My LOL for the day. Not disagreeing at all, just struck me funny. The sad fact is that we toiled in Machine Intelligence for years and barely had anything to show for it–then business came along and now we have working intelligent agents that can converse.
1 Like
I’ll try to summarize why I don’t think cortex is passive. Everything else I said doesn’t matter much.
- The motor output (L5tt) anchors locations.
A. The apical dendrites enhance perceptual detection, so a weaker stimulus will be noticed [1].
B. Upon active touch, motor copy signals sent to the apical dendrites cause firing strongly correlated with location in a coordinate-transformed spatial system [2, 3, also see 4].
Those are both consistent with L5tt cells sending signals which anchor locations being tracked by other neurons. - The same cells trigger anchoring and perceptual detection by apical firing modes. So anchoring is equivalent to perceptual detection.
- So that aspect of attention happens by anchoring, not by thalamic filtering. Also, anchoring is a crucial part of perception, so I think it has to happen automatically, e.g. without reward.
- So the cortex can cause things which volition can control (at least attentional shifts), so I don’t think the cortex is entirely passive.
I also argued for motor commands driven by purely perceptual processes. Remember, the cells I’ve been talking about are the cortical motor output. They’re absolutely involved in perception (for many reasons), so I expect that they produce perceptually-driven, non-motivated behaviors. Motivations can constrain / guide those behaviors and attention, of course.
I think that’s just filtering out sounds, not frequencies. Thalamic gain control is probably a thing, though. I guess my point was just that it feels like attending a visual object means you’re filtering out the rest, but it’s actually about anchoring egocentric locations to the object.
I think generic cortex is a thing. That doesn’t mean we experience vision the same way we experience touch. They’re different senses. We perceive things on the retina from a distance, so we need to deal with retinotopic things, like occlusion / line of sight. So it feels like it’s fundamentally retinotopic, but if we assume generic cortex, it’s not, because we don’t experience other senses topographically.
So attending objects by topographical filtering wouldn’t work well for generic cortex. So attending objects requires another mechanism, which I think is anchoring.
A link between anchoring and attention isn’t a new idea. I think the references thing brought up in some Numenta meetings is related to that, but I don’t really understand that so dunno.
To clarify, I wasn’t talking about motor control, just describing anchoring and its link to perceptual detection.
It probably doesn’t just modify behavior originating subcortically. vS1 has a direct pathway to a muscle [5]. That study did polysynaptic retrograde tracing, pretty cool.
Even if it just modifies ongoing behavior, the cortex would still have that parallel activations issue.
2 Likes
Passive: the proposition as it seems to me is that the cortex:
- receives a flow of pre-processed sensory input (from sub-cortex)
- performs a range of functions including anomaly detection, developing goals, history and prediction, etc
- sends a flow of high-level instructions (to sub-cortext) for execution as motor commands, new goals, etc.
- sub-cortex can and will do everything cortex does except for that modest range of extremely valuable ‘higher’ functions (‘AGI’).
Is that the proposition you argue against?
[As well cortex presumably initiates consciousness, inner voice.]
My concerns with this post:
- introspection is not science. Hypotheses must be tested against experiment (thus mostly in animals)
- the neuroscience quoted IMHO does not support the conclusions drawn from it, and in general cannot
These grounds are insufficient to reject the proposition.
1 Like
I think we need to be careful to keep things in perspective. On this point, no, no we don’t. As somebody who’s written a few chatbots, all that we have are very brittle detection models and a few neat tricks.
Not to say I don’t appreciate being able to ask some smart device what time it is, the weather, or to set a timer/alarm… those are all neat and useful. But I would shy away from saying that we have anything “intelligent” or with a real ability to “converse” beyond a tightly controlled and narrow dialogue/script. ELIZA is a lot of fun though. 
Eliza, Computer Therapist
1 Like
I apologize. Perhaps it is naïve and aspirational. However, after having written firmware for a couple medical device projects, I see that there certainly is a bureaucratic system in place to gatekeep code which has the potential to directly impact peoples’ health; given that perspective and background, combined with the negative sentiment that many people have gathered about AI, facial recognition, the abuse of personal data for widespread abuse/manipulation via social media, it wouldn’t say yet that it’s beyond the realm of possibility that some system of review is imposed on systems that impact people outside of finance. As a matter of fact, because of the stringent requirements that medical devices are required to prove their code execution, I suspect that is one of the reasons we don’t see any real explosion of “AI” medical devices aside from some ‘smart’ watches which log your activity to the cloud for potential analysis… even still, these generally aren’t allow to be officially sold as ‘medical’ devices.
As for those stupid institutional high speed trading bots… I think you’re right. They’d need to cause another global financial catastrophe before government around the world might consider some restrictions on their activity. Even then, they’d be quietly rolled back a few years later, just as we saw with exotic derivatives trading activity.
2 Likes
I’m only arguing that the cortex can produce attentional shifts as part of sensory processing. We can do that intentionally too, so I don’t consider that a passive thing.
Discussions aren’t science, but they still have value.
Which conclusions and why?
I want to help refine it, not reject it.
I think everyone missed one of the base assertions: input of sensory data via the thalamus. This certainly can initiate activity and part of this activation can be recall, and loops through higher-dimensional search space. Continued inputs could certainly give rise to a continuous response. If you think about it, the cortex is actually passively responding but it could be a very rich response.
