Causal Inference in the perception of self-motion
Consider the well-known illusion of self-motion that arises when a person, seated in a stationary train, sees a neighboring train depart: often, this gives a fleeting sensation of self-motion, while the other train is perceived to be stationary. What is happening in the brain that may explain this illusory perception?
People possess multiple sensory systems that provide the brain with information on motion stimuli: visual motion is seen with the eyes, and physical motion can be felt via the inertial sensory system, which is comprised of the vestibular system of the inner ear and various kinds of pressure-sensitive neurons distributed throughout the body.
At the onset of the illusion, the visual system signals to the brain that there is translational motion, whereas the inertial system outputs only background noise. In itself, the visual information is ambiguous: the motion that is seen can just as well have been caused by motion of the own train (self-motion), as by motion of the other train (object-motion). Furthermore, when a motion is of small intensity, it may be below the detection threshold of the inertial system. Given that there is noise in the system, the absence of inertial information is thus not necessarily incompatible with the small visual signal.
To infer the cause of the visual stimulation, the brain may resort to prior knowledge about the world. In this case, this knowledge may reflect that coherent motion of the entire visual field generally results from self-motion. Hence, taking all the information into account, the brain decides that it is more likely that there is self-motion than object-motion. However, as the other train keeps accelerating, the intensity of the visual signal keeps increasing, while the inertial signal fails to appear. Ultimately, the discrepancy between the two sensory systems becomes incompatible with the interpretation of self-motion, resolving the illusion and resulting in a perception of object-motion: i.e., the other train moving
The mechanism described here, of forming perceptions out of sensory signals taking into account inferences on their causality as well as prior knowledge, is described by a class of statistical models of perception known in the scientific literature as Causal Inference models.
Researchers in the Motion Perception and Simulation research group have set out to test the hypothesis that the brain performs Causal Inference in the process of forming perceptions of self-motion and spatial orientation.
Although the Causal Inference model appears to be consistent with the illusion outlined above, it is not known whether the brain actually works like this. The models take into account characteristics of the visual and inertial sensory information itself, and make specific predictions on the perceptions that result in response to all possible combinations of visual and inertial stimuli –including discrepant combinations that would not normally arise.
To test the applicability of the models thus requires the characterization of the output of the individual sensory systems in response to specific stimuli; making predictions on perceptions for all possible multisensory stimuli –as well as measuring these perceptions; and evaluation of model performance.
Each step poses a number of problems.
First, how does one measure perceptions? It is not yet possible to record perceptions directly from the brain; the only way is to ask participants to provide estimates of their perceptions. This is no mean feat, as the concept of self-motion perception is multisensory, and comprises of many characteristics: translations and rotations in multiple dimensions, with variable magnitudes and directions. Here, perceptions were evaluated one characteristic at a time. This was done for each sensory system in isolation as well as for their combination.
Second, how does one present participants with discrepant combinations of visual and inertial stimuli experimentally? To simulate situations analogous to the train illusion for different characteristics of motion, the researchers used the Max Planck motion simulator facilities (the CyberMotion Simulator and the Cyberpod) in combination with high-end visualization equipment (Figure 1).With this equipment, participants could be presented with any desired combination of inertial and visual stimuli with any discrepancy, while maintaining a high degree of ecological validity.
Third, how does one determine from the data if the brain performs Causal Inference? Given that human perception is inherently noisy, the predictions and measurements of perception cannot be expected to perfectly coincide. Here, the researchers compared the performance of models covering the spectrum of possible mechanisms of perception. When the Causal Inference model performed best, this was taken as evidence supporting the hypothesis.
By this methodology, the tenability of the Causal Inference model was tested for the perception of heading[1,2,3], i.e., the direction of horizontal linear self-motion; and for perception of verticality –the upright.
In the first study on heading perception[1,2], tolerances for discrepancies between visual and inertial heading were found to be quite large (Figure 2). In some cases, even discrepancies as large as 90° appeared to go unnoticed. The results thus suggested that perceptions could be best described by a model whose behavior resembles weighted averaging of the sensory information –negating the need for a Causal Inference model. However, by increasing the range of discrepancies in subsequent work, decisive evidence in favor of Causal Inference was obtained. Similarly, recent work on perception of verticality, with discrepancies between the visual and physical upright up to 27°, suggests that perceptions of orientation are constructed as a weighted average. However, a closer look at individual results suggests that this conclusion does not hold for all participants, and favors a Causal Inference model in particular cases.
The results of the different studies are not unequivocal. It turns out that it is not straightforward to tune an experiment such that the chosen stimuli provide the proper resolution to reveal the mechanisms by which perceptions of self-motion are made. However, by evaluation of data from individual participants, the evidence over a number of studies is now converging towards a conclusion that perceptions of self-motion reflect Causal Inference.
More studies will need to be performed to assess the mechanisms by which other characteristics of self-motion are formed. Ultimately, the results of these studies need to be fused into a comprehensive model of self-motion perception.
Apart from its neuroscientific value, this model could for instance provide a tool to assess the efficacy of motion simulations, and could predict whether humans can tell up from down on Mars.
- De Winkel, K.N.; Katliar, M.; Bülthoff, H.H. Forced Fusion in Multisensory Heading Estimation. PLoS One. 2015; 10.5
- De Winkel, K.N.; Katliar, M.; Bülthoff, H.H.
Heading Coherence Zone from Causal Inference Modelling
Proceedings DSC 2015 Europe: Driving Simulation Conference & Exhibition, 67-70 (2015)
- De Winkel, K.N.; Katliar, M.; Bülthoff, H.H. Causal Inference in Multisensory Heading Estimation. PloS One. 2017; 12.1
- De Winkel, K.N.; Katliar, M.; Diers, D., Bülthoff, H.H. What's Up: an assessment of Causal Inference in the Perception of Verticality. bioRxiv. 2017