Dan Hendrycks

@DanHendrycks

UC Berkeley PhD student. Focusing on Trustworthy Machine Learning via value modeling, robustness, and uncertainty. Maximizing utility.

Berkeley, CA • Marshfield, MO
people.eecs.berkeley.edu/~hendrycks/
Vrijeme pridruživanja: kolovoz 2009.
29 fotografija ili videozapisa Fotografije i videozapisi

Tweetovi

Blokirali ste korisnika/cu @DanHendrycks

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

  1. proslijedio/la je Tweet
    5. pro 2019.

    Excited to announce our new paper "AugMix", which proposes a simple yet surprisingly effective method to improve robustness & uncertainty particularly under dataset shift :)  Joint work with . More details below:

    74 proslijeđena tweeta 240 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  2. 26. stu 2019.

    ML systems that can _act cautiously_ will require detailed uncertainty estimates. An important piece of uncertainty information is an anomaly's location. Consequently we created a benchmark for locating anomalies. Paper: Dataset:

    1 proslijeđeni tweet 15 korisnika označava da im se sviđa
    Poništi
  3. 9. lis 2019.

    A high-level look at neural network robustness, with many mentions of research done at Berkeley.

    When a mushroom is a pretzel and a sloth is a race car: deep neural networks can break in unpredictable ways, with implications way beyond misclassifying images.
    2 proslijeđena tweeta 9 korisnika označava da im se sviđa
    Poništi
  4. proslijedio/la je Tweet
    22. kol 2019.

    We're releasing a new method to test for model robustness against adversaries not seen during training, and open-sourcing a new metric, UAR (Unforeseen Attack Robustness), which measures how robust a model is to an unanticipated attack:

    15 replies 221 proslijeđeni tweet 645 korisnika označava da im se sviđa
    Poništi
  5. 21. kol 2019.

    While adversaries evolve their attack strategies and do not limit themselves to l_p perturbations, most academic research assumes just the opposite. We introduce several new attacks beyond l_p perturbations and measure robustness to unforeseen attacks:

    1 reply 3 proslijeđena tweeta 11 korisnika označava da im se sviđa
    Poništi
  6. 8. kol 2019.

    PyTorch 1.2 implemented the GELU activation function proposed in my first undergrad paper with : While the ReLU is x * <step function>, the GELU is x * <smoothed step function>. Now the GELU the default activation in GPT-2 and in BERT.

    [v1.2.0] New TorchScript API with Improved Python Language Coverage, Expanded ONNX Export, NN.Transformer Blog: Release notes: Install commands on (the non-default installs have changed)
    Prikaži ovu nit
    15 korisnika označava da im se sviđa
    Poništi
  7. 17. srp 2019.

    Natural Adversarial Examples are real-world and unmodified examples which cause classifiers to be consistently confused. The new dataset has 7,500 images, which we personally labeled over several months. Paper: Dataset and code:

    11 replies 194 proslijeđena tweeta 564 korisnika označavaju da im se sviđa
    Poništi
  8. 2. srp 2019.

    (2) is currently best done by spatial and temporal consistency checks ( work at ECCV 2018 and ICLR 2019), ensembling during evaluation (leaving dropout on during testing), and extreme input randomization (E-LPIPS, Barrage of Random Transformations). (3/3)

    2 korisnika označavaju da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  9. 2. srp 2019.

    (1) involves training on adversarial noise (which requires harnessing more training information through pre-training, self-supervised learning, mixup augmentation, and semi-supervised learning) or automatically smoothing the loss surface (Guided Complement Entropy). (2/3)

    1 reply 3 korisnika označavaju da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  10. 2. srp 2019.

    Currently, adversarial robustness can be improved by (1) increasing stability on noise or by (2) reducing the freedom of the attacker. (tweet 1/3)

    1 reply 5 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  11. 2. srp 2019.

    Our new work shows how self-supervised learning can surpass supervised learning for out-of-distribution detection, and that it can improve adversarial, common corruption, and label poisoning robustness:

    1 proslijeđeni tweet 11 korisnika označava da im se sviđa
    Poništi
  12. proslijedio/la je Tweet
    22. lip 2019.

    A dragonfly mislabeled as a manhole cover? What about a squirrel confused for a sea lion? These images are challenging algorithms to be more resilient to attacks.

    1 reply 11 proslijeđenih tweetova 20 korisnika označava da im se sviđa
    Poništi
  13. 1. lip 2019.

    Translated images are enough to break ConvNets. Richard Zhang proposes a max pooling modification that dramatically increases stability, even on images changed by ImageNet-P perturbations such as Gaussian noise, resizing, and viewpoint variation.

    9 proslijeđenih tweetova 31 korisnik označava da mu se sviđa
    Poništi
  14. 21. tra 2019.

    Our paper _Using Pre-Training Can Improve Model Robustness and Uncertainty_ will be presented at ICML 2019. In the paper we bring adversarial robustness from 46% to 57%, take issue with a previous characterization of pre-training, and more. Pre-print here:

    8 korisnika označava da im se sviđa
    Poništi
  15. proslijedio/la je Tweet
    18. ožu 2019.

    We ( ,, , & myself) are organizing a workshop on "Uncertainty & Robustness in Deep Learning" at . See for more info. Please submit your work (deadline: April 30, 2019) and attend the workshop! :)

    1 reply 34 proslijeđena tweeta 107 korisnika označava da im se sviđa
    Poništi

@DanHendrycks još nije objavio nijedan Tweet.

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:

    ·