Ishwar Purushotham (Data scientist) et Laurie Riguccini (Data Scientist & Astrophysicienne) chez Soladis Digital,
“In very simple terms, Machine Learning is a collection of techniques that capitalizes on pattern detection to transform your data into actionable insights.
There are four main techniques of Machine Learning namely supervised, unsupervised, semi-supervised and reinforcement Learning. Two techniques of Machine learning are presented above to prove the emergence of Deep Learning as a robust method.
In supervised Machine Learning, a classification or a regression model is trained on a training set called “gold standard training set” which is then used to predict or make a prediction on non-training data. Good models are able to maximize the limited available training samples and generalize the predictions on unseen data sets. Some of the better models are able to adapt to the subtle evolution of the underlying data patterns over time and can learn the underlying data distribution from one problem and adapt it to another different but related problem.
At Soladis, supervised Machine Learning is used to solve a number of tasks for our clients. Some of them may include missing data imputation and also making models for the detection of diseases and anomalies. We provide our client some consulting on those topics or perform the analysis on their request.
In contrast with unsupervised Machine Learning the goal is to find hidden properties in data sets without explicitly being given gold standard training data set. One of the main technique used at Soladis for unsupervised learning is called clustering similar data is grouped together according to their hidden characteristics. Solutions offered by Soladis frequently take advantage of such methods.
Deep learning is a specialized branch of Machine Learning where any of the four domains mentioned above can used. Very briefly it consists of an Artificial Neural Network comprising of thousands, if not millions, of layers and each layer consisting of thousands of neurons where each neuron represents an electronic weight. As information passes from the input through the layers, the data is transformed into abstract feature representation and the data only relevant features are transmitted from one layer to the next until finally the prediction of the problem is the output.
If you think that data is becoming more and more available nowadays and also the fact that the cost of computation is decreasing, then you realize that deep learning is actually becoming more and more efficient in predicting trends in your data set. At Soladis, we have understood the power of Deep Learning and we are using such techniques to serve our clients interests with a full range of expertise.”