According to TechCrunch and Tech Xplore, the chip was developed by a team led by MIT graduate student Avishek Biswas and its greatest strength is its ability to run neural networks on smartphones, home appliances, and other portable devices rather than on high-power servers road.
This means that in the future handsets using the chip can use neural networks for speech and face recognition and native depth learning instead of using a more coarse, rule-based algorithm or sending information to the cloud for analysis and returning results.
Biswas said that in general AI chip designs have a memory and a processor that will move data back and forth between the memory and the processor.Machine learning algorithms require a large number of operations and thus data is transmitted most power back and forth The operation of these algorithms can be simplified to a specific operation called a dot product, which eliminates the need to transfer data back and forth if the dot product operation can be performed directly in the memory.
Neural networks are usually divided into many layers. A single processing node in a network layer usually receives data from several nodes in the lower layer and passes the data to multiple nodes in the upper layer. Each connection between nodes Has its weight.And the process of training neural network is mainly to adjust these weights.
When a node gets the data of multiple nodes in the lower layer, it multiplies each data by its weight and adds these results. This process is called dot product. If the dot product exceeds a certain threshold, the result will be Is sent to the upper node.
In fact, these nodes are only the weights stored in the computer's memory.Calculating the dot product usually involves reading the weight from the memory, getting the relevant data, multiplying the two, and storing the result somewhere and at the node's This is repeated for all input data, and given that there will be thousands or even millions of nodes in a neural network, large amounts of data must be moved in the process.
But this series of operations is to digitize the events in the brain where signals travel along multiple neurons and meet at the synapse.And the firing rate of neurons and the voltage across synapses Of the electrochemical signal corresponds to the data values and weights in the neural network.MIT researchers' new wafers increase neural network efficiency by more faithfully replicating brain activity.
In this chip, the node's input value is converted to voltage, multiplied by the appropriate weight. Only the combined voltage will be converted back to the data, and stored in memory for further processing. Therefore, the prototype chip can simultaneously calculate 16 Dot product of the nodes without having to move the data between the processor and the memory every time the operation is performed.
Dario Gil, vice president of IBM AI, said the results of the study are expected to open up the possibility of using more complex convolutional neural networks for image and video classification in the Internet of Things (IoT).