2015-12-22 Deep Learning and others readings
I came accross the following article Evaluation of Deep Learning Toolkits which studies a short list of libraries for deep learning: Caffe, CNTK, TensorFlow, Theano, Torch, and various angles: modeling capability, interfaces, model deployment, performance, architecture, ecosystem, cross-platform. It gives a nice overview and helps choosing the library which fits your needs. Once your deep models has been trained, how to use it? This question should be the first one to be answered.
As machine learning and big data become more and more popular, people look for ways to simplify the implementation of complex chains of processings. Python is quite popular so here is one suggestion in that language for deep learning: Blocks and Fuel: Frameworks for deep learning (Bart van Merriënboer, Dzmitry Bahdanau, Vincent Dumoulin, Dmitriy Serdyuk, David Warde-Farley, Jan Chorowski, Yoshua Bengio). It introduces Fuel which models pipelines of data processing.
Finally, a nice tutorial on machine learning with Python: PyData Seattle 2015 Scikit-learn Tutorial. The author's blog is nice too: Pythonic Perambulations. See Out-of-Core Dataframes in Python: Dask and OpenStreetMap. Some modules are hidden in his blog posts such as gatspy which plots timeseries in many ways or supersmoother to smooth timeseries or line_profiler in Optimizing Python in the Real World: NumPy, Numba, and the NUFFT. Two other readings to conclude: Why Python is Slow: Looking Under the Hood and Frequentism and Bayesianism: A Practical Introduction still from the same source.
01/06/2015 Comparative Study of Caffe, Neon, Theano, and Torch for Deep Learning
2015-12-11 Machine learning automatique
Et si plutôt que d'essayer de caler le meilleur modèle sur votre jeu de données, vous esssayiez d'apprendre un modèle qui le fait pour vous... Il existe une conférence pour cela : AutoML workshop @ ICML'15 et un module auto-sklearn.
Et toujours awesome-machine-learning
2015-12-01 Quelques articles de blog, Rodeo, TensorFlow, Tableau, Autoreload, RLPy
Rodeo facilite l'écriture de rapports avec des équations, du code et des graphes. Il est convertit en markdown et PDF : Rodeo 1.1 - Markdown, Autoupdates, Feedback.
Lorsqu'on met à jour un module, les modifications ne sont pas prises en compte automatiquement dans un notebook. Il faut le recharger. Il existe une extension qui fait ça pour vous : Autoreload des modules sous iPython.
Tableau est une application gratuite dans certains cas qui permet de réaliser facilement des dashboards afin de visualiser rapidement des données avec des graphiques animés.
Module pour faire de l'apprentissage par renforcement : RLPy: A Value-Function-Based Reinforcement Learning Framework for Education and Research. Lire également : Batch Learning from Logged Bandit Feedback through Counterfactual Risk Minimization.
Un article sur TensorFlow : What you wanted to know about TensorFlow.
|
Xavier Dupré
|
