Researchers from TU/e, led by Yoeri van de Burgt and Marco Fattori, claim to have solved an important problem regarding neuromorphic chips. The new research has been published in Science Advances.
Large-scale neural network models form the basis of many AI-based technologies such as neuromorphic chips, which are inspired by the human brain. Training these networks can be cumbersome, time-consuming and energy-inefficient, as the model is often first trained on a computer and then transferred to the chip. This limits the application and efficiency of neuromorphic chips.
TU/e researchers have solved this problem by developing a neuromorphic device that can perform on-chip training, eliminating the need to transfer trained models to the chip. This could lead to more efficient AI chips in the future.
Neural networks can help solve complex problems with large amounts of data, but as networks grow larger, they come with increasing energy costs and hardware limitations. Neuromorphic chips offer a solution.
Like neural networks, neuromorphic chips are inspired by how the brain works, but the chips take the imitation to a whole new level. When the electrical charge in a neuron changes in the brain, it can send electrical charges to connected neurons. Neuromorphic chips mimic this process.
But there’s a neuromorphic catch—and it has to do with the two ways people train hardware on neuromorphic chips. In the first way, training is done on a computer, and the network’s weights are assigned to the chip’s hardware. The alternative is to do the training in situ or on the hardware, but current memristors have to be programmed one at a time and then checked for errors. This is necessary because most memristors are stochastic, and it’s impossible to update the device without checking it.
While the researchers have shown that the new training approach works, the next logical step is to make the neural networks bigger, bolder, and better.
Photo: Bart van Overbeeke
Source: www.emerce.nl
