Building Convolutional Neural Networks with Tensorflow

Building Convolutional Neural Networks with #Tensorflow #abdsc

  • In the past year I have also worked with Deep Learning techniques, and I would like to share with you how to make and train a Convolutional Neural Network from scratch, using tensorflow.
  • Later on we can use this knowledge as a building block to make interesting Deep Learning applications.
  • Source code is also provided.
  • Before you continue, make sure you understand how a convolutional neural network works.
  • The code is also available in my GitHub repository, so feel free to use it on your own dataset(s).

In the past year I have also worked with Deep Learning techniques, and I would like to share with you how to make and train a Convolutional Neural Network from…
Continue reading “Building Convolutional Neural Networks with Tensorflow”

Artificial intelligence now powers all of Facebook’s translation

Artificial intelligence now powers all of Facebook’s translation

  • On Thursday, Facebook announced that all of its user translation services—those little magic tricks that happen when you click “see translation” beneath a post or comment—are now powered by neural networks, which are a form of artificial intelligence.
  • Back in May, the company’s artificial intelligence division, called Facebook AI Research, announced that they had developed a kind of neural network called a CNN (that stands for convolutional neural network, not the news organization where Wolf Blitzer works) that was a fast, accurate translator.
  • Now, Facebook says that they have incorporated that CNN tech into their translation system, as well as another type of neural network, called an RNN (the R is for recurrent).
  • Facebook says that the new AI-powered translation is 11 percent more accurate than the old-school approach, which is what they call a “phrase-based machine translation” technique that wasn’t powered by neural networks.
  • As an example of the difference between the two translation systems, Facebook demonstrated how the old approach would have translated a sentence from Turkish into English, and then showed how the new AI-powered system would do it.

On Thursday, Facebook announced that all of its user translation services—those little magic tricks that happen when you click “see translation” beneath a post or comment—are now powered by neural networks, which are a form of artificial intelligence.
Continue reading “Artificial intelligence now powers all of Facebook’s translation”

Receptive Field Calculator

We built a tool for calculating the receptive field of convolutional filters:  #deeplearning

  • A convolutional layer operates over a local region of the input to that layer with the size of this local region usually specified directly.
  • You can also compute the effective receptive field of a convolutional layer which is the size of the input region to the network that contributes to a layers’ activations.
  • For example, if the first convolutional layer has a receptive field of 3×3 then it’s effective receptive field is also 3×3 since it operates directly on the input.
  • However if the second layer of a convolutional network also has a 3×3 filter, then it’s (local) receptive field is 3×3, but it’s effective receptive field is 5×5.

A convolutional layer operates over a local region of the input to that layer with the size of this local region usually specified directly. You can also compute the effective receptive field of a convolutional layer which is the size of the input region to the network that contributes to a layers’ activations. For example, if the first convolutional layer has a receptive field of 3×3 then it’s effective receptive field is also 3×3 since it operates directly on the input.
Continue reading “Receptive Field Calculator”

Google is using machine learning to sort good apps from bad on the Play Store

Google is using machine learning to sort good apps from bad on the Play Store

  • Google is using machine learning to group apps by function and spot the bad apples.
  • Image: Google

    With machine learning, Google can use peer grouping to scan apps that are being loaded on to the Play Store en masse.

  • “We focus on signals that can negatively affect user privacy, such as permission requests that are not related to core app functionality, and the actual, observed behaviors,” explains Martin Pelikan of Google’s security and privacy team over email.
  • In its most recent annual Android security review, the percentage of users who had installed harmful apps from the official Play Store fell from 0.15 percent in 2015 to 0.05 percent in 2016.
  • Many users — particularly those in China — install Android apps from alternative app stores, which the company doesn’t have control over.

Security on Android has always been a challenge for Google due to the operating system’s open nature. But in recent years, the company has been gaining ground in its fight against malware and…
Continue reading “Google is using machine learning to sort good apps from bad on the Play Store”

Learning Deep Learning with Keras

Learning #DeepLearning with #Keras  #NeuralNetworks @pmigdal

  • For that reason, I suggest starting with image recognition tasks in Keras, a popular neural network library in Python.
  • Deep learning is a name for machine learning techniques using many-layered artificial neural networks.
  • See a plot of AUC score for logistic regression, random forest and deep learning on Higgs dataset (data points are in millions):

    In general there is no guarantee that, even with a lot of data, deep learning does better than other techniques, for example tree-based such as random forest or boosted trees.

  • Deep learning (that is – neural networks with many layers) uses mostly very simple mathematical operations – just many of them.
  • Its mathematics is simple to the point that a convolutional neural network for digit recognition can be implemented in a spreadsheet (with no macros), see: Deep Spreadsheets with ExcelNet.

I teach deep learning both for a living (as the main deepsense.io instructor, in a Kaggle-winning team1) and as a part of my volunteering with the Polish Chi…
Continue reading “Learning Deep Learning with Keras”

GitHub

Our implementation of graph auto-encoders (in TensorFlow) is now available on GitHub:

  • This is a TensorFlow implementation of the (Variational) Graph Auto-Encoder model as described in our paper:

    Graph Auto-Encoders (GAEs) are end-to-end trainable neural network models for unsupervised learning, clustering and link prediction on graphs.

  • GAEs have successfully been used for:

    GAEs are based on Graph Convolutional Networks (GCNs), a recent class of models for end-to-end (semi-)supervised learning on graphs:

    A high-level introduction is given in our blog post:

    In order to use your own data, you have to provide

    Have a look at the function in for an example.

  • In this example, we load citation network data (Cora, Citeseer or Pubmed).
  • The original datasets can be found here: and here (in a different format): can specify a dataset as follows:

    You can choose between the following models:

    Please cite our paper if you use this code in your own work:

gae – Implementation of Graph Auto-Encoders in TensorFlow
Continue reading “GitHub”

Microsoft is using 400 million PCs to build antivirus protection

Microsoft is using 400 million PCs to build antivirus protection

  • To prevent the next global malware crisis, an upcoming update will rely on machine learning from more than 400 million computers running Windows 10, Microsoft said Tuesday.
  • In its Fall Creators Update, Microsoft will use a wide range of data coming from its cloud programs such as Azure, Endpoint and Office to create an artificial intelligence antivirus that can pick up on malware behavior, said Rob Lefferts, director of program management for Windows Enterprise and Security.
  • If new malware is detected on any computer running Windows 10 in the world, Microsoft said it will be able to develop a signature for it and protect all the other users worldwide.
  • Microsoft sees artificial intelligence as the next solution for security as attacks get more sophisticated.
  • Using cloud data from Microsoft Office to develop malware signatures is crucial, for example, because recent attacks relied on Word vulnerabilities.

An upcoming security update will incorporate machine learning from millions of computers fending off malware, the company says.
Continue reading “Microsoft is using 400 million PCs to build antivirus protection”