Two constraints on the neuroscience of content

Charles Rathkopf

Center for Philosophical Psychology

March 21st, 2024

What are the theoretical limits on mind-reading?

  1. The relevant data is too hard to acquire
  2. The relevant data is too hard to interpret
  3. The relevant data is not (exclusively) located in the brain

What is mind-reading?

Decoding mental content from brain data such that:

  1. No conventional symbols
  2. Driven by neural data

No conventional symbols

Rule out the intentional production of conventional symbols like words.

  1. Thought-to-text devices
  2. fNIR for ALS patients

Driven by neural data

  1. Social cognition takes behavior as input
  2. Mind-reading takes exclusively neural data as input
    • e.g. predictive algorithm cannot take behavioral response latency as input

Typical decoding experiment

  1. Each measurement site in the brain is represented as one dimension in a high-dimensional space
  2. ML used to associate each point with some property of the experimental condition in which it was elicited
  3. Typically stimulus classification, sometimes reconstruction

From stimulus prediction to attitude ascription

  1. Speeded one-back repetition detection task.
  2. Algorithm predicts house/face & correct/incorrect from brain data
  3. Where t is a correct house trial, we infer “S believes that the image on trial t was a house”

Too much rational reconstruction?

  1. Images of gesture
  2. Dutch hand-ear gesture
  3. Infer: S believes that the person in the photo likes the food

Decoding propositional content

  • Tang et al. (2023) reconstructs continuous natural langauge from both perceived and imagined speech
  • Relies on neural representations of semantic meaning, rather than acoustic or phonetic information
  • Decoder trained within-subject on 16 hours of fMRI data, with narrative audio

Results from perceived speech experiment

Decoded Language from Tang et al. (2023)
from Tang et al. (2023)

Rational reconstruction in a computer program

Experimental pipeline
from Tang et al. (2023)

Need for generative models

  1. LLM generates new language based on plausibility
  2. This is a form of rational reconstruction
  3. Conjecture: some form of rational reconstruction will always be necessary

Deep theories of belief

1. Representationalist

    - Beliefs are relations to structured mental 
2. Psychofunctionalist 

    - Representations are explicitly stored in the brain 
    - Yes-or-no phenomenon
    - To be discovered by mind/brain sciences 

Superficialist theories of belief

Whether an agent has the belief that P depends only on 
whether their patterns of actual and potential behavior 
come sufficiently close to matching a stereotype of 
rational action for agents with the belief that P

Superficialists are committed to the conjecture

  1. Stereotypes of rational action are essential to belief
  2. Stereotypes depend on population-level facts about behavior
  3. Therefore, belief-determining facts not (only) in the brain
  4. Therefore, no empirical discovery will relieve us of the need for rational reconstruction.

Trumping cases

1. Behavioral data, including biographical information, 
support the claim that the subject believes that P, but

2. Brain data supports the claim that the subject 
believes that ~P, and 

3. The brain data constitutes stronger evidence 
about what the subject really believes than the 
behavioral evidence

Have we made progress?

  1. Added empirical commitment to superficialist theory of belief
  2. Sharpened the distinction between the deep and superficialist views
    • worry: Mandelbaum & Schwitzgebel talking past each other
    • Trumping cases show that disambiguation of “belief” does not dissolve the disagreement

Thanks for your attention!