Temas interesantes a investigar en Machine Learning

Avatar de Usuario
Dalamar
Site Admin
Mensajes: 8927
Registrado: 09 May 2012 01:38

Temas interesantes a investigar en Machine Learning

Mensajepor Dalamar » 05 Jun 2017 16:17

Temas interesantes a investigar en Machine Learning, los ire agregando por prioridad:


Vision: Yolov2 (https://arxiv.org/abs/1612.08242)

Githubs interesantes:


Papers and gihubs: http://www.gitxiv.com/
Papers: http://www.arxiv-sanity.com

Se admiten sugerencias.

Nota: Recomiendo seguir los videos de Siraj Raval en youtube, cortos, al grano y entretenidos.
Adjuntos
1710.09829v1.pdf
(894.86 KiB) Descargado 3 veces
99b7cb0c78706ad8e91c13a2242bb15b7de325ad.pdf
(627.47 KiB) Descargado 6 veces
GAN.pdf
(3.46 MiB) Descargado 15 veces
¿Te ha gustado este hilo? Compartelo en las redes sociales para que se sume mas gente a la conversacion!

girado007
Especulador Listillo
Mensajes: 330
Registrado: 16 Ene 2013 10:40

Re: Temas interesantes a investigar en Machine Learning

Mensajepor girado007 » 05 Jun 2017 19:51

Siraj es un crack! Yo pasé las navidades con sus videos, aunque muy elemental, lo hace todo muy ameno.

Avatar de Usuario
Dalamar
Site Admin
Mensajes: 8927
Registrado: 09 May 2012 01:38

Re: Temas interesantes a investigar en Machine Learning

Mensajepor Dalamar » 06 Jun 2017 09:12

Es todo un actor, se lo curra mucho con los videos y son amenos!

Algun otro que recomiendes?

Lo que hago ultimamente es monitorizar https://arxiv.org/list/stat.ML/recent para ver que nuevos papers hay y si alguno de ellos merece la pena ser investigado e implementado, si te animas a investigar un poco avisa!
¿Te ha gustado este hilo? Compartelo en las redes sociales para que se sume mas gente a la conversacion!

girado007
Especulador Listillo
Mensajes: 330
Registrado: 16 Ene 2013 10:40

Re: Temas interesantes a investigar en Machine Learning

Mensajepor girado007 » 06 Jun 2017 11:25

Tengo a medio leer este: http://www.deeplearningbook.org/ y me parece que está bien.

En videos, me gusta mucho este curso: http://course.fast.ai/lessons/lesson1.html, una máquina tiene que saber distinguir a los mejores amigos del hombre.

Avatar de Usuario
Dalamar
Site Admin
Mensajes: 8927
Registrado: 09 May 2012 01:38

Re: Temas interesantes a investigar en Machine Learning

Mensajepor Dalamar » 07 Jun 2017 17:13

Con respecto a "Deep Learning book", si, yo sigo a Yoshua Bengio en Quora, muy interesante todo lo que dice, por ejemplo:

What are the pros and cons of Generative Adversarial Networks vs Variational Autoencoders?

An advantage for VAEs (Variational AutoEncoders) is that there is a clear and recognized way to evaluate the quality of the model (log-likelihood, either estimated by importance sampling or lower-bounded). Right now it’s not clear how to compare two GANs (Generative Adversarial Networks) or compare a GAN and other generative models except by visualizing samples.

A disadvantage of VAEs is that, because of the injected noise and imperfect reconstruction, and with the standard decoder (with factorized output distribution), the generated samples are much more blurred than those coming from GANs.

The fact that VAEs basically optimize likelihood while GANs optimize something else can be viewed both as an advantage or a disadvantage for either one. Maximizing likelihood yields an estimated density that always bleeds probability mass away from the estimated data manifold. GANs can be happy with a very sharp estimated density function even if it does not perfectly coincide with the data density (i.e. some training examples may come close to the generated images but might still have nearly zero probability under the generator, which would be infinitely bad in terms of likelihood).

GANs tend to be much more finicky to train than VAEs, not to mention that we do not have a clear objective function to optimize, but they tend to yield nicer images.



What would be a practical use case for Generative models?

Because if you are able to generate the data generating distribution, you probably captured the underlying causal factors. Now, in principle, you are in the best possible position to answer any question about that data. That means AI.

