41 Matching Annotations
  1. Jun 2017
    1. his single model is trained concurrently onImageNet, multiple translation tasks, image captioning (COCO dataset), a speechrecognition corpus, and an English parsing task. Our model architecture incor-porates building blocks from multiple domains. It contains convolutional layers,an attention mechanism, and sparsely-gated layers

      Massively multitask learning

  2. May 2017
    1. Notice that with a few exceptions, the non-tank areas are largely dark

      Yes, but the brighted point is actually in the sky - seems strange to a reader

    2. like partial occlusion and saliency maps exist to provide a sanity check on the learning process.

      You've already linked these. Don't need to do it again

    3. al occlusion Occlude parts of the image and see how the classification changes [1].Saliency map Compute derivatives of class predictions with respect to the input image [2].Class Prediction Not a visualization per se, but can be a handy, simple check on the learning process.

      Can you include visuals directly here?

    4. At Merantix,

      Your homepage takes too long to load (image took a few seconds)

    1. The Talking Machines Podcast by Katherine Gorman and Ryan Adams The Spectator by Shakir Mohamed Andrew Gelman’s Blog Christian Robert’s Blog OpenAI Blog Andrej Karpathy’s Blog Dustin Tran’s Blog Alex Smola’s Blog Radford Neal’s Blog Hal Daume III’s NLP Blog inFERENCe

      Solid set of machine learning blogs

  3. Apr 2017
  4. terralang.org terralang.org
    1. Terra is a low-level system programming language that is embedded in and meta-programmed by the Lua programming language
    1. Persistent homology is a method for computing topological features of a space at different spatial resolutions. More persistent features are detected over a wide range of length and are deemed more likely to represent true features of the underlying space, rather than artifacts of sampling, noise, or particular choice of parameters.

      Sounds like feature extraction to me. Wonder if it is relevan to NNs

      Clayton Mcdonald pointed me to this

    1. var myChart = chart().width(720).height(80); Modifying an existing chart is similarly easy: myChart.height(500); As is inspecting it: myChart.height(); // 500 Internally, the chart implementation becomes slightly more complex to support getter-setter methods, but convenience for the user merits additional developer effort! (And besides, this pattern becomes natural after you’ve used it for a while.) function chart() { var width = 720, // default width height = 80; // default height function my() { // generate chart here, using `width` and `height` } my.width = function(value) { if (!arguments.length) return width; width = value; return my; }; my.height = function(value) { if (!arguments.length) return height; height = value; return my; }; return my; } To sum up: implement charts as closures with getter-setter methods.
    1. Good page for quick overview of es6 that's relevant to d3

  5. pdfs.semanticscholar.org pdfs.semanticscholar.org
    1. AI and HCI: Two Fields Divided by a Common Focus

      Article on relationship between HCI & AI

    1. A few equations scattered on a single page of paper explain most of what goes on in the physical world. This miracle speaks of the organizing principles of the universe: symmetry, invariance, and regularity—precisely the stuff on which mathematics feasts.
    2. Natural Algorithms and Influence Systems By Bernard Chazelle
    1. (One promising area of current research is Generative Adversarial Networks (GANs), which seem to be well-suited to generating realistic data.)

      Related to Pixel Domain Adaptation work: https://arxiv.org/abs/1612.05424

    2. We put this orientation predictor into our pipeline, using its detected orientation to rotate the image to upright before doing word detection and OCRing.

      I wonder how hard it would be to replace this with a small spatial transformer network

    3. , being very careful about their privacy,

      They say then whenever user data is mentioned in any way whatsoever

    4. Our jail infrastructure allows us to efficiently set up expensive resources a single time at startup, such as loading our trained models, then have these resources be cloned into a jail to satisfy a single OCR request. The resources are cloned Copy-on-Write into the forked jail and are read-only for how we use our models so it’s quite efficient and fast.
    5. We generated about a million synthetic words, trained our deep net, and then tested our accuracy, which was around 79%.

      I expected this to be higher. 1 million is a lot of images in data land, although I'm not sure how much adding more data would help (considering they can generate unlimited samples)

      Was the test set real images?

    1. Three years experience building machine learning or AI systems

      Wonder why they chose 3

    1. opml2json A simple tool to convert opml files exported by Mindnode Pro to JSON consumable by D3 Javascript library.
  6. Mar 2017
    1. "augmentation of human intellect" via an interactive vehicle navigating through "thought vectors in concept space."

      Haven't seen many good interfaces for navigating through e.g. word vectors

    2. He actually traveled with his own 16mm projector with a remote control for starting and stopping it to show what was going on (people were not used to seeing and following cursors back then).
    3. needs for large scope, reduction in complexity, and end-user literacy would require that data and control structures be done away with in favor of a more biological scheme of protected universal cells interacting only through messages that could mimic any desired behavior.


    4. Bruner
    1. Architecture guidelines for stable Deep Convolutional GANsReplace any pooling layers with strided convolutions (discriminator) and fractional-stridedconvolutions (generator).Use batchnorm in both the generator and the discriminator.Remove fully connected hidden layers for deeper architectures.Use ReLU activation in generator for all layers except for the output, which uses Tanh.Use LeakyReLU activation in the discriminator for all layers

      Concrete guidelines for DCGAN architecture

  7. Feb 2017
  8. Jan 2017
    1. The “can you hear me” con is actually a variation on earlier scams aimed at getting the victim to say the word “yes” in a phone conversation. That affirmative response is recorded by the fraudster and used to authorize unwanted charges on a phone or utility bill or on a purloined credit card.
    1. Adversarial Variational Bayes: Unifying Variational Autoencoders andGenerative Adversarial Networks