Cezanne Camacho

@cezannecam

Machine and deep learning educator. Former curriculum lead . Electrical Eng. MS . Certainly uncertain. (say-zahn)

Vrijeme pridruživanja: veljača 2018.

Tweetovi

Blokirali ste korisnika/cu @cezannecam

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @cezannecam

  1. Prikvačeni tweet
    17. stu 2018.

    Completed an implementation of a simple Capsule Network! You can take a look at the code, here: . The trained network has some cool traits, and I tried to replicate some experiments they did in the og paper, like feature visualization via reconstructions

    Perturbed, reconstructed images that show how an image of a "6" can be transformed in width, skew, and other localized features.
    Prikaži ovu nit
    Poništi
  2. prije 15 sati

    I have been taking pictures of clouds for a while and I recently gathered a bunch of photos and used them to train a low-res GAN. Here are some cute fake-clouds for

    Poništi
  3. 24. ruj 2019.

    The blog post has some great visualizations such as this one, which shows a network, starting out with randomly-weighted connections then learning to strengthen important connections during training (thanks for the clarification )!

    Prikaži ovu nit
    Poništi
  4. 24. ruj 2019.

    Very cool work on a neural net that can jointly learn the weights and *connections* between nodes in an NN during training! A simple update rule relates the "importance" of a node-node connection (wrt decreasing the training error) to a weight assigned to that connection

    Prikaži ovu nit
    Poništi
  5. 19. kol 2019.

    On auditing as a tool to make AI systems more transparent and accountable to the populations they affect. The brilliant discussed strategies for designing an actionable audit of facial recognition systems, taking inspiration from the social and information sciences

    A slide that describes design considerations for creating an actionable audit. Four considerations are listed: benchmark bias, access to target, [lack of] public pressure, and hostile corporate reaction. These considerations were mitigated by the design of a test dataset for evaluating facial recognition models, and the design of a targeted audit of these models.
    Poništi
  6. 14. kol 2019.

    The is devastating and affecting. I’m linking to the livestream in which historians and writers use objects and the creative imagination to tell the history and legacy of slavery. I’m learning much for the first time and will continue to learn

    Poništi
  7. 13. kol 2019.

    I'd be so curious to see how different video embeddings relate to one another in this embedding space

    Prikaži ovu nit
    Poništi
  8. 13. kol 2019.

    One more detail: the system learns embeddings for each video (in a style similar to learned, word embeddings) based on a number of features including “freshness” or the time a video was uploaded. A sequence of these embeddings = your user watch history

    Prikaži ovu nit
    Poništi
  9. 13. kol 2019.

    After reading about these rankings being biased towards extreme results, I wanted to see how the "watch time" optimization worked. The linked paper is "Deep Neural Networks for YouTube Recommendations" [Covington, Adams, Sargin (2016)]

    Prikaži ovu nit
    Poništi
  10. 13. kol 2019.

    Reading the paper on the YouTube recommendation system; it's made up of 2 neural nets: one for video-candidate generation and one for ranking those videos. The ranking net predicts your "watch time" for each candidate and shows you the highest-ranked vids

    Recommendation system architecture demonstrating the "funnel" where candidate videos are retrieved and ranked before presenting only a few to the user. Caption and image from the paper "Deep Neural Networks for YouTube Recommendations" (Covington et. al., 2016).
    Prikaži ovu nit
    Poništi
  11. 13. kol 2019.

    "YouTube’s recommendation system is engineered to maximize watchtime, among other factors... As the system suggests more provocative videos to keep users watching, it can direct them toward extreme content they might otherwise never find." -

    Prikaži ovu nit
    Poništi
  12. 13. kol 2019.

    On YouTube’s recommendation system & the spread of misinformation and radicalization. Paraphrasing: The emotions that draw people into videos (which YT's system learns to surface and highly-recommend) are often central features of conspiracy theories, and of right-wing extremism.

    Prikaži ovu nit
    Poništi
  13. 19. srp 2019.

    A good lunch-time read about a PizzaGAN! Using a structure similar to a CycleGAN, researchers trained models to add and remove pizza toppings. They show that the PizzaGAN can learn to segment pizza toppings, and remove them (via inpainting)

    Poništi
  14. proslijedio/la je Tweet

    , , , Gary Cottrell, and I have released a full version of our workshop paper on capsule networks and adversarial attack detection! Check it out, or read this thread if you are busy :) 1/7

    Prikaži ovu nit
    Poništi
  15. 16. lip 2019.

    It's in beta, but still very exciting!

    Highlighted function call for the deep learning model ResNet50.
    Prikaži ovu nit
    Poništi
  16. 16. lip 2019.

    A great tool for exploring open source repositories, recently released a "jump to definition" feature which allows you to hover over a function, and jump to its original definition within that same repo!

    Prikaži ovu nit
    Poništi
  17. 14. lip 2019.

    I was thinking that fitting to a dataset average does build in bias. For example: would a general, diagnostic classifier perform better on people of certain genders or age groups? Yes, and it depends on the data source, which cannot be directly investigated if it’s private

    Prikaži ovu nit
    Poništi
  18. 14. lip 2019.

    The goal of techniques like differential privacy is to formalize this idea and develop methods for 1. quantifying information leakage and 2. proving that a model is not leaking too much of one individual's information (so, someone cannot reconstruct an individual’s data)

    Prikaži ovu nit
    Poništi
  19. 14. lip 2019.

    The most helpful patterns for a classifier (that generalizes well) may be thought of as *general truths* that are buried in private, individual data

    Prikaži ovu nit
    Poništi
  20. 14. lip 2019.

    Imagine training a model to do cancer recognition. This ML model aims to learn information that is consistent across individuals and can help identify any new cases of cancer; in fact, we explicitly do *not* want our model to overfit to any one piece of data

    Prikaži ovu nit
    Poništi
  21. 14. lip 2019.

    Specifically, we often want to train an ML model to generalize well to new tasks, which means we want the model to recognize general patterns and not any patterns unique to an individual person (and this personal data should be private!)

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·