Fri, 19 May 2017
CNNs are characterized by their use of a group of neurons typically referred to as a filter or kernel. In image recognition, this kernel is repeated over the entire image. In this way, CNNs may achieve the property of translational invariance - once trained to recognize certain things, changing the position of that thing in an image should not disrupt the CNN's ability to recognize it. In this episode, we discuss a few high-level details of this important architecture.
Fri, 12 May 2017
Despite the success of GANs in imaging, one of its major drawbacks is the problem of 'mode collapse,' where the generator learns to produce samples with extremely low variety.
To address this issue, today's guests Arnab Ghosh and Viveka Kulharia proposed two different extensions. The first involves tweaking the generator's objective function with a diversity enforcing term that would assess similarities between the different samples generated by different generators. The second comprises modifying the discriminator objective function, pushing generations corresponding to different generators towards different identifiable modes.
Direct download: mutli-agent-diverse-generative-adversarial-networks.mp3
Category:general -- posted at: 8:00am PDT
Fri, 5 May 2017
GANs are an unsupervised learning method involving two neural networks iteratively competing. The discriminator is a typical learning system. It attempts to develop the ability to recognize members of a certain class, such as all photos which have birds in them. The generator attempts to create false examples which the discriminator incorrectly classifies. In successive training rounds, the networks examine each and play a mini-max game of trying to harm the performance of the other.
In addition to being a useful way of training networks in the absence of a large body of labeled data, there are additional benefits. The discriminator may end up learning more about edge cases than it otherwise would be given typical examples. Also, the generator's false images can be novel and interesting on their own.
The concept was first introduced in the paper Generative Adversarial Networks.
Fri, 28 April 2017
Recently, we've seen opinion polls come under some skepticism. But is that skepticism truly justified? The recent Brexit referendum and US 2016 Presidential Election are examples where some claims the polls "got it wrong". This episode explores this idea.
Fri, 21 April 2017
No reliable, complete database cataloging home sales data at a transaction level is available for the average person to access. To a data scientist interesting in studying this data, our hands are complete tied. Opportunities like testing sociological theories, exploring economic impacts, study market forces, or simply research the value of an investment when buying a home are all blocked by the lack of easy access to this dataset. OpenHouse seeks to correct that by centralizing and standardizing all publicly available home sales transactional data. In this episode, we discuss the achievements of OpenHouse to date, and what plans exist for the future.
Announcements and details
Thanks to our sponsor for this episode Periscope Data. The blog post demoing their maps option is on our blog titled Periscope Data Maps.
To start a free trial of their dashboarding too, visit http://periscopedata.com/skeptics
Kyle recently did a youtube video exploring the Data Skeptic podcast download numbers using Periscope Data. Check it out at https://youtu.be/aglpJrMp0M4.
Fri, 14 April 2017
There's more than one type of computer processor. The central processing unit (CPU) is typically what one means when they say "processor". GPUs were introduced to be highly optimized for doing floating point computations in parallel. These types of operations were very useful for high end video games, but as it turns out, those same processors are extremely useful for machine learning. In this mini-episode we discuss why.
Fri, 7 April 2017
Backpropagation is a common algorithm for training a neural network. It works by computing the gradient of each weight with respect to the overall error, and using stochastic gradient descent to iteratively fine tune the weights of the network. In this episode, we compare this concept to finding a location on a map, marble maze games, and golf.
Fri, 31 March 2017
In this week's episode of Data Skeptic, host Kyle Polich talks with guest Maura Church, Patreon's data science manager. Patreon is a fast-growing crowdfunding platform that allows artists and creators of all kinds build their own subscription content service. The platform allows fans to become patrons of their favorite artists- an idea similar the Renaissance times, when musicians would rely on benefactors to become their patrons so they could make more art. At Patreon, Maura's data science team strives to provide creators with insight, information, and tools, so that creators can focus on what they do best-- making art.
On the show, Maura talks about some of her projects with the data science team at Patreon. Among the several topics discussed during the episode include: optical music recognition (OMR) to translate musical scores to electronic format, network analysis to understand the connection between creators and patrons, growth forecasting and modeling in a new market, and churn modeling to determine predictors of long time support.
A more detailed explanation of Patreon's A/B testing framework can be found here
Other useful links to topics mentioned during the show:
Fri, 24 March 2017
Feed Forward Neural Networks
In a feed forward neural network, neurons cannot form a cycle. In this episode, we explore how such a network would be able to represent three common logical operators: OR, AND, and XOR. The XOR operation is the interesting case.
Below are the truth tables that describe each of these functions.
AND Truth Table
OR Truth Table
XOR Truth Table
The AND and OR functions should seem very intuitive. Exclusive or (XOR) if true if and only if exactly single input is 1. Could a neural network learn these mathematical functions?
Let's consider the perceptron described below. First we see the visual representation, then the Activation function , followed by the formula for calculating the output.
Can this perceptron learn the AND function?
Sure. Let and
What about OR?
Yup. Let and
An infinite number of possible solutions exist, I just picked values that hopefully seem intuitive. This is also a good example of why the bias term is important. Without it, the AND function could not be represented.
How about XOR?
No. It is not possible to represent XOR with a single layer. It requires two layers. The image below shows how it could be done with two laters.
In the above example, the weights computed for the middle hidden node capture the essence of why this works. This node activates when recieving two positive inputs, thus contributing a heavy penalty to be summed by the output node. If a single input is 1, this node will not activate.
Universal approximation theorem tells us that any continuous function can be tightly approximated using a neural network with only a single hidden layer and a finite number of neurons. With this in mind, a feed forward neural network should be adaquet for any applications. However, in practice, other network architectures and the allowance of more hidden layers are empirically motivated.
Other types neural networks have less strict structal definitions. The various ways one might relax this constraint generate other classes of neural networks that often have interesting properties. We'll get into some of these in future mini-episodes.
Check out our recent blog post on how we're using Periscope Data cohort charts.
Thanks to Periscope Data for sponsoring this episode. More about them at periscopedata.com/skeptics
Fri, 17 March 2017
In this Data Skeptic episode, Kyle is joined by guest Ruggiero Cavallo to discuss his latest efforts to mitigate the problems presented in this new world of online advertising. Working with his collaborators, Ruggiero reconsiders the search ad allocation and pricing problems from the ground up and redesigns a search ad selling system. He discusses a mechanism that optimizes an entire page of ads globally based on efficiency-maximizing search allocation and a novel technical approach to computing prices.