While it may be somewhat sensational to say that he is working on superintelligent cyborg mice, it may not be inaccurate. He definitely is working on what folks around here call “intelligence amplification.” But enough preamble.
Kellendonk’s lab has an experiment where they are inserting electrical stimulation probes into the medio-dorsal thalamus of rats which enhances their ability to complete a cognitive task. Kellendonk believes that his cybernetic modifications enhance the rat’s ability to retain a memory for completion of a memory task, but I have a different interpretation. Regardless, there is a definitive improvement in the ability of the rats to complete the task under stimulation.
For those unfamiliar with the thalamus, it is the place all coritco-thalamic loops pass through and in many ways is the heart of the brain’s circulatory system, pumping waves of electrochemical cascades through various parts of the cortex, as I explained last week. Different specific parts of the thalamus are linked architecturally in these thalamo-coritcal loops with different areas of the cortex. The medio-dorsal part of the thalamus links mainly to the prefrontal cortex although there is some linkage to the anterior cingulate cortex.
Linkages to the Prefrontal Cortex from various parts of the Medio-Dorsal Thalamus from Li, Kaixin, et al. “The human mediodorsal thalamus: Organization, connectivity, and function.” NeuroImage 249 (2022): 118876.
In his experiments Kellendonk’s lab first showed a link between a prefrontal area and the thalamus evidenced by simultaneous activation patterns; he then uses genetic designer mice which can have their neural connections inhibited with green light to show that a probe inhibiting the prefrontal areas directly or the region they are connected to in the thalamus will cause the mouse to perform worse on a cognitive task.
In particular the mice were sent through a T-maze twice; on the first trial a reward was available on either left or right arm of the T maze, the mouse retrieves it and is trained to return to the start position and wait through a delay; on the second trial they had to choose the arm of the T-maze opposite to the arm where the reward was available on the first trial (or else be trapped in a rewardless arm by a quickly closed gate). The delay between the two trials was an adjustable variable.
Kellendonk’s mouse maze.
The mice were shown to perform worse on this task when the light from the probe was activated, inhibiting the neurons in either the prefrontal cortex or the thalamus which was connected to it [1]. In a separate experiment Kellendonk then demonstrated that direct electrical stimulation to the medio-dorsal thalamus produced the opposite effect: an improvement in performance on the task.
Difficulties with working memory are a common symptom of schizophrenia; Kellendonk’s lab is part of a medical psychiatry center at Columbia studying schizophrenia. He interprets these results as improving the working memory of the mice and deep brain thalamic stimulation of schizophrenic patients is being explored as a cutting edge treatment. But I think I may have a different interpretation.
Rather than working memory encoding in which direction lay the reward in the previous trial I think what may be going on here is the prefrontal areas are encoding the rule “choose the opposite direction,” which is accessed when formulating a plan. I would hypothesize that upon detecting the reward in the first trial a plan is formulated, based on this rule, to go the opposite direction either immediately or upon some reflection. This is stored in working memory somewhere either in some kind of frequently reused system which is flushed of contents for use in a new task frequently or some other way which naturally degrades quickly [2]. I hypothesize that these prefrontal areas are merely accessed when rewriting the plan to working memory but themselves are an element of long term memory.
So what I propose happens is the mouse retrieves a reward down the left pathway and formulates a plan to return to the start and go right next time in working memory. The mouse then returns to start and waits. This plan is compressed and does not record why the choice to go right this time is made, just that the plan is to go right. At some point, depending on wait time, the place storing the working memory of what to do is wiped likely because the mouse started thinking about something else and needed that bit of scratchpad [2]. Then the mouse rewrites the simple plan after retrieving the memory of which path it took on the first trial, probably from elsewhere in the brain with more connections to sensory and motor areas rather than the prefrontal areas, and combines that with this prefrontal learned rule to reformulate the simple plan.
And I think this fits better with the general functions of the prefrontal cortex than assuming the prefrontal cortex is involved with retaining sensory experiences and recording motor output; these are functions more likely associated with the occipital, temporal, or parietal lobes. The prefrontal cortex contains the most abstract neural architecture for high level decision making and stores things like patterns of plans or understandings of things like procedures or laws: i.e. “rules.”
I suppose this doesn’t fully contradict Kellendonk’s findings. You can conceive of this as an improvement of memory in some extended sense. And it likely would represent an improvement for schizophrenic patients. But I think electrical medio-dorsal thalamic stimulation is better understood as an improvement in writing or rewriting memory/plans through constantly retriggering thoughts about the higher level rules that structure our plans rather than allowing some sensory impression to linger on for a longer period. I would expect thalamic stimulation of this kind to be like continually having the thought “Remember the rule: [x: where x is some rule, in the Kellendonk experimental case ‘go the opposite way’]” inserted into your consciousness. On trials where this is inhibited, as in the green light genetically modified mice, the mouse may just be forgetting to apply the rule when reformulating its plan. Arguably this is more about keeping the cyborg focused on the important rule rather than extending a sensory impression.
And so that is one area where we are currently able to show something like amplified intelligence in a mouse model. But I speculate that rather than changing us into human hard drives with extraordinary precise memory this stimulation of the thalamus may be better understood as promoting our ability to focus on and attend to relevant planning rules.
To be clear: Kellendonk’s lab ran trials using the inhibitory green light probe first on the cortical prefrontal area and then separately on the medio-dorsal thalamus. Both these variations produced similar results.