As to the volitional part - this is entirely done through the sub-cortex. A loop from the cortex to sub and back again is involved in command exchange. With the relatively small number of subcortical nodes and a large variety of levers to pull on the body, the subcortex can have some pretty powerful built-in patterns in some sort of genetic coding. I seem to recall reading that the large variety of neurotransmitters gives behavior shaping. I have read that we make decisions that we “become aware of” after the fact. “It” has to have a cortical presence to enter consciousness.
I have been chasing this for years and I keep reading paper after paper that is consistent with a very close partnership between all the brain elements. There is programming that takes place outside of cortical computations and this is where common sense and motivations come from.
This is the missing “ghost in the machine.”
With DSP processing it is relatively common to think of processing pipelines When considering what the cortical column is doing the bigger picture of what computations are being supported may shed some light. That has been helpful for me to understand HTM/TBT.
2 Likes
If the cortex can generate behaviors for non-volitional reasons, would you still consider the cortex passive? I’m just trying to understand what you’re saying.
Introspection is not science. Can you demonstrate empirically that this is so?
Discussions lead to hypotheses. How would you prove yours empirically?
Neuroanatomy is not the way to do it. What you’re looking for is a situation where an organism initiates a course of action that (a) does not rely on any current sensory input (b) does rely on a cortical capability such as constructing a timeline. Doesn’t sound too hard.
My perspective on that comment was as a grad student in the early 80’s and my guess is that at that time you were just a twinkle in your dad’s eye.
2 Likes
Do you have a cite for this? I would seriously like to see it, like you I have been ‘chasing this for years’. TIA
I think the discussion of active vs passive is misleading. The cortex is actively doing the things that it does, and passively not doing the things that it doesn’t do. The HTM theory says it is building a model of the world. There are complimentary theories explaining how attention (via apical dendrites) can be used to interrogate/manipulate the cortex’s model of the world.
The cortex is passive with respect to the process of action initiation. That is mostly done by the basal ganglia and brainstem.
An interesting note: The locus coeruleus is part of the brain stem, and it contains norepinephrine (aka the adrenaline rush hormone) releasing neurons. IIRC the release of norepinephrine is critical for starting an action. The following review article (which i have not yet read beyond the abstract) says that the frontal cortex projects directly to the locus coeruleus and controls it in-sync with the animals behavior.
1 Like
I agree introspection isn’t science. I just think hypotheses matter a lot for figuring out how intelligence works, considering the history of AI winters and what I know about neuroscience. That’s just an opinion.
I can’t because I’m not a neuroscientist, but if you’re asking whether it’s provable, “L5tt sends signals to location-representing neurons which anchor locations” is very specific. The second part, “anchoring locations is equivalent to perceptual detection mechanism-wise” is also very specific, because it’s really about whatever the exact mechanism is. I don’t know what it is, but something apical, if it exists. The rest is basically just semantics and why I think cortex isn’t passive.
If you really need exact examples of ways to test the hypothesis
- Do L5tt cells have a synaptic pathway by which they could anchor cortical grid cells?
- Do L5tt cells correlate with locations upon perceptual detection in other cortical regions and species? This doesn’t require figuring out the spatial system because cortical grid cells are a thing. L5tt cells shouldn’t be grid cells, I think, because they’re producing a signal which strongly correlates with location by combining sensory / motor information. That means they should include extraneous information, and better predict the resulting activity of grid cells better than vice-versa.
- Do cortical grid cells no longer anchor their locations upon perceptual detection if the apical dendrites of L5tt cells are silenced?
- What’s the exact mechanism, and does it work when implemented in code? Does the exact mechanism exist in generic cortex?
I never said the organism initiates a course of action which doesn’t rely on any current sensory input.
Maybe the cortex is passive in the sense that it reacts to sensory input (and external inputs from subcortex, whereas subcortex wants stuff and has emotion things like amygdala)?
I was thinking about it in the sense that, we can intentionally do xyz, so the cortex isn’t passive if it can do any of that, because it’s sending something like a command.
2 Likes
Have you considered that the cortex is being ordered to make that command?
1 Like
I guess if by passive you mean lacking its own volition, intention, desires then it would make more sense to everybody.
It might be quite active otherwise by doing … its stuff, which probably can be summarized as … knowing
Well, I’m not sure that’s what you mean, this is actually a question.
1 Like
Not a question - an assertion.
Without a driving input, the cortex does nothing.
I start with that when considering the common criticisms of DL and tangentially the HTM/TBT models in regards to AGI. That is the original thrust of this thread.
In most (all?) of these models, the “animal spirit” that is brought to the table by the subcortex is the secret sauce of the missing element. Volition, desire, attention, common sense. The cortex adds so much more but it is not the driver in control. Dumb boss, smart advisor.
1 Like
To maybe put a different way… Freud may have had the insight to recognise that :
Dumb boss = ID = subcortex
Smart advisor = EGO = cortex
During sleep the smart advisor teaches the dumb boss.
The smart advisor is THE fast learner
Deep learning methods at the moment are the dumb boss.
2 Likes
I guess some people like dumb bosses. I fired mine a lifetime ago.
2 Likes
So the key question IMO is how to put that in the form of a testable hypothesis.
Start with a general lack of sensory input: quiet, nothing happening, not hungry, etc. Then the experimental subject initiates some behaviour: explore, make noise, demand attention, etc. How do you determine whether such unsolicited action is triggered by cortex or sub-cortex?
This should actually be accessible experimentally. There are a variety of methods to localize activity in the skull.
One of the earliest tests of volition is the Libet experiments:
The conclusions are not a slam-dunk:
Adding viewing tools that tie neural correlates to the basic Libet experiment should serve to further refine this experiment and localize decision making.
1 Like