Abstract: Visual perception is closely associated with consciousness, and an important part of visual perception is the ability to infer the properties of objects from time-varying changes in retinal stimulation. Two key properties that are of obvious behavioral relevance are their identity and their velocity relative to the observer. In this presentation I will review old results that show how certain kinds of anesthetics impair the ability of individual neurons in the visual cortex to extract important physical quantities from retinal input. I will then present more recent work that provides a computational account of how these neurons integrate their inputs so as to become selective for important stimulus properties. Finally, I will show how the same computations appear to be at work in visual cortical regions that are responsible for different functions, including the estimation of object identity and velocity.
Mineault, P.J., Khawaja, F.A., Butts, D.A., and Pack, C.C. (2012) Hierarchical processing of complex motion along the primate dorsal visual pathway. Proceedings of the National Academy of Sciences of the USA, 109, E972-980. http://packlab.mcgill.ca/mineaultetal2012.pdf
Pack, C.C., Berezovskii, V.K., and Born, R.T. (2001) Dynamic properties of neurons in cortical area MT in alert and anaesthetized macaque monkeys. Nature, 414, 905-908.http://packlab.mcgill.ca/packetal2001.pdf
Comments invited
Dear Professor Pack,
ReplyDeleteAnesthesia vs. waking consciousness is obviously an important source of empirical constraints on our understanding of consciousness. Having said that, we have various natural and artificial sources of unconsciousness (as a state), including different anesthetics. Some anesthetics could conceivably have quite "unnatural" effects on neuronal functioning by minor toxic effects on cell metabolism or membrane properties, for example. You can stop a car by putting sugar in the tank, but sugar does not give a lot of insight into the normal running of the engine.
I would like to raise some questions.
1. How do different global anesthetics affect visual neurons of the kind you studied?
2. How do natural unconscious states, like slow wave sleep, influence those neurons?
3. Some recent work by Massimini and Tononi seems to show that the peak of the slow delta wave may show waking-like activity during SWS, but only for fractions of a second, while the trough blocks fast signaling in cortex. Is there any difference in the neurons you studied between the UP and the DOWN phase of the delta wave?
4. During normal waking a number of methods have been used to inhibit conscious, but not unconscious visual processes. Including binocular rivalry, backward masking, attentional blink, and more grossly, TMS, etc.
Do your results generally converge on the findings from those kinds of studies?
Thanks very much. I'm trying to make sense of a growing mountain of evidence, and if you have any suggestions for unifying principles that explain a lot of those phenomena that would be very valuable.
Sincerely,
Bernard Baars
Dr. Baars-
DeleteUnfortunately the anesthesia result that I mentioned in my talk was something that we haven't followed up on directly. But I'll try to answer your questions:
1) The results I showed involved isoflurane. We did repeat the experiment with sufentanil, which has become the standard for anesthetized primate experiments. The results looked very much like those in the alert animals, which is consistent with my understanding that opioids generally have much weaker hypnotic properties than inhalational anesthetics like isoflurane.
2) I have no direct experience in this area. There are a few reports (e.g., Livingstone & Hubel 1981, Worgotter et al., Nature, 1998) on the effects of different sleep states on visual receptive fields. I think these could be consistent with our findings if there are state-dependent effects on the nonlinearities that we have found to be crucial for stimulus integration in the extrastriate cortex. But I haven't pursued this rigorously.
3) Unfortunately we didn't record the EEGs in our old experiments! We are now spending a lot of time looking at LFP-spike relationships in the alert animal in my lab. We should have results on this soon.
4) These are excellent ideas, but we haven't pursued them. The one thing we did explore in Rick Born's lab was reversible (cooling) inactivation of MST feedback while recording from MT. Unfortunately nothing interesting happened, but the experimental design may have been suboptimal.
I wish I could provide more insight into the interesting points you raised, but I'm not much of an expert on anesthesia or consciousness. My overall sense is that the effects of isoflurane in our experiment could indeed be like sugar in your analogy -- they could affect low-level mechanisms that are not intimately associated with consciousness per se. In fact the entire output of the extrastriate cortex is in my view something that consciousness, memory, behavior, etc. can act upon, as it provides a very compact and robust code for interpreting one's surroundings, but that it's not necessarily a great way to understand consciousness. But I'm happy to be convinced otherwise!
Best wishes,
Chris
Because the sophisticated motion stimuli used in Dr. Pack's experiments were presented in a continuous motion movie (the one that he showed using youtube), I would have guessed that the neuronal firing in MST would reflect some sort of re-entrant/feedback processing. For example, does the neuron fire in a different way depending on which stimulus was presented to the monkey immediately before the current stimulus? In other words, do contents of the monkey's short term/working memory ever get featured in the firing rate? Or do you have to configure the task to be specifically about memory in order to have such effects? Are their direct connections to MST from the hippocampus (this is just for my own curiosity as a hippocamp-ologist!).
ReplyDeleteHello -
DeleteI think it's quite likely that the MST responses reflect some feedback inputs -- the anatomy tells that they exist in abundance, so they likely contribute something. However, we don't know what that something is, and so it's difficult to incorporate them into our modeling. We can say from our data that stimulus history effects in MST are minimal (see also Paolini et al., 2000). This likely reflects in part the fact that the monkey had no incentive to think about or attend to the stimulus. I don't recall whether there are direct MST-hippocampus connections, but the answer can probably be found in the old work by Boussaoud, Desimone and Ungerleider from 1992.
