The team found that when the cartoon was set up to work on the edge of the throw mode, the use of gate voltage was possible to control the charging bloodup in the silicon, either closing the device or enabled the spiral of the neuron duplicate activity. Adjustment in this voltage can allow different frequencies of spiking. These adjustments can also be made using spikes, which allows spiking activity to adjust the weight of different inputs.
With the basic concept, the team found out how to run hardware in two ways. In one of them, it acts like an artificial synapse, which is capable of adding any six (and potentially high) weight, which means that the strength of synthetic neurons in the front layer of the nerve network. This weight is an important feature of nerve networks such as large language models.
But when it helps to modify its behavior, together with other transistor, it was possible to act like a neuron, which affects the inputs in a way that affects the frequency of spikes that send it to other artificial neurons. Spiking frequency can be a result of a thousand element in the intensity. And this behavior was stable for more than 10 million watch cycles.
All this, developed with the CMOS process, is easily required by standard transistors, so this is something that can potentially be put in place very fast.
The profession and in agreement
So what are the benefits of it? It only requires two transactions, meaning that many of these devices are possible to keep the same chip. Researchers say “from the synaptic point of view,” in principle, in principle, can replace static access memory in nerve networks of bicarized weight (a fluctuating memory cell consisting of at least six transistors), or a multi -level synaptic rows, which has a significant reduction in the synaptic rows. “