After explaining Markov and Friston blankets in particular, Bruineberg et al. demonstrate how difficult it is for these to adequately enclose real-world examples. One reason is the assumption of conditional independence, which they are based on. To underline their point of view, they choose an example from clinical medicine: To model the patellar reflex, they construe a simple Bayesian network and a Friston blanket.
I agree with Bruineberg and colleagues that the assumption of conditional independence in Friston blankets is problematic for real-world examples. Consequently, their theoretical debate on Friston blankets makes sense. However, their example of the patellar reflex is insufficiently based on knowledge from neurophysiology or clinical medicine. This makes the example in this context problematic. As a result, I argue that the authors should refrain from applying this example as a proof of concept for their theoretical arguments. To make it clear what is problematic about their choice of example, let me first describe how they try to simulate the patellar reflex.
They propose a Bayesian network where different nodes are dedicated to different purposes: The patient's intention to move the leg, the doctor's intention to move the patient's leg, the spinal neurons, the intervention of striking the patellar tendon with a hammer, the motor command from the central nervous system sent to the spinal neurons, and finally, a third way (e.g., someone else) moving the leg. An example for their Bayesian network can be found in Figure 7a of their target article. Subsequently, they present an elaboration of the Bayesian network, where the nodes are partitioned into external states, internal states, sensory states, and active states. This partitioning is done in what they call a “Friston blanket,” which is a transfer of the Markov blanket idea from statistics (Pearl, Reference Pearl1988) to the life sciences (e.g., Friston, Reference Friston2013).
The Friston blanket from this example is problematic because of the following reasons:
Bruineberg et al. use nodes that incorporate the patient's intention. However, the patient's intention as well as the central motor commands are not part of a monosynaptic reflex arc. Consequently, they are not part of the patellar reflex, which is a monosynaptic reflex arc. Rather, the incorporation of intention would involve several neurons, interneurons and therefore several synapses between neurons. A reflex arc with several synapses, however, would be termed a polysynaptic reflex arc. Therefore, Bruineberg et al. have construed a Friston blanket for a polysynaptic reflex arc, although their aim was actually to model a monosynaptic reflex arc. As a result, they cannot argue that their Friston blanket does not adequately enclose the patellar reflex. To understand why their example cannot be applied as an argument against Friston blankets, let us consider the patellar reflex, which should not incorporate the patient's intention, as the synaptic transmission happens at the level of the spinal cord: After striking the patellar tendon with a hammer, the muscle spindle in the quadriceps femoris muscle is activated, followed by an afferent signal traveling along the sensory neuron to the dorsal root of the spinal cord. In the spinal cord, a monosynaptic transmission to an alpha-motor neuron takes place, which produces an efferent signal traveling along this alpha-motoneuron to the quadriceps femoris muscle, eliciting the movement (e.g., Ginanneschi, Mondelli, Piu, and Rossi, Reference Ginanneschi, Mondelli, Piu and Rossi2015). Although some interaction with interneurons at the level of the spinal cord is possible (Ginanneschi et al., Reference Ginanneschi, Mondelli, Piu and Rossi2015), it is obvious that the Friston blankets construed by Bruineberg et al., Figures 7b and 7c, involve intentional leg movements, which should not be the case.
If Bruineberg et al. had built a Friston blanket for a truly monosynaptic reflex, they could easily avoid the counterintuitive sensorimotor boundaries that they consider problematic for Friston blankets. I will attempt to do this by using the same terms and the same states as Bruineberg et al. in Figures 7b and 7c. I will only get rid of those parts that do not belong to a monosynaptic reflex arc: The patient's intention ip is not part of the monosynaptic patellar reflex, nor is the motor command sent from the central nervous system c, nor someone else kicking the patient's leg k. Rather, there would be nodes for the doctor's intention id – external state, the hammer h – sensory state, the spinal neuron s – internal state, and the motor command m active state. Given that all nodes represent different states and each state can lie on a different Friston blanket, there would be no counterintuitive sensorimotor boundaries in this example. Because Bruineberg et al. have not modeled a truly monosynaptic reflex arc with their choice of nodes, they see a problem. This problem vanishes when omitting the nodes that are not part of a monosynaptic reflex arc. Does this imply that their critique on counterintuitive sensorimotor boundaries with Friston blankets is not justified? I would argue that their critique is still justified and here is the reason why: There are several conditions in clinical medicine that would challenge Friston blankets. For example, the patellar reflex would require additional nodes for the muscle spindle / the quadriceps muscle. If these were added, as well as other internal states that change the extent of reflexes, for example, endocrinopathies (Rodriguez-Beato & De Jesus, Reference Rodriguez-Beato and De Jesus2021), electrolyte derangements (Espay, Reference Espay, Biller and Ferro2014; Hensle & Lambert, Reference Hensle, Lambert, Gearhart, Rink and Mouriquand2010), there would be several internal states. Consequently, there would be several Friston blankets with counterintuitive boundaries all impacting the extent of the patellar reflex, which would challenge Friston blankets.
To conclude, Bruineberg et al. contribute substantially to the theoretical debate on Friston blankets, but their idea to challenge Friston blankets with the patellar reflex example does not work due to the aforementioned shortcomings of their chosen model. If their example is modified and other relevant nodes for the patellar reflex are added, one can easily find counterintuitive sensorimotor boundaries, which would be a challenge for Friston blankets.