Best wishes,
Chris Pack
Can someone clarify on the point that the core mechanism of the model has been used in other research to model insects. Model what in insects exactly, and how is that relevant to the validity of the model?
ReplyDeleteMartha -
DeleteThe locust has a well-studied type of neuron for estimating the time to contact of looming stimuli. This is something that MST does for the primate. Some of the mechanisms in the locust model are very much like our MST model (see for example Hatsopoulos et al., Science, 1995). This suggests either a common evolutionary origin or a convergent solution to a similar problem. I actually think we can understand a lot about primate vision by studying bugs.
-Chris
Thanks for the reply :)
DeleteDr. Pack claimed that his model, which used non-linear activation functions to explain how MST reacted to signals from MT, was able to account for about 50% of the data. Maybe I didn't pay enough attention (or maybe I don't know enough about neuroscience), but how did he exactly measure that proportion? What was the measuring rod? I assume he trained artificial neural networks to develop these specific pattern-recognition neurons, but how was the behavior of those networks assessed?
ReplyDeleteAlexandre -
DeleteThe 50% figure is just a simple cross-correlation between the predicted output of the model for each stimulus and the measured response to that stimulus. If you're familiar with regression analysis, it's the standard metric used to assess the quality of the fit.
-Chris
An incredibly impressive approach to model a high-level function - one we perform so effortlessly yet were until now at a complete loss to explain.
ReplyDeleteI want to echo, however, the point made in discussion... Maybe it is just my math-incompetence, but is it presumptuous to say that our brian 'does' math? To what extent can we model the highly stereotyped and probabalistically-weighted functions our higher-order visual neurons perform (that is: multi-synaptic integration and computation) with functions? Isn't this a bit like saying an artichoke 'computes' its own fibonacci numbers during its process of leaf budding?
The literal answer is that we can, up until Dr. Pack's study, model MST function 50% of the time. I can't help be left wondering whether this other 50% represents MST computations done via other transformative models, or whether the entire MT-MST relay occurs via entirely different mechanisms.
- WakeupSheeple 1111
Dr. Sheeple -
DeleteI wouldn't say that the brain does math (except when you actually do have to solve a math problem), but just that you can use certain mathematical approaches to model brain functions. Your artichoke example is spot on.
The question for empirical science is whether one's model maps in a useful way onto entities that can be studied experimentally. That's why we use realistic subunits modeled after well-studied neurons as the entities in our models. The outcome of the effort is then a set of predictions that can be tested in the next round of experiments. If it turns out that we were wrong in the initial formulation, so be it -- as you say, the other 50% of the variance in the data could be where all the important mechanisms are. But we can only test models that we have thought of, and we naturally want to test those that are at least plausible based on the data at hand.
-Chris
This presentation was very interesting! So it seems to point to the fact that conscious states are related to integration of information. How does this evidence relate to the integration consensus reported by Morsella? Are we talking about the same kind of integration here? If I understood correctly the integration consensus was more about a multimodal / multi-region information integration.
ReplyDeleteI'm afraid I missed Dr. Morsella's talk, but would be interested to hear more about it.
DeleteIt would also be interesting to compare Dr. Pack's view with views with theories of consciousness where synchrony of the signal is key!
DeleteAs presented earlier this week by Prof. Freedman, the MT region, which encodes information in a continuous way, communicates with the LIP region, which encodes information in an abstract binary (categorical) way.
ReplyDeleteProf. Park presented a model in which MT (and MST) neurons' activations were predicted by applying non-linear (compressive) parameters to subunits' activations.
I wonder if this non-linear best fitting model reflects (in part) the dynamics between the LIP and the MT. Indeed, LIP reflects a higher-order organization, while non-linear dynamics offers an account of the emergence of new structure in complex open systems, which are predicted by an increase in entropy, followed by a decrease in entropy.
Thus, categorization might play a role in explaining this non-linear model.
Moreover, I notice that categorization has been little addressed through the summer school. I would have thought that given even though we receive continuous information we perceive categorical information (although this might be contested), we would have been more inclined to look for consciousness through more empirical work on categorization.
Etienne Dumesnil
Etienne -
DeleteI have to say that I hadn't thought about the link between a compressive nonlinearity and categorical processing, but it's an interesting idea. The outputs of the subunits in our model are close to binary in many cases, so they might be thought of as a sort of automatic category processing. The obvious difference with Dr. Freedman's work is that the categories in that case were imposed by the experimenter, whereas in our case they would represent a task-independent way of warping the processing of the stimulus itself. I don't know of any evidence for or against the involvement of LIP in MST stimulus selectivity, though it is in principle testable with our existing data if you can formulate it as a mathematical statement.
-Chris
I found Dr. Pack's talk very entertaining, especially the TEO cell which responds specifically to O.J. Simpson! I was wondering if Dr. Pack saw any changes to the firing pattern in the MST over time. Were fewer neurons recruited with the same stimulus over time?
ReplyDelete