Ezequiel López-Rubio (2018), Minds and Machines 28(4), 667–688 — DOI-10.1007/s11023-018-9480-7
Key takeaways
- Deep networks’ success comes from extracting successively higher-level features across multiple layers. López-Rubio argues that this calls for a new version of computational functionalism, keyed to deep neural computation rather than to classic program-based functional organization.
Seeded from the abstract alone, not a reading. What follows is what the work says about itself.
Deep learning, machine learning in deep artificial neural networks, has repeatedly reached human-like performance on intelligent tasks, and these networks resemble biological ones in many respects. That raises the question whether the resemblance explains the performance. The paper reviews the structure and learning processes of artificial and biological networks. It names two closely linked reasons for deep learning’s success: hierarchical extraction of increasingly high-level features, and multi-layer structure. After remarks on how the debate should be framed, it assesses deep networks as models of the brain from the standpoint of how similar the representations are. It ends by proposing a version of computational functionalism fitted to deep neural computation instead of classical program-based functionalism.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 0 seconds of active reading over 0 sessions. Deferred because nobody has read it closely here yet, not on merit.
Priority for a deeper reading: low — t=0, and the paper is closed access. It is a background reference, not load-bearing for current notes.
What a deeper reading should check:
- A pre-LLM attempt to update functionalism for connectionist and deep systems. Useful background for LIT-135 (Seth’s critique of computational functionalism) and LIT-131 (Freeborn on understanding in deep learning).
- A deeper reading should find out what the “new version” actually says: is it a functionalism of representational hierarchies, of learning processes, or of architectures? It should also check whether it avoids the substrate-flexibility objections Seth raises.
- Check how “similarity of model representation” is operationalized, and whether it anticipates later representational-similarity work comparing deep networks with the brain.
Access when seeded: Only the Springer landing page (abstract, keywords, dates) was reachable. The article is subscription-only (Semantic Scholar lists it as CLOSED), and no green copy was found: ResearchGate and PhilPapers were not reachable, and no institutional repository copy turned up. Published online 5 Oct 2018 per Crossref and the Springer page; issue December 2018.