July 8 (Fri)
|
09:00-10:00 |
Breakfast |
10:00-10:30 |
Invited talk (abstract)
H. Yin - SOM’s Relationship with Manifold and Deep Features
Recent decades have witnessed a much-increased demand for effective and efficient methods for dealing with, analysing and understanding fast growing amount of data of increasing complexity, dimensionality and volume. Whether it is in biology, social sciences, engineering, robotics or computer vision, data is being sampled, collected and cumulated in an unprecedented scale. A systematic and automated way of utilizing data and representing it efficiently has become a great challenge facing all these fields. Research in this emerging area has flourished and deep networks have emerged to the frontline. Deep learning has become the main methodology for many vision, machine learning and data analytics tasks, with abundant deep networks developed. However, it's largely unclear how they come to the decision. Making sense of deep learning with manifolds and graphs can help elucidate the underlying relationships that the network is uncovering and reveal possible shortfalls or holes and unstable areas that may cause the network to fail. SOM has long been associated with manifold. Manifold concept plays important role in the representations in deep learning, not only because of the manifold hypothesis, but also its practicality as the basis for machine learning, task representation and recognition. How well these feature maps can capture the intrinsic properties of data will determine the performance of the deep network; and crude, enduring training may not always guarantee good representation. Instead, feature maps especially organised feature maps can help explain the output of the networks, where SOM can play a useful role in elucidating relationship among features. For instance, filters in the layers of convolutional neural network can be seen as basis of multiple manifolds or feature maps and can be further associated by SOM or graphs for probing robustness and sensitivities of the networks. Better representation can lead to better representation and better learning performance. Examples and case studies will be used to illustrate that manifold concept provides an underlying framework to the study of wide range data-driven methods and learning techniques.
|
10:30-11:00 |
Invited talk (abstract)
M. Olteanu -
R-SOMbrero for SOM visualisation
Over the last forty years, self-organising maps have been proven as an insightful method for clustering, mapping, and visualising complex data.
Numerous softwares are available, and implement the original SOM algorithm or different versions of it.
The aim of this talk is to illustrate a specific tool, called SOMbrero, and available as an R package on the CRAN.
SOMbrero builds on the powerful R visualisation environment (ggplot, shiny, …) and is able to handle different structures of data: numerical, categorical, relational.
It also handles missing data and performs missing data imputation.
We will use several real-life datasets to illustrate the functionalities of the package, and its user interface.
|
11:00-11:30 |
Session 4
11:00-11:15 |
#5734 -
SOM Visualization Framework in Python, including SOMStreamVis, a Time Series Visualization
-
Sergei Mnishko and Andreas Rauber
|
11:15-11:30 |
#4092 -
Application of Kohonen Maps in Predicting and Characterizing VAT Fraud in Southern Mozambique
-
Ricardo Santos, Ricardo Moura and Victor Lobo
|
|
11:30-13:15 |
Lunch break (lunch provided) |
13:15-13:45 |
Transfer to Smíchovské nádraží |
13:45-14:58 |
Travels to Pilsen |
15:30-21:30 |
Guided tour Pilsner Urquell Brewery (with refreshments and ending with dinner) |
21:30-23:30 |
Travel to Prague (Expected ending at Resslova street (nearby the WSOM+22 venue) |