Machine learning has come a long way, but it is still very different from how our brains work. AI has changed into Artificial General Intelligence (AGI), but Machine Learning still has a long way to go before it can match the human brain.
The only thing they have in common is that they are both made of cells called neurons linked together by connections called synapses. ML algorithms can’t work in biological neurons because the signals are different, and the time scale is different, among other things.
Unlike an artificial neural network, biological neurons are slow but work differently. Yes, there are a lot of nerve cells that connect and work in a manner that is comparable to how an artificial neural network would work.
But it’s clear that seeing and hearing happen one after the other, and the slowness of neurons limits the number of steps employed in processing. If you can see or hear something like that and react to it in a split second, the number of brain layers that can process data is limited by how quickly each layer can perform its function.
The slow processing speed makes it unlikely that great training examples like those used in Computer Vision would work in a natural setting. Layers in artificial neural networks must set up a certain way to work. If messages from later layers could go back to previous layers, the basic backpropagation won’t work because the gradient descent ground of the network is no longer constant.
Looping connections give a neural network some internal memory, so it can create different outputs from the same input data depending on how it is set up on the inside. This is great for the idea of making machines think more like humans but not ideally great for a graze artificial neural network.
What does a perceptron mean?
The perceptron’s method is different from what we know about real neurons and doesn’t make sense. Most of the time, neurons can’t do the most basic addition that perceptrons can do. After each spike, a neuron has a refractory period that lasts about 3 ms. This causes neurons to miss inbound spikes, which leads to mistaken summations.
The sum of 6 and 1 is always 6, not 7. Total sum only works whenever the network is too slow, and there aren’t sufficient values to make it useful. In general, it doesn’t work to think that the value of a perceptron is the number of times a biological nervous spikes.
Machine learning relies on accurate neuron values and synapse weights. However, neither of these things makes more sense in a natural setting. The slower each network layer runs, the more a neuron’s value will represented. A neuron could stand for about 10 different values instead of the floating-point numbers that most artificial neural networks use. Setting precise synaptic weights is even worse.
What are synapses in the field of machine learning?
Some theoretical methods for determining synapse weights may be useful. However, the available biological data indicate that chance largely determines synapse weights. Most synapses are digital, with a mass of either 0 or 1, and any values only demonstrate how sure the value is or how simple it is to remember or ignore that particular piece of data.
How is machine learning different from the way your brain works?
The bp algorithm needs to set specific synaptic connections to precise weights is the single most problematic aspect of the idea that Machine Learning is analogous to the brain. There is not a single biological pathway that we are aware of that may result in this. The synaptic weights shift whenever the neurons connected by those synapses fire nearly simultaneously.
However, the infrastructure necessary to build a particular synapse will require the cooperation of numerous neurons, making it inefficient to store the information in synapse weights. The neurons in the human brain only activate on average once every two seconds, which means that the human brain does not require much power.
Because of its modest firing rates, your brain cannot constantly imitate an artificial neural network designed to trigger activity. Even while the field of neuromorphic engineering has made significant headway in the right direction in recent years, most of its success still comes from training algorithms and accurately determining synaptic weights, which are entirely reasonable practices.
One Response