Group of professionals from different fields meets twice per month in the heart of Seoul to discuss about AI or implement various AI applications.

Martin Kersner
Cinyoung Hur
Adel Sohbi

Call for presenters!

Presentations from 2017


Keras and Tensorboard

Martin Kersner, 2018 Feb 06

Keras, as a high-level neural networks API, tries to simplify procedure of training neural networks by introducing numerous constraints. Recently, I have stumbled upon one of them when I tried to visualize training metrics with TensoBoard. Keras offers TensorBoard callback that is able to save and visualize graph, training and test metrics and activation histograms for different layers in model. [read full post]

Kaggle - How to use pretrained Keras models?

Martin Kersner, 2018 Jan 23

Training deep learning models from scratch is never a good way of aproaching new problem. Moreover, if you want to obtain reasonable baseline results fast. One of the solutions is to use pretrained models, which are available for every major deep learning framework (e.g. Tensorflow, Keras) and finetune them for your particular task or train separate model on outputs from pretrained model. [read full post]

Temperature in softmax layer

Martin Kersner, 2017 Dec 23

In Distilling the Knowledge in a Neural Network paper, there is described how soft targets of the final softmax layer can be modified by additional denominator called temperature. This blog posts aims to visualize effect of temperature magnitude on output of modified softmax layer. [read full post]