You have
Access
After explaining Markov and Friston blankets in particular, Bruineberg et al. demonstrate how difficult it is for these to adequately enclose real-world examples. One reason is the assumption of conditional independence, which they are based on. To underline their point of view, they choose an example from clinical medicine: To model the patellar reflex, they construe a simple Bayesian network and a Friston blanket.
I agree with Bruineberg and colleagues that the assumption of conditional independence in Friston blankets is problematic for real-world examples. Consequently, their theoretical debate on Friston blankets makes sense. However, their example of the patellar reflex is insufficiently based on knowledge from neurophysiology or clinical medicine. This makes the example in this context problematic. As a result, I argue that the authors should refrain from applying this example as a proof of concept for their theoretical arguments. To make it clear what is problematic about their choice of example, let me first describe how they try to simulate the patellar reflex.
They propose a Bayesian network where different nodes are dedicated to different purposes: The patient's intention to move the leg, the doctor's intention to move the patient's leg, the spinal neurons, the intervention of striking the patellar tendon with a hammer, the motor command from the central nervous system sent to the spinal neurons, and finally, a third way (e.g., someone else) moving the leg. An example for their Bayesian network can be found in Figure 7a of their target article. Subsequently, they present an elaboration of the Bayesian network, where the nodes are partitioned into external states, internal states, sensory states, and active states. This partitioning is done in what they call a “Friston blanket,” which is a transfer of the Markov blanket idea from statistics (Pearl, Reference Pearl1988) to the life sciences (e.g., Friston, Reference Friston2013).
The Friston blanket from this example is problematic because of the following reasons:
Bruineberg et al. use nodes that incorporate the patient's intention. However, the patient's intention as well as the central motor commands are not part of a monosynaptic reflex arc. Consequently, they are not part of the patellar reflex, which is a monosynaptic reflex arc. Rather, the incorporation of intention would involve several neurons, interneurons and therefore several synapses between neurons. A reflex arc with several synapses, however, would be termed a polysynaptic reflex arc. Therefore, Bruineberg et al. have construed a Friston blanket for a polysynaptic reflex arc, although their aim was actually to model a monosynaptic reflex arc. As a result, they cannot argue that their Friston blanket does not adequately enclose the patellar reflex. To understand why their example cannot be applied as an argument against Friston blankets, let us consider the patellar reflex, which should not incorporate the patient's intention, as the synaptic transmission happens at the level of the spinal cord: After striking the patellar tendon with a hammer, the muscle spindle in the quadriceps femoris muscle is activated, followed by an afferent signal traveling along the sensory neuron to the dorsal root of the spinal cord. In the spinal cord, a monosynaptic transmission to an alpha-motor neuron takes place, which produces an efferent signal traveling along this alpha-motoneuron to the quadriceps femoris muscle, eliciting the movement (e.g., Ginanneschi, Mondelli, Piu, and Rossi, Reference Ginanneschi, Mondelli, Piu and Rossi2015). Although some interaction with interneurons at the level of the spinal cord is possible (Ginanneschi et al., Reference Ginanneschi, Mondelli, Piu and Rossi2015), it is obvious that the Friston blankets construed by Bruineberg et al., Figures 7b and 7c, involve intentional leg movements, which should not be the case.
If Bruineberg et al. had built a Friston blanket for a truly monosynaptic reflex, they could easily avoid the counterintuitive sensorimotor boundaries that they consider problematic for Friston blankets. I will attempt to do this by using the same terms and the same states as Bruineberg et al. in Figures 7b and 7c. I will only get rid of those parts that do not belong to a monosynaptic reflex arc: The patient's intention ip is not part of the monosynaptic patellar reflex, nor is the motor command sent from the central nervous system c, nor someone else kicking the patient's leg k. Rather, there would be nodes for the doctor's intention id – external state, the hammer h – sensory state, the spinal neuron s – internal state, and the motor command m active state. Given that all nodes represent different states and each state can lie on a different Friston blanket, there would be no counterintuitive sensorimotor boundaries in this example. Because Bruineberg et al. have not modeled a truly monosynaptic reflex arc with their choice of nodes, they see a problem. This problem vanishes when omitting the nodes that are not part of a monosynaptic reflex arc. Does this imply that their critique on counterintuitive sensorimotor boundaries with Friston blankets is not justified? I would argue that their critique is still justified and here is the reason why: There are several conditions in clinical medicine that would challenge Friston blankets. For example, the patellar reflex would require additional nodes for the muscle spindle / the quadriceps muscle. If these were added, as well as other internal states that change the extent of reflexes, for example, endocrinopathies (Rodriguez-Beato & De Jesus, Reference Rodriguez-Beato and De Jesus2021), electrolyte derangements (Espay, Reference Espay, Biller and Ferro2014; Hensle & Lambert, Reference Hensle, Lambert, Gearhart, Rink and Mouriquand2010), there would be several internal states. Consequently, there would be several Friston blankets with counterintuitive boundaries all impacting the extent of the patellar reflex, which would challenge Friston blankets.
To conclude, Bruineberg et al. contribute substantially to the theoretical debate on Friston blankets, but their idea to challenge Friston blankets with the patellar reflex example does not work due to the aforementioned shortcomings of their chosen model. If their example is modified and other relevant nodes for the patellar reflex are added, one can easily find counterintuitive sensorimotor boundaries, which would be a challenge for Friston blankets.
Financial support
No funding received in relation to this manuscript.
Conflict of interest
No conflict of interest in terms of this commentary. Considering all possible conflicts of interest, Rainer Spiegel owns a small number of stocks from the respirator/ventilator company Draeger in his private portfolio.