Yanny or Laurel? A perspective from the science of mind and brain

I really like the Yanny versus Laurel meme, which exploded yesterday. It helps illustrate some key points about human perception:

  1. In some situations people can differ wildly in their experience of low-level perception.
  2. Active top-down expectations (and other, weirder processes) have a strong effect on low-level perception.

So basically, it’s an auditory version of #ThatDress.

Continue reading

Advertisements

What is a biological model? Here’s a useful categorization system for people interested in neuroscience, cognitive science, and biology

I found an excellent classification of models in a paper on neurogenesis: Using theoretical models to analyse neural development.

I think this should be illuminating for anyone interested in theoretical, mathematical and/or computational approaches in neuroscience, cognitive science, and biology.

There are several ways in which models of biological processes can be classified. 

Formal or informal models

Informal models are expressed in words or diagrams, whereas formal models — which this Review is concerned with — are described in mathematical equations or computer instructions. Using formal language forces a model to be precise and self-consistent. The process of constructing a formal model can therefore identify inconsistencies, hidden assumptions and missing pieces of experimental data. Formal models allow us to deduce the consequences of the postulated interactions among the components of a given system, and thus to test the plausibility of hypothetical mechanisms. Models can generate new hypotheses and make testable predictions, thereby guiding further experimental research. Equally importantly, models can explain and integrate existing data.

 Phenomenological or mechanistic models 

Most formal models lie on a continuum between two extreme categories: phenomenological and mechanistic. A phenomenological model attempts to replicate the experimental data without requiring the variables, parameters and mathematical relationships in the model to have any direct correspondence in the underlying biology. In a mechanistic model, the mathematical equations directly represent biological elements and their actions. Solving the equations then shows how the system behaves. We understand which processes in the model are mechanistically responsible for the observed behaviour, the variables and parameters have a direct biological meaning and the model lends itself better to testing hypotheses and making predictions. Although mechanistic models are often considered superior, both types of model can be informative. For example, a phenomenological model can be useful as a forerunner to a more mechanistic model in which the variables are given explicit biological interpretations. This is particularly important considering that a complete mechanistic model may be difficult to construct because of the great amount of information it should incorporate. Mechanistic models therefore often focus on exploring the consequences of a selected set of processes, or try to capture the essential aspects of the mechanisms, with a more abstract reference to underlying biological processes. 

Top-down or bottom-up models 

Formal models can be constructed using a top-down or a bottom-up approach. In a top-down approach, a model is created that contains the elements and interactions that enable it to have specific behaviours or properties. In a bottom-up approach, instead of starting with a pre-described, desired behaviour, the properties that arise from the interactions among the elements of the model are investigated. Although it is a strategy and not a type of model, the top-down approach resembles phenomenological modelling because it is generally easier to generate the desired behaviour without all of the elements of the model having a clear biological interpretation. Conversely, the bottom-up approach is related to mechanistic modelling, as it is usual to start with model elements that have a biological meaning. Both approaches have their strengths and weaknesses.

(I removed citation numbers for clarity.)

One point might be relevant here: a model is neither true nor false — ideally it’s an internally consistent mini-world. A theory is the assertion that a model corresponds with reality.

How is depression (and other mental disorders) addressed with computational neuroscience?

Nowadays there are broadly two sorts of computational neuroscientist: those who analyze experimental data using statistical methods, and those who propose computational models aimed at a theoretical understanding. I belong to the latter category, so what I say doesn’t really apply to data analysis issues.

Computational or mathematical modeling/theory is still at a very rudimentary stage when it comes to addressing clinical neuroscientific questions. The sheer complexity of the brain’s neurons, connections, and dynamics prevents a direct “brute force” modeling of how the cells interact to generate behaviors and disorders.

But highly simplified models can be used to explore hypotheses, or the implications of experimental findings. So a schematic computational model can test out how a particular brain circuit might work, and how “breaking” the circuit in various ways could correspond with particular disorders.

For instance, depression could be modeled as a weakening of cells in the hippocampus that help a person discover new possibilities in life. Depression is often described as an inability to move out of the apathetic “place” you find yourself in. Metaphorically speaking, discovering new “places” to go to may involve the formation of new neurons (neurogenesis) in the hippocampal circuit — a process that may be disrupted in some depressed patients. (For more on the evidence read this excellent New York Times article on depression The Science and History of Treating Depression)

The (ideal!) steps involved in computational modeling of disorders might go something like this:

  1. A network of artificial neurons can be arranged in a circuit inspired by anatomical data.
  2. The dynamic properties can then be derived from physiological data. (Cell firing, fMRI, EEG etc.)
  3. The global activity or output of the network is related to some behavior, such as goal-directed decision making
  4. Changing parameters in the network or in the artificial neurons can be investigated. Weakening some parameters — such as inhibitory or excitory strength — may make the system resemble a “depressed” person or animal who is incapable of making a decision. Enhancing some other parameter may appear equivalent to the action of a drug.
  5. The modeling results can be discussed with experimentalists, and ways to verify or falsify the results can be invented.
  6. In the event of a falsification (which is pretty much all the time!) you return to step 1!
  7. If all goes well, a well-tested model can be introduced to clinicians and the general public, and can inform treatment strategies. As far as I know this has yet to happen! 🙂

A computational model I have developed in conjunction with anatomists and another modeler looks at a circuit centering on the amygdala, which is believed to play a major role in emotion, and in emotional disorders. The model suggests that prefrontal cortical modulation of the amygdala can push the system into a “cautious” state, or a more “reckless” state. And this modulation may be related to pathological brain states. We also show that weak attention could relate to overgeneralization, a problem that seems to occur in phobic patients.

Our model is highly simplified, but hopefully it will throw up some ideas for experimentalists to take further.

You can read more about our model here: Anatomy and computational modeling of networks underlying cognitive-emotional interaction | Frontiers in Human Neuroscience

(Pardon the self-promotion!)

View Answer on Quora