Concepts Are Compressed Models

A Concept is an interface to expose and connect some Models’ salient features.

It extracts key distinctions from existing models, amplifies them and “fades out” the rest to reduce granularity and “boost the signal”, i.e. what is important or interesting about the models in a specific context.

A concept is thus a “lossy compression” of these models: it reduces bits by removing unnecessary or less important information, and thus reduces the resources required to store and transmit data about the models.

Extracting distinctions from models

To extract key distinctions from existing models, we first need to identify them as relevant. There are three ways in which we do this:

1. We import and implicitly adopt distinctions from other areas with Conceptual Metaphors. The critique of their metaphorical content can be an emancipatorySee e.g. Haslanger (2020).
   
   and therapeutic actSee e.g. Lakoff and Johnson (1999).
   
   .
2. We identify distinctions that track differences in the world in a way that makes the models (more) useful, i.e. enables successful predictions. These differences are the Affordances of the modelling System’s environment.
3. We rely on distinctions that support existing systems of social relations by enabling, justifying or making invisible Power structures and dynamics. This is an ideological choice of distinctions.

There is no sharp distinction between the latter two: The criterion is usefulness in both cases, and its operationalisation is determined by the wider social system – “objectivity” is always a matter of perspective.

The ideological choice of distinctions can nonetheless be uncovered by genealogy, ethnography, and political or systems analysis. This is where Critical Theory and Conceptual Engineering come in: To disrupt and expose the identification of distinctions as politically motivated, unjust, unhelpful, and, first and foremost, socially constructed.

Working with models

We also use concepts as templates to more easily

• generate new models using the existing distinctions (“scaffolding”);
• retrieve existing models, the “label” making accessing memory easier;
• connect models, i.e. relate compressed models so they form larger networks. This enables Network Effects: Concepts (and thus models) become more valuable the larger the network; the Concept Network’s value grows exponentially with new concepts. This is how concepts facilitate the generation of compound Knowledge.

Related conceptions

The outlined approach is similar to Chris Eliasmith et al.’s conception of concepts as “semantic pointers”:See e.g. Blouw et al. (2015).

• They use vectors and vector spaces as modelling tools for multidimensional concepts, multimodality, and the distinction between occurrent and conceptual representation.
• In their framework, a semantic pointer is “a compressed representation that captures summary information about a particular domain”ibid., 5
  
  , “a vector encoded by the spiking activity in a population of neurons”ibid., 7
  
  , and “an entity that enables the occurrence of a concept rather than … an entity that is equivalent to a concept”ibid.
  
  .
• The brain uses a “(de)compression network” to translate between perceptual representations and pointers. The pointers are independently usable from occurrent representation and thus enable multimodal representation and the combination into higher-level concepts.

References

• Blouw et al. (2015): “Concepts as Semantic Pointers: A Framework and Computational Model”
• Fauconnier & Turner (2002): The Way We Think: Conceptual Blending and the Mind’s Hidden Complexities
• Haslanger (2020): “How Not to Change the Subject”
• Lakoff and Johnson (1999 ): Philosophy in the Flesh