13 Rebranding Sciences
13.1 Key Message
đź‘Ť Consequence 2: Towards a rebranding
Hard & Soft must be replaced with meaningful & accurate terms for both skills & sciences
13.2 Consider what unites hard & soft, not what divides
False divisions: - Hard & soft characterises skills & scientific disciplines using dichotomous, binary metrics which necessitates false divisions
Real unifiers across disciplines: - The entire scientific enterprise, encompasing all disciplines, either “hard” or “soft”, is driven by our need/desire to understand the world around us. - Each discipline focuses on a part of the story. - The “hard” sciences: Physics, chemistry, geology, astronomy, etc. - The “soft” sciences: Biology, philosophy, sociology, anthropology, theology, etc. - But these seemingly disparate disciplines are united by the use of models that simplify the complexity of some part of the “real” world. - They all try to capture the complexity of the world in some explainable terms. - These models, in all their varied forms, allow us to explore and experiment. They expand our understanding of the world and help explain the world in terms we can comprehend and experience.
13.3 Instead of hard & soft, describe the models used with words that convey something more meaningful
Features the distinguish & unite disciplines: - The subject matter, which serves as input and dictates the output, is a distinguishing feature - The models themselves can be described using a set of metrics that range from static to dynamic. - The idiosyncratic composition of these metrics makes up the models specific to a discipline. - All disciplines have their model, but the idea is unifying
Hard Skills & Hard Science | Soft Skills & Soft Science |
---|---|
Static | Dynamic |
Finite | Infinite |
Defined | Vague |
Certain | Uncertain |
Simple | Complex |
Definable | Undefinable |
Knowable | Unknowable |
Sharp | Fuzzy |
13.4 So what is a model: Imagine a generic instance
A naked model, generically defined, is:An instrument (read tool, algorithm) that receives input and, through processing, returns an output of a different form A model is made up of :Hyper-parametersThe model’s overall structure and functional definition, including the total number of parameters and how they interact with each other. ParametersThe variables/metrics/features that describe the input CoefficientsThe fitted values for each parameter In dressing a naked model, we provide …Input - Measured or predicted values for each parameter. … and expect to receive a …Response - The output, input in a new state
13.5 Properties to describe static and dynamic models
Organizational Level | Dimension | Static Models | Dynamic Models |
---|---|---|---|
Model | Definition | Fixed, pre-defined | Dynamic, context-specific |
Applicability | Generic, adaptable to many situations | Specific, designed for a specific purpose | |
Structure | Mostly simple | Complex, undefined, or black box | |
Execution | linear | reitative | |
Amount of data | small to very large | typically very large | |
Ensemble | possible | typical for larger models | |
On-the-fly change | No | Yes (e.g. learning rate, epoch) | |
# of influencing variables | Typically small & countable | Large to unknowably large, poorly defined, Changing | |
Knowability of material | Completely known or knowable | Incompletely known | |
Applicability | Universal | Contextual | |
Accuracy | 100% or inapplicable | <100% or unknowable | |
Scope | Complete | Partial | |
Mindset | Analytical thinking (reductionist) | Critical thinking (systems thinking) | |
Scope | Functional, individual and isolated units | Systems | |
Novelty | In pre-production (design and experiment) | In post-productions (Emergent properties) | |
Thinking Fast & Slow | System 2 slow | System 1 fast | |
Perception | Rational | Intuitive | |
Potential to surprise | Unlikely | Likely | |
Hyperparameters | Structure | Mostly simple | complex, undefined or black box |
Number | low | large to very large | |
Parameters | Composition | Fixed | Flexible |
Amount | Small, countable | Large, potentially uncountable | |
Coefficients | Accuracy | As high as possible | High, or unknowable |
Number of iterations | 1 | Determined by structure and input | |
Depth | Shallow | Deep | |
Defined by observation | Yes | Less likely | |
Changeable in number | No | Yes | |
Named | Yes | No | |
Defined by observation | Yes | Less likely | |
Potential for systematic Bias | High, if assumptions are not met | Low, fewer or no assumptions | |
Importance of preprocessing | High, can invalidate model | Low, potentially unnecessary |
13.6 Finally, binary distinctions are replaced with information-rich scales
- As we saw, binary metrics reinforce a false dichotomy and value hierarchy in STEM
- And the entire world can’t be reduced to binary variables, the lowest information content possible (True/False, 1/0)
- Thus, for each dimension defining the model, hyperparameters, parameters & coefficients above, we need metrics that retain more information:
- Continuous, ordinal, nominal, binary
- A lower and upper limit
- Single, range or multipoint descriptions.