like a human brain.

So start with the neuron, first in the neural network we can, for the most part, say it as a combination of a linear equation and activation function, it yields some weight (probability ratio) and give to the neuron of the next layer. So after each forward propagation, it looks at for loss and back propagates to update the weights, and this process continues to minimize the loss and those connections will remain intact, just weights are getting updated. Presently in any event that we change those connections and attempt to think of an alternate kind of network which really reduces that learning(training) time gives us lead to having an approach of any kind of algorithm letting neural network work as same as of human brain. As per biological perspective, the synapses of neurons change the connection, so changes ought to be done to node simply in the neural network, in the event that we are discussing neurons connections, one ought to know about "drop-out layers".

There’s a term in deep learning ‘overfitting’ and one have to prevent it, so it actually done by dropping out several random neurons in the neural network at every training step, now that dropped out neuron is no longer permitted to give or receive any information from the other corresponding neurons, one can assume that the neural network is working as if that neuron didn’t exist after dropping it. The decision that which neuron will drop out at each training step is solely random. A probability factor is set to have the specific ratio of dropout at a particular training iteration. It actually allows one to train the model in an ensemble fashion, so a single neural network can have a better performance instead of training multiple because each training step results in different subsets of neural network. It adds on the performance on unseen data as well but training time per epoch gets increased and computational intensity also gets increased. So here we are not satisfied by modifying the learning algorithm using drop-out but there are other such techniques that can have a better hand in dealing with the training time and computation power.

Extensively talking artificial neural network is like a trained pattern recognition system, simply input something (numerical values) at one end and outputs at another, it’s an interpretable fashion while a human brain is an example of intelligence, where we thought different mechanisms to interact with the environment. As indicated by the figures, there are around 14 to 16 billion neurons in the cerebral cortex (back of the brain), so attempting to have the equivalent no. of neurons in a neural network is a very complex task and computation rich, using some regularization methods can be an option but not a pure solution of the idea.

The information furnished above is true to author's knowledge but do not consider it as a professional advice.