Colloqium #13

August 2019


An Introduction, and 1-2 Lectures on neural networks and machine learning technologies and their importance in economy and in general.

We are going to draw on what we have discussed in previous coloquia on AI, deep learning, what algorithms are involved and talk about some of these technologies in more detail.

  1. Introduction by Michael Scheppat

Title: Astonishing phenomena.

Summary: On previous colloquia on AI at PhenCoCo, examples for AI creating ‘art’, like the nvida-alogrithm, an image processing engine, which can transform rudimentary pictures into beautiful landscapes, or the sound aiva engine, creating elaborate synfonies not unlike movie soundtracks.

  1. Joaquin Gomez Prats (electronic engineer)

Title: Machine Learning (ML) fundamentals and types of ML algorithms

Outline Machine Learning (ML):

Introduction, definitions, classic programming vs ML, fundamental concepts, common terms. Types of ML algorithms: criteria categorisation, supervised learning, unsupervised learning, reinforcement learning, classification, regression, common algorithms, use cases, examples. The machine learning process, methods, applied examples, cost function, gradient descent. Fundamentals of Neural Networks.


Machine learning (ML) is the science of getting computers to learn to act, based on data, and without explicitly programming the rules. In the past years, machine learning provided us with a wide range of applications such as effective web search, speech recognition, autonomous driving, advanced image recognition and processing capabilities, and allowed achievements previously thought impossible (Deep Learning, AlphaGo Zero). Most of us use ML several times in our daily lives without even noticing it. The current presentation aims to clarify the fundamental concepts of the matter by categorising, organising and presenting the available information and to provide insight on the fundamental ML processes.


  • gradient descent (finding the right optimum - maxima/minima)

  • underfitting and overfitting: when neural networks get imbalanced, under- and overfitting leads to an ovrerabundance or lack of percision, which distortes the optimal result for the task at hand. Reserving some training data for tests can help to avoid this distortion.

  • Is underfitting and bias the same or do the concepts just overlap? To call underfitting bias, even though it can be parallelized, is not as neutral. It therefore depends on the focus one wants to put in language.

  • example for a popular convolutional neural network in use: CNN Mobile Net V2



Joachím Gomez Prats: Dipl.-Ing. Joaquín Gomez Prats, Key Skills: Data Science, Machine Learning, Elektrotechnik, Prototyping, KI-Beratung, Vorträge. Wichtigste angewandte Tools: Python (numpy, pandas, matplotlib, scikit-learn, Keras, FastAPI, OpenCv), Matlab, Linux, Git, Docker, Electronics, Data Science / Machine Learning Skills und Kenntnisse: Supervised, Unsupervised and Reinforcement Learning, Machine Learning System Design, Neural Networks, Classification: Multiclass, Multilabel, Linear / Logistic Regression, Recommender Systems, Sentiment Analysis, Support Vector Machines, Principal Component Analysis (PCA), Anomaly Detection, K-Means (Clustering), Decision Trees. Random Forests u.v.m; Deep Learning Specialization, Neural Networks and Deep Learning, Hyperparameter tuning, Regularization and Optimization, Machine Learning Projects Structurization, Convolutional Neural Networks, Sequence Models

Colloqium Material