Ann vs BNN

Ann vs BNN Mo Huafeng

Without a doubt, Ann simulates BNN. (Some people do not think so. They intend to be independent. Not to mention the copyright issue, it is not necessary to think that you can pass the Lord's family ). These two days of the Olympics, we were excited. We wanted Ann and BNN to compete and see how much difference we were between us and God's old man.

In BNN, neural cells are the foundation. Although neural cells are also called neurons, we have to differentiate them for the convenience of judgment: BNN is called neural cells and Ann is called neurons.
Neural Cells have an important feature, which is charged. (Don't worry, even a crawler won't die ). The neural cell membrane has a potential difference of-70mV. However, this is just a quiet state of neural cells. Under certain circumstances, neural cells will be excited and the membrane potential will suddenly reach + 40mV. It will spread until it passes through the whole cell. In the process of propagation, each place will trigger a reversal of the membrane potential. In some specific parts, many special switches on the cell membrane will emit chemicals that accumulate for a long time to stimulate other nerve cells. One Nerve Cell collects stimuli from other nerve cells whenever possible. When the stimulus accumulation reaches a certain level and exceeds a specific value (threshold), it will become excited and stimulate other nerve cells. (It sounds like sending a gossip :)).
Come back and have a look at Ann. In ANN, neurons correspond to neural cells.
Ann is an abstract model, and neurons naturally do not have cell membranes. However, the membrane potential still exists, but it is expressed by a variable. A neuron collects data from other neurons through a mathematical formula
Is your status value. Of course, not simply adding up a brain, but adding a "weight" value to the state of each neuron. In this way, the state of the neuron itself can be accumulated. Every God
Each element has a threshold value. When the status value exceeds the threshold, it is "activated" and the status is 1; otherwise, it is 0.
In this case, Ann's simulation of BNN is quite close: neural cells have a membrane potential indicating their own State, neurons have State values, neural cell membrane potentials will spread, and the state of neurons can be weighted; neural Cells have syncs and neurons have weights. Neural Cells gather stimulation from other neural cells, and neurons accumulate weights.
It is too early to conclude that Ann is already a complete simulation of BNN. This is only the first section.

