ISIC 2018 - Skin Lesion Segmentation using a LinkNet Derived Network

27 July 2018

Introduction The purpose of this document is the description of the architecture used in the 2018 ISIC Challenge for segmenting skin lesions present in images. We use a deep learning based solution for solving the problem. U-Net (Ronneberger, Fischer, and Brox 2015) derived architectures have been proven to be a...

Bellman Equation Derivation

19 July 2018

Using math equations as magic has never been my way. For deep concept understanding, a key point is to first understand all the prior information related with the new concept that you are trying to catch. In Reinforcement Learning, a important equation is the Bellman Equation. In this video Constantin...

Cross Validation alternatives for model selection

10 June 2018

Hold Out method Splitting original training dataset using random sampling without repetition into 2 subsets. The first called training set is used for fitting the model/s, the second one, called valitation set is used for hyper-parameter optimization and for model selection, not only between hyper-parameters but also between different types...

Bagging, Boosting and Decision Trees Based Predictors

04 June 2018

Bagging vs Boosting Error sources are noise, bias and variance. Model ensembles are a very effective way of reducing prediction errors. Bagging and Boosting are two ways of combining classifiers. They are able to convert a weak classifier into a very powerful one, just averaging multiple individual weak predictors. Bagging...

Summary of the most successful Imagenet prediction standardized architectures

20 April 2018

Table below shows a summary of the most successful Imagenet prediction standardized architectures sorted by its performance for the Imagenet classification task. The more successful network is InceptionResNetV2 but at the cost of having more than 55 million of parameters. Xception has a similar accuracy using less than a half...

Covariate shift in Machine Learning inference

09 March 2018

Machine Learning paradigm gives computer systems the ability to learn from data. These algorithms allow the identification of the more relevant features present in data and the way of combining them to maximize the performance in a particular classification or regression problem. Identifying such a set of important features allow...

Testing the stability of a diabetic retinopathy classifier against image disturbances

08 March 2018

In this post we show how a deep learning diabetic retinopathy classifier respond against disturbances in the input image. The classifier showed in the videos is a deep convolutional neural network of 16 layers with human expert performance in the classification of the disease into the 5 standarised classes. In...

Visualizing feature extraction disentangling

20 February 2018

Deep learning convolutional network architectures are used widely for as image classification models. These parametric architectures are composed by a set of layers with differentiated functions. We can consider mainly two differentiated functions: feature extraction layers and classification layers. Feature extraction takes place as a first step, extracting simple localized...

f-GAN a generalization of GANs

20 February 2018

In this blog intro I want to write some notes about the paper f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization and the fantastic presentation of the NIPS 2016 Workshop on Adversarial Training that I had the great honor to attend in person on December 2016. A probabilistic model...

Deep reinforcement learning intro

19 February 2018

In this post I’m going to a make a small intro to the Deep Reinforcement Learning field. It is based on the topics given in the MIT 6.S094: Deep Reinforcement Learning. Reinforcement learning is a general purpose framework for decision making used as part of the design of artificial devices...

Prior & Posterior: two concepts not always well understood

18 February 2018

Intuitions are the best way to get into the concepts, also in math. Prior and Posterior are two commonly used concepts in Bayesian Statistics. For the profan sometimes the specialized jargon can be confusing. This was my case until I found the equation that connected them. The two concepts are...

A deep learning interpretable classifier for diabetic retinopathy disease grading

10 February 2018

Introduction In this blog entry I would like to show you a step by step sample evaluation of the classifier published in arxiv.org paper A Deep Learning Interpretable Classifier for Diabetic Retinopathy Disease Grading and submitted for peer review to a Journal. We show the steps for finding this last...

ELO català vs ELO FIDE

06 February 2018

Motivació Moltes de les fonts d’estudi en el mon dels escacs fan referència a l’ELO FIDE. Una gran majoria de competidors disposen d’una mesura fiable d’ELO català, però no pas d’una mesura fiable d’ELO FIDE, bàsicament degut a que majoritàriament juguen tornejos avaluats mitjançant ELO català. Realitzem un estudi estadístic...

Relationship between Mutual Information and Kullback-Leibler divergence

22 February 2017

Mutual information between two random variables and can be expressed mathematically (by definition) as the Kullback-Leibler divergence between the joint distribution of both variables and the distribution . Mutual information definition is written normally as a function of the entropy but I find more intuitive the first formulation. One can...

How do we define human performance ?

05 December 2016

A key factor for evaluation of machine vs humans performance is the standard used to measure the last one. If we are talking for example about a medical classification task, we can use as a human standard the next different measures: A person (not a doctor) A general doctor A...

On Lasso model regularization

13 October 2015

Lasso regularization is also known as L1-regularization. This regularizer is more aggresive that the more usual L2-regularizer, also known as Rigde regularizer. Both methods are derived from adding to the chosen cost function an additional term, that penalizes the parameter values, forcing them to be as small as possible. In...

On model ensembling

30 September 2015

Model ensembling is a way to improve accuracy predictions using a combination a set of classifiers that are built using different models, data, hyperparameters or parameter initializations. Every difference introduced in the ensembling increases diversity and improves generalization capabilities of the final model. The drawbacks are the loss of interpretability...

Prediction error: bias vs variance

22 September 2015

Sometimes prediction error sources are not well understood. These two important concepts are easily understood after visualizing them: Bias is the difference between the expected average prediction and the real value. Variance measures the variability of the model prediction for a given data point. A low bias - high variance...

On the type of questions that Data Science tries to answer

21 September 2015

Data Scientist use its analysis tools for different purposes. Problems can be classified in different categories according to the result that we want ot achieve: Descriptive: In this case the goal is the description of data. Usually it is not generalizable without further analysis. Exploratory: Here the goal is to...