Intelligence at the Edge
22
Threshold
One of the keys to correctly employ BNNs is properly
setting a threshold. Threshold setting determines what
falls into a 0 or a 1. Take the following example (Table 1).
The figure shows a list of people and their respective ages.
Say the goal is to classify them into one of two categories:
adult or child. The first correct decision that must be made
is what is the transition (threshold) age when one crosses
into adulthood. In this example, the ages of 18 and 21 were
employed. The threshold plays a role in determining what
the results will be. In a manner, this is analogous to the
perceptual illusion from before—same data, different reality.
Table 1: BNN threshold setting example of classifying adulthood
at two distinct ages, 18 and 21.
k-means clustering
Well, the example above is quite simple. The real world
is a lot more complex. How can all this complexity be
handled? In other words, what happens if the data cannot
be so easily categorized into nice groups as shown
above. What if the task is to determine what is real and
significant in an overwhelming sea of data that has few
set parameters from which to categorize?
The solution is done by math. k-means clustering is a
method of vector quantization (sophisticated math). This
process endeavors to group n observations into k clusters
(categories). This is done by putting every observation
into the cluster with the nearest mean (centroid). It then
runs another analysis readjusting the mean in real-time
and resorting any observations into another cluster until
optimization is achieved. It is a form of recursive learning
(Figure 3).
ST Solutions
In a world that requires clarity to get the correct
reality, STMicroelectronics has the solutions.
STMicroelectronics offers the B-U585I-IOT02A Discovery
Kit. It is a complete demonstration and development
platform for the STM32U585AI microcontroller. The
microcontroller features an Arm® Cortex®-M33 core
with Arm TrustZone® and Armv8-M mainline security
extension. It also has 2Mbytes of Flash memory,
786Kbytes of SRAM, and smart peripheral resources.
ST Solutions For AI
