Gary Marcus

Rank 28 of 47
|
Score -1
8/1/2024, 5:58:40 PM
X
In reply to:
Thomas G. Dietterich
@tdietterich
·
90d
Alex Guyer and I explored this a bit in this paper: https://arxiv.org/abs/2203.02486
Thomas G. Dietterich
@tdietterich
·
90d
Of course, we can never guarantee to detect novelty. E.g., a novel obstacle visible only in UV will not be detected by an RGB camera. But foundation models are the best method we have invented to date. 6/
Thomas G. Dietterich
@tdietterich
·
90d
But the foundation model approach trains computer vision systems on everything we can find -- including synthetic and adversarial imagery. The resulting representations are much more capable of detecting outliers. 5/
Thomas G. Dietterich
@tdietterich
·
90d
Deep learning only learns to represent what it sees in the training data. Traditional supervised learning (including Bayesian methods) fails on outliers because the training data only contains the known object classes. 4/
Thomas G. Dietterich
@tdietterich
·
90d
The second error is claiming that generative AI can't handle this. In fact, generative AI is an important way to learn to represent the vast diversity of the world, and hence, novelty. 3/
Thomas G. Dietterich
@tdietterich
·
90d
UQ requires the system to be able to represent the novel situation well enough to detect that it is novel. For example, if the scene aliases with a normal scene, it will not appear novel. This applies to symbolic as well as neural systems. 2/

The statement constitutes public discourse as it engages substantively with a public issue related to the capabilities and limitations of deep learning and computer vision systems. The tone is constructive and aims to contribute to the ongoing discussion by suggesting additional resources and perspectives for consideration.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement strives to do no harm and is constructive in nature, suggesting further reading and additional perspectives without any negative or harmful language. [+2]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding by suggesting additional resources that could provide a more comprehensive view of the topic. It encourages a deeper look into the limitations of current approaches, which can foster empathy and compassion for the challenges faced in the field. [+2]
  3. Principle 4:
    I will engage in constructive criticism and dialogue with those in disagreement and will not engage in personal attacks or ad hominem arguments.
    The statement engages in constructive criticism by suggesting additional resources and perspectives without resorting to personal attacks or ad hominem arguments. It maintains a respectful tone and focuses on the content of the discussion rather than the individuals involved. [+2]