A MICROWAVE oven that uses a neural network to decide when reheated food or
drink is ready will go on sale in Britain next year. It works by detecting the
humidity of the air inside the oven, and deducing from that how 鈥渄one鈥� the
contents are.
The user only has to press a button to indicate whether the contents are
solid or liquid. As the food heats up, the oven continually decides what power
level to use and when to stop.
The oven is the result of three years鈥� research at the European
laboratories of the Japanese electronics company Sharp and the University of
Oxford鈥檚 department of engineering science.
Advertisement
The aim was to build a 鈥渙ne-touch鈥� system that would cook foods according
to their type. But according to Toshio Nomura, research manager at Sharp, the
researchers soon realised that there are too many variables. 鈥淭here are thin
soups, thick lumpy soups, thin meats, vegetables,鈥� he says. 鈥淭here were just
too many types.鈥�
So the team decided to see if the system could be programmed to recognise
the 鈥渄oneness鈥� of food using a sensor system. Given that, they reasoned, the
oven could decide how much longer to heat it.
The researchers started with an microwave that had a built-in sensor to
monitor the oven鈥檚 humidity and temperature, and the food鈥檚 gas production,
height and weight. But that oven sells for around 拢700, and to get the
price down below 拢500 the researchers had to reduce the number of
sensors.
After analysing data from the sensors as the oven reheated hundreds of
samples of food and drink, they found to their surprise that only a humidity
sensor was essential. A neural network compares data from the sensor with data
determined in the laboratory to decide how much more heating is needed. It
needs to be told only whether the food is solid or liquid.
鈥淭he input to the network is the shape of the humidity curve so far,鈥� says
Nomura. 鈥淚t works out the slope of the curve, and when that reaches a maximum,
it can match that against the basic data.鈥� This is especially important with
frozen food, where there may be a rapid increase in the humidity level as the
food thaws, followed by a slow rise.
The neural net鈥檚 output to the oven controls is in the form of a 鈥渄oneness鈥�
percentage, indicating how much more heating is required. 鈥淭his makes it an
unusual neural network, since they are usually used for classifying things
into groups,鈥� says Nomura. 鈥淭hey usually have multiple inputs and outputs.
This has only one of each.鈥�
The system鈥檚 humidity sensor samples the air in the oven every two seconds.
But while the sensor is simple, the oven has an 8-bit processor running at 10
megahertz, making it as powerful as personal computers of a decade ago.
For cooking rather than reheating, users will have to program the oven just
like older models: 鈥淭he humidity characteristic for cooking a meat joint, for
example, is too different [from that of precooked meat] to be incorporated,鈥�
says Clive Bradley of Sharp.