BNN
There are also many interesting features at the neural cell level. First, the expression of nerve cell excitement. Although neural cells have a 0-1 feature (Excited-resting), this does not indicate that they cannot express their excited processes.
Degree. This is important in the operation of the nervous system. For example, assume that a group of nerve cells represents the concept of dogs, and the other group represents cats. When you see a dog, the dog's nerve cells are more excited than cats.
Neural cells are more powerful. (Otherwise, how do you know it's a dog ?) But the cat's nerve cells are not excited, because dogs and cats are still somewhat similar. Although cat nerve cells are less excited than dog nerve cells, Ken
It will definitely be more excited than bovine nerve cells. After all, cats are more like dogs than cows.
The most direct way to express the excitement of nerve cells is the membrane potential. That is to say, the 20mV Ratio
10mV is even more excited. Many ANN models also use state values of neurons to indicate the degree of activation of neurons. This is the most intuitive and simple. But actually, neural cells adopt another solution --
Time encoding. In layman's terms, the frequency of distribution of neurons represents the excitement of neural cells. The membrane potential is the same for each excitement, no matter how excited the nerve cells are.
Phase
In contrast, it seems that neural cells are not very clever. If the membrane potential is used to indicate the degree of excitement, it can be much faster than the time encoding method in information processing. There is a reason why neural cells adopt this "Unsatisfactory" SOLUTION
. You must know that the neural system is not made up of circuits and cannot accurately control the potential. An electrochemical reaction is used to maintain the operating of the nervous system and the voltage cannot be precisely controlled as the circuit does. On the other hand, for organisms
In other words, it is much easier for cells to maintain a fixed membrane potential, which is more efficient, energy-saving, environmentally friendly, and green ....
In this case, the old man is not very clever.
The cell design is not ideal. It seems that human design is better. But never think so, if you don't want to be hacked. In fact, the time encoding scheme has many advantages. If Neural Cells
With membrane potential, how can a nerve cell continuously track the status changes of other nerve cells? Two nerve cells communicate with each other through a SYN, while the SYN is stimulated by a neurotransmitter.
. To continue learning about the status of other nerve cells, these cells must continuously distribute the transmitters and indicate the status of the nerve cells by controlling the number of delivery of the transmitters. The consumption of such high-intensity transmitters is
What organisms cannot afford.
In fact, this situation is also a serious problem for Ann. If a neuron tries to continue learning about the exciting processes of other neurons
The weighted summary must be executed endlessly. This is true for all neurons. Generally, computing resources cannot support such operations. Therefore, many ANN models actually give up the interaction between neurons.
Instead, they use snapshots to periodically calculate the state of all neurons. This can still be applied in small-scale Ann systems, but it cannot achieve good results in large systems.
Reverse
If neural cells use the distribution frequency to show their excitement, the problem is simple. Because the frequency encoding also contains changes in the excited state and state of the neural cells over time. Only one neural cell needs
Honestly, we distribute neural pulses according to our own State (which is actually a Summary of the stimulation of other nerve cells. When external stimulus increases, the distribution will naturally become frequent. When a nerve cell keeps receiving
Another high-frequency distribution, then it will know that "that guy is very excited ". This accumulation effect of high-frequency stimulation in time also increases the excitement of the target nerve cells.
In BNN, this dynamic interaction between neural cells plays a vital role. A neural cell can constantly know the excited status and changes of other nerve cells and respond continuously. This feature can help with time-related information. For example, motion detection.
Hour
The implementation of inter-coding also relies on another design of God's family. The membrane potential of nerve cells increases after a single stimulus, then gradually drops back to the resting potential. This "leakage" phenomenon ensures different frequencies
Rate stimulation may have different effects on nerve cells. When the frequency of stimulation exceeds the speed at which the membrane potential drops, the membrane potential continues to rise until the threshold is exceeded, causing nerve cells to be excited.
This
And there is an interesting phenomenon related to this. Usually, when excited nerve cells are distributed to a nerve cell, all the stimuli are combined to make the nerve cell excited. But this overlay and
It does not have to happen at the same time. Sometimes, a series of off-threshold stimuli that are insufficient to trigger nerve cell excitation will overlay the time and eventually lead to nerve cell excitation. This feature plays an important role in the mechanisms related to neural system learning, such as LTP and Ltd.
Most
After that, any neural cell has a so-called "seasonal ". That is, when a nerve cell is released, it will enter a short "sleep period" and will not be affected by other nerve cells until the end of this phase. This
As a result, the maximum excitation frequency of a nerve cell is the reciprocal of the length of the delayed phase. This has practical significance. Each release of nerve cells consumes a variety of nutrients and energy and therefore must be supplemented. Cannot imagine
Neural Cells are excited and can survive. This is equivalent to Ann. Although Ann is composed of numbers or circuits, resources are not infinite. We certainly cannot tolerate individual
Neurons occupy computing resources for a long time, making other neurons unavailable. God's design is always the most streamlined and efficient.

Upper
The halftime is over. In fact, Ann and BNN can be further compared in many aspects. So far, we have only compared the basic features at the cell layer. In general, BNN is more flexible and flexible, and
BNN focuses more on resource utilization. Although it seems that the behavior of neural cells is more complex, it actually uses a simpler and more reliable way to achieve the desired goal. Perhaps this form is not the fastest, or the most
Simple, but it is the best choice under the needs and constraints.
The PK will be extended later, because BNN is obviously above Ann at a higher level, such as neural cell/neuron tissue and functional module composition. We should not doubt the design of the old man's family. Arrogance often won't produce good results. Don't forget, we are designed by God.