"
Training data provides the knowledge and
contextual basis for ML algorithm reasoning.
STMicroelectronics 2022
9
STM32U5 MCU
SERIES
STEVAL-
STWINKT1B
Learn More
Learn More
ML and Data
Whether supervised or unsupervised, training data
provides the knowledge and contextual basis for ML
algorithm reasoning. Data availability and quality are the
foundation of any ML project. Although data collection
and training are not the most complex ML development
step, they are often underestimated. Both directly impact
ML model performance.
ML algorithms are a "good data fit" if they generalize
well from training data to accurately classify and predict
from future data. "Underfitting" means the algorithm or
model do not fit the data accurately enough, oftentimes
because of insufficient training data. "Overfitting" means
the algorithm is too complex and does not generalize the
training data well enough. Underfitting requires additional
model features whereas overfitting requires fewer model
features and balancing of the training data set (Figure 4).
ML at The Edge
Until recently, ML was
restricted to servers and
personal computers with
fast CPU, GPU, and NPU cores. But now ML
is moving onto handheld and embedded
devices at the edge, close to users and
acquired data to reduce latency, increase
efficiency, and ensure data privacy.
This rise of ML on edge devices is being
driven by new methods of algorithm
optimization for resources-constrained
applications, ever more powerful and
efficient MCUs like ST's STM32U5 MCU
series, and easy-to-use developer tools such
as ST's NanoEdge™ AI Studio for machine
learning and STM32Cube.AI toolkit for deep
learning. As a result, Edge-AI is the promising
future for ML everywhere. ■