But maybe this is too abstract of an explanation. A practical use-case is for simulating possible futures when planning a decision and reasoning. As I wrote earlier, I know what to do to avoid making a car accident even though I never experienced one. I actually have zero training example of that category! Nor anything close to it (thankfully). I am able to do so only because I can generate the sequence of events and their consequence for me, if I chose to do some (fatal) action. Self-driving cars? Robots? Dialogue systems? etc.

Another practical example is structured outputs, where you want to generate Y conditionally on X. If you have good algorithms for generating Y's in the first place, the conditional extension is pretty straightforward. When Y is a very high-dimensional object (image, sentence, data structure, complex set of actions, choice of a combination of drug treatments, etc.), then these techniques can be useful.

We are using images because they are fun and tell us a lot (humans are highly visual animals), which helps to debug and understand the limitations of these algorithms.


Un paper que escribe y recomienda: http://papers.nips.cc/paper/5422-on-the ... tworks.pdf
TODO:
- Explicar que significa Deep Rectifier Networks
- Explicar que son Deep Maxout Networks

Usos de Unsupervised Learning by Yoshua Bengio:

* Predict one variable given the others (pseudolikelihood)
* Predict a subset of variables given the others (generalized pseudolikelihood)
* Predict a variable given the previous ones in some order (fully-visible Bayes nets, autoregressive nets, NADE, generative RNNs)
* Given a corrupted observation, recover the original clean point (denoising)
* Predict whether the input comes from the data generating distribution or some other distribution (as a probabilistic classifier) (Noise-Constrastive Estimation)
* Learn an invertible function such that the transformed distribution is as factorial as possible (NICE, and when considering approximately invertible functions, the variational autoencoders)
* Learn a stochastic transformation so that if we were to apply it many times we would converge to something close to the data generating distribution (Generative Stochastic Networks, generative denoising autoencoders, diffusion inversion = nonequilibrium thermodynamics)
* Learn to generate samples that cannot be distinguished by a classifier from the training samples (GAN = generative adversarial networks)
* Maximize the likelihood of the data under some probabilistic model!
¿Te ha gustado este hilo? Compartelo en las redes sociales para que se sume mas gente a la conversacion!

Avatar de Usuario
Dalamar
Site Admin
Mensajes: 8927
Registrado: 09 May 2012 01:38

Re: Temas interesantes a investigar en Machine Learning

Mensajepor Dalamar » 09 Jun 2017 17:08

girado007 escribió:En videos, me gusta mucho este curso: http://course.fast.ai/lessons/lesson1.html, una máquina tiene que saber distinguir a los mejores amigos del hombre.


http://course.fast.ai acabo de pegarle un vistazo... muuuy buena pinta! Muchas gracias!
¿Te ha gustado este hilo? Compartelo en las redes sociales para que se sume mas gente a la conversacion!

girado007
Especulador Listillo
Mensajes: 330
Registrado: 16 Ene 2013 10:40

Re: Temas interesantes a investigar en Machine Learning

Mensajepor girado007 » 09 Jun 2017 19:12

Que usar dependiendo de que datos sean y que queremos averiguar:

Imagen

Avatar de Usuario
Dalamar
Site Admin
Mensajes: 8927
Registrado: 09 May 2012 01:38

Re: Temas interesantes a investigar en Machine Learning

Mensajepor Dalamar » 10 Jun 2017 06:35

Otro Cheat-sheet.
Adjuntos
MLCheatSheet.jpg
¿Te ha gustado este hilo? Compartelo en las redes sociales para que se sume mas gente a la conversacion!

girado007
Especulador Listillo
Mensajes: 330
Registrado: 16 Ene 2013 10:40

Re: Temas interesantes a investigar en Machine Learning

Mensajepor girado007 » 10 Jun 2017 12:37

Mala suerte :D

Imagen

Avatar de Usuario
Dalamar
Site Admin
Mensajes: 8927
Registrado: 09 May 2012 01:38

Re: Temas interesantes a investigar en Machine Learning

Mensajepor Dalamar » 10 Jun 2017 15:57

Adjuntos
Scikit_Learn_Cheat_Sheet_Python.pdf
(145.68 KiB) Descargado 16 veces
Keras_Cheat_Sheet_Python.pdf
(136.71 KiB) Descargado 12 veces
¿Te ha gustado este hilo? Compartelo en las redes sociales para que se sume mas gente a la conversacion!


Volver a “Inteligencia Artificial”

cron

Ingresar