Transcript
Most recent advances in artificial-intelligence systems have
come courtesy of neural networks. These are densely interconnected meshes of
simple information processors that learn to perform tasks by analyzing huge
sets of training data.
Until now, neural nets have been large, and their computations
have been energy intensive. Therefore, so they're not very practical for
handheld devices. Most smartphone apps that rely on neural nets simply upload
data to internet servers, which process it and send the results back to the
phone.
But, according to new research presented at the International Solid-State Circuits Conference, MIT
researchers have developed a special-purpose chip that increases the speed of
neural-network computations by three to seven times, while reducing power
consumption 93 to 96 percent. That could make it practical to run neural
networks locally on smartphones or even to embed them in household appliances.
Neural networks are typically arranged into layers. A single
processing node in one layer of the network will generally receive data from
several nodes in the layer below and pass data to several nodes in the layer
above. Each connection between nodes has its own "weight," which indicates how
large a role the output of one node will play in the computation performed by
the next.
Training the network is a matter of setting those weights.
The MIT researchers' new chip improves efficiency by replicating
the brain more faithfully than prior designs. In the chip, a node's input
values are converted into electrical voltages and then multiplied by the
appropriate weights. Only the combined voltages are converted back into a
digital representation and stored for further processing.
The chip can thus calculate dot products for multiple nodes-6 at
a time, in the prototype in a single step, instead of shuttling between a
processor and memory for every computation.
One of the keys to the system is that all the weights are either
1 or -1. That means that they can be implemented within the memory itself as
simple switches that either close a circuit or leave it open. Recent
theoretical work suggests that neural nets trained with only two weights should
lose little accuracy-somewhere between 1 and 2 percent.
In experiments, the MIT researchers ran the full implementation
of a neural network on a conventional computer and the binary-weight equivalent
on their chip. Their chip's results were generally within 2 to 3 percent of the
conventional networks.
|