Workshop GRIN – Ethical Knowledge and Deep Learning

Preston Werner (Hebrew U, Jerusalem); Jonna Vance (Northern Arizona U); Ana Gantman (Brooklyn College).

Date: 12 mars 2020, 13h30 – 16h
Emplacement: MILA, auditorium 1, 6650 Saint-Urbain, 2e étage


1:30-2:15 ANA GANTMAN (Brooklyn College, New York)

The Moral Pop-out Effect: Why, How, and Who cares?

Abstract: Moralization is the process by which preferences become values. Once a behavior enters into the moral domain, there are observable changes in how we process information related to that behavior. For example, our moral values feel more objective, widely shared, and central to our identities. I will argue that morally relevant stimuli are also more likely to reach our conscious awareness. Evidence suggests that visual experience is tuned to morally relevant stimuli, a phenomenon we term the moral pop-out effect. Specifically, people detect moral words (e.g., kill, moral, should) with greater frequency than non-moral words (e.g., die, useful, could) when the words are presented ambiguously. The moral pop-out effect is motivationally sensitive, and likely due to a lower threshold for reaching conscious awareness (as evidenced by enhanced P3 activity compared to non-moral words). Finally, morally relevant words that capture attention are also more likely to be shared on the social media platform Twitter.  Taken together, this work provides evidence for thinking about moral psychology not only in terms of what kinds of content fall into the moral domain, but how domain-general processes are tuned and utilized to create a specific kind of signature for moral cognition.


2:20-3:05 PRESTON WERNER (Hebrew U of Jerusalem)

Moral Perception and the Epistemic Structure of High-Level Experience

Abstract: Philosophers of perception have disputed whether visual perceptual experience is limited to representations of low-level properties like color, shape, and motion, or whether it can also represent natural kind properties, social properties, and moral properties. What is widely agreed between both sides, however, is that *if* high level properties can be experienced, their ability to justify beliefs is epistemically dependent on our experience of low level properties. While this thought is intuitive, I argue that the intuition in its favor turns on conflating representation that occurs in perceptual processing and representation that is (agentially) experienced. Once the distinction is made clear, we can see that the question is partially empirical. I sketch a few models about how this might work, with an eye toward moral perception in particular.


3:10-3:55 JONNA VANCE (Northern Arizona U)

Moral knowledge by abstraction

Abstract. Deep learning neural networks provide a promising model for human perceptual and affective learning. Yet the question of how far these models can be extended to higher cognition remains controversial and untested. In this paper, I argue that social and moral knowledge is obtainable from lower-level perceptual and affective feature spaces via transformational abstraction in layered networks. This empiricist approach is then contrasted with Lawrence Kohlberg’s rationalist approach, in which the logical notion of reversibility is the key to moral development.