Machine Learning
Machine learning (ML) is a branch of computer science that enables computers to learn from data.
The learning process involves the preparation of data sets, the selection of appropriate models
and learning algorithms, training the model with the respective data sets through the learning algorithm,
and the validation of the model. The produced model , which can be regarded as an approximation model,
is then used to make data-driven decisions on unseen novel data.



There are various classes of ML algorithms that are categorized according to the way they work.
Examples of the ML algorithms adopted widely in applications are: clustering, decision trees,
and artificial neural networks. Many effective techniques have also been studied to construct
the models that are more reliable and more accurate. The tools include regularization and ensemble algorithms.



ML algorithms have been successfully applied to a variety of problems: object detection/recognition,
speech recognition, image retrieval, medical diagnosis, and bioinformatics.
Based on the class of a problem, e.g., classification or regression, an appropriate ML algorithm is
chosen and used to build an approximation model.
ML algorithms can provide us with a good model when problems have enough data. The more data, the better performance.
Deep Learning
Deep learning is a subfield of machine learning concerned with the artificial neural networks that are
inspired by the function and structure of the brain. Deep networks have attracted much attention from many
application domains since the remarkable improvements in image classification and speech recognition,
leading to the recent revival of artificial neural networks.



Deep networks are deeper, larger, and more complex artificial neural networks than traditional ones.
Deep networks usually consist of more than twenty layers while traditional ones have just two or three layers.
The introduction of GPUs (Graphic Processing Units) made it possible to accelerate the learning process of deep
networks, that otherwise would take a huge amount of time due to the large number of parameters in deep networks,
dramatically.



The most popular deep networks are Deep Belief Networks (DBN), Convolutional Neural Networks (CNN),
and autoencoders. CNNs have phenomenal performance in computer vision tasks such as image classification and
recognition. The architecture of CNNs, which models the human visual system, allows us to almost eliminate
the feature engineering that are crucial and inevitable to the performance in computer vision tasks.
CNNs become the de facto standard in image recognition and understanding. DBNs are graphical models that are
capable of learning reconstruction of its inputs probabilistically, recently adopted in drug discovery because
they allow efficient search and optimization in the space of chemical compounds.
Pattern recognition
Pattern recognition is a branch of machine learning that focuses on detecting and classifying
patterns and ordered structures in input data. Pattern recognition is linked closely to machine
learning but pattern recognition is concerned with finding, extracting, and visualizing patterns
from input data while the focus lies on building models from data by learning algorithms in machine learning.



A major application of pattern recognition is classification, which attempts to assign each input
to one of a given set of classes. In general, pattern recognition systems consist of the components
of pre-processing, feature extraction, classification, and post-processing.
The pre-processor performs operations on input sensor data such as noise reduction, filtering, and
segmentation to isolate the pattern from background and to enhance it. In feature extraction,
problem-specific processing is applied to extract salient features that are useful for classification.
Efforts are taken in the feature extraction step as the recognition performance relies on whether or not the
features are effective. That is why the step is called “feature engineering”. The problem-specific processing
requires dimensional reduction, transformations, and spatiotemporal computation. The classifier,
constructed based on machine learning algorithms, then uses the extracted features to assign the input to a category.



Address : Rm 113, SNU Research Park Main Center, 1 Gwanak-ro, Gwanak-gu, Seoul, Korea, 08826
Tel : +82-70-7701-0064 | info@medysapiens.com
COPYRIGHT 2017 MEDYSAPIENS ALL RESERVED