Latent cause inference (LCI) refers to the cognitive process of inferring hidden or unobservable causes that explain observed patterns of events. It explains how our brain groups similar experiences together and uses them to predict what might happen next.
How Does It Work? A Real-World Example
Let's consider a common example: encounters with dogs. Imagine someone who had a frightening experience with a dog - perhaps they were bitten after hearing aggressive barking. Their brain creates a "latent cause" that links together:
- The sight of dogs
- The sound of barking
- The experience of pain
This becomes a mental category that we might call “Dogs are dangerous”. However, the same person might have entirely different experiences with their friend's friendly dog. This positive experience gets grouped into a separate "Some dogs are safe" latent cause - one that associates this dog with no barking and a pleasant experience. This is why someone can be generally afraid of dogs but comfortable petting specific dogs they know well.
Digging Deeper
As its core, LCI is a mathematical framework that uses Bayesian probability theory. LCI is powerful because it formalizes how humans and animals can discover underlying structures in the world without being explicitly told what to look for. To do so, LCI determines the boundaries of generalization in learning and decision making – which association is updated based on new information, and which association is used to determine the next action.
In this framework, experiences that are similar are assigned to a single latent cause, and a new cause is created when an event differs substantially from past experiences, such as when meeting a friendly dog. This gives rise to a causal model of the world that is parsimonious, assuming as few latent causes as possible, and at the same time flexible, expanding to create new latent causes as needed to explain observations in a potentially changing environment.
What Do You Mean by Cause?
Although we refer to latent causes, these are not necessarily causal. In our example, the latent cause “Dogs are dangerous” did not cause the barking or biting. But because LCI is a (Bayesian) generative model, it assumes that there is a hidden structure - the latent cause - that “emits” an observable event according to the statistics of that cause. Thus, if the latent cause “Dogs are dangerous” emits the next dog we encounter, the dog is likely to bark and try to bite, whereas if “Some dogs are safe” is responsible for the next dog, it likely won’t bark.
A Probabilistic Model of The World
LCI is the process of inferring, based on the observations and their similarity to past events, which latent cause was active. It is a probabilistic model, giving rise to a distribution over possible latent causes: a non-barking dog that does not look like the friend’s dog has, perhaps, a 60% chance of having come from “Dogs are dangerous”, a 30% chance of coming from “Some dogs are safe”, and a 10% chance of coming from a new latent cause.
The model also takes into account the person's belief about how likely each latent cause is, which is called a prior belief. If someone thinks that most dogs belong to “Dogs are dangerous”, they are more likely to think that the next dog they see is doing to bark and bite.
Why This Matters for Mental Health
Understanding LCI is particularly valuable to understand fear conditioning, a prominent animal model for anxiety disorders, which are driven by strong emotional memories. Back to our example: the person who was once bitten by a barking dog. They have learned the latent cause “Dogs are dangerous”, that they persistently recall when seeing dogs. If, when a dog doesn’t bark or bite, they infer that this particular dog must belong to a different latent cause “Some dogs are safe”, that will leave their “scary” latent cause unchanged and they may develop a phobia of dogs.
Understanding LCI can also be valuable for mental health treatment, especially in approaches like cognitive behavioral therapy. In seeking treatment for their dog phobia, the person is asked to “expose” themselves to dogs. For example, they have therapy sessions at the dog park. If these exposures update the old latent “Dogs are dangerous” - which we will rename “Dogs are safe” - to no longer predict barking and pain, they might be cured of their phobia. However, if they create a new latent cause “During therapy, dogs are safe”, and associate the dog-park experiences with that cause, the new experience will not modify their original fear construct, and their phobia will likely relapse.
LCI shapes how we learn from experience and adapt to new situations. By understanding this process, we can better grasp why people respond differently to similar experiences and why some mental health treatments might work better for some individuals than others.