On Columbia University’s Superintelligent Cyborg Mice
I recently wrote a post explaining how the thalamus is very likely responsible for our string of discrete temporally bound moments of subjective awareness as well as where our brain comes to consensus, making it arguably the seat of consciousness. I was surprised at the relative lack of discussion of this brain structure on LessWrong. During research for that post I came across another surprising omission here: the work of Christoph Kellendonk and his lab at Columbia University’s Irving Medical Center’s psychiatry lab.
While it may be somewhat sensational to say that he is working on superintelligent cyborg mice, it may not be inaccurate. He definitely is working on what folks around here call “intelligence amplification.” But enough preamble.
Kellendonk’s lab has an experiment where they are inserting electrical stimulation probes into the medio-dorsal thalamus of rats which enhances their ability to complete a cognitive task. Kellendonk believes that his cybernetic modifications enhance the rat’s ability to retain a memory for completion of a memory task, but I have a different interpretation. Regardless, there is a definitive improvement in the ability of the rats to complete the task under stimulation.
For those unfamiliar with the thalamus, it is the place all coritco-thalamic loops pass through and in many ways is the heart of the brain’s circulatory system, pumping waves of electrochemical cascades through various parts of the cortex, as I explained last week. Different specific parts of the thalamus are linked architecturally in these thalamo-coritcal loops with different areas of the cortex. The medio-dorsal part of the thalamus links mainly to the prefrontal cortex although there is some linkage to the anterior cingulate cortex.
In his experiments Kellendonk’s lab first showed a link between a prefrontal area and the thalamus evidenced by simultaneous activation patterns; he then uses genetic designer mice which can have their neural connections inhibited with green light to show that a probe inhibiting the prefrontal areas directly or the region they are connected to in the thalamus will cause the mouse to perform worse on a cognitive task.
In particular the mice were sent through a T-maze twice; on the first trial a reward was available on either left or right arm of the T maze, the mouse retrieves it and is trained to return to the start position and wait through a delay; on the second trial they had to choose the arm of the T-maze opposite to the arm where the reward was available on the first trial (or else be trapped in a rewardless arm by a quickly closed gate). The delay between the two trials was an adjustable variable.
The mice were shown to perform worse on this task when the light from the probe was activated, inhibiting the neurons in either the prefrontal cortex or the thalamus which was connected to it [1]. In a separate experiment Kellendonk then demonstrated that direct electrical stimulation to the medio-dorsal thalamus produced the opposite effect: an improvement in performance on the task.
Difficulties with working memory are a common symptom of schizophrenia; Kellendonk’s lab is part of a medical psychiatry center at Columbia studying schizophrenia. He interprets these results as improving the working memory of the mice and deep brain thalamic stimulation of schizophrenic patients is being explored as a cutting edge treatment. But I think I may have a different interpretation.
Rather than working memory encoding in which direction lay the reward in the previous trial I think what may be going on here is the prefrontal areas are encoding the rule “choose the opposite direction,” which is accessed when formulating a plan. I would hypothesize that upon detecting the reward in the first trial a plan is formulated, based on this rule, to go the opposite direction either immediately or upon some reflection. This is stored in working memory somewhere either in some kind of frequently reused system which is flushed of contents for use in a new task frequently or some other way which naturally degrades quickly [2]. I hypothesize that these prefrontal areas are merely accessed when rewriting the plan to working memory but themselves are an element of long term memory.
So what I propose happens is the mouse retrieves a reward down the left pathway and formulates a plan to return to the start and go right next time in working memory. The mouse then returns to start and waits. This plan is compressed and does not record why the choice to go right this time is made, just that the plan is to go right. At some point, depending on wait time, the place storing the working memory of what to do is wiped likely because the mouse started thinking about something else and needed that bit of scratchpad [2]. Then the mouse rewrites the simple plan after retrieving the memory of which path it took on the first trial, probably from elsewhere in the brain with more connections to sensory and motor areas rather than the prefrontal areas, and combines that with this prefrontal learned rule to reformulate the simple plan.
And I think this fits better with the general functions of the prefrontal cortex than assuming the prefrontal cortex is involved with retaining sensory experiences and recording motor output; these are functions more likely associated with the occipital, temporal, or parietal lobes. The prefrontal cortex contains the most abstract neural architecture for high level decision making and stores things like patterns of plans or understandings of things like procedures or laws: i.e. “rules.”
I suppose this doesn’t fully contradict Kellendonk’s findings. You can conceive of this as an improvement of memory in some extended sense. And it likely would represent an improvement for schizophrenic patients. But I think electrical medio-dorsal thalamic stimulation is better understood as an improvement in writing or rewriting memory/plans through constantly retriggering thoughts about the higher level rules that structure our plans rather than allowing some sensory impression to linger on for a longer period. I would expect thalamic stimulation of this kind to be like continually having the thought “Remember the rule: [x: where x is some rule, in the Kellendonk experimental case ‘go the opposite way’]” inserted into your consciousness. On trials where this is inhibited, as in the green light genetically modified mice, the mouse may just be forgetting to apply the rule when reformulating its plan. Arguably this is more about keeping the cyborg focused on the important rule rather than extending a sensory impression.
And so that is one area where we are currently able to show something like amplified intelligence in a mouse model. But I speculate that rather than changing us into human hard drives with extraordinary precise memory this stimulation of the thalamus may be better understood as promoting our ability to focus on and attend to relevant planning rules.
To be clear: Kellendonk’s lab ran trials using the inhibitory green light probe first on the cortical prefrontal area and then separately on the medio-dorsal thalamus. Both these variations produced similar results.
This is probably the hippocampus or possibly a circuit in the cerebrum.