Yehouda Harpaz
yh@maldoo.com
last updated 23 Oct 2022
related texts 

Progress in understanding of how the brain works

Most of the discussion in this site of current research are highly critical (Errors, Myths). This page is intended to highlight what looks to me like progress.

1. The brain is always active

1.1) In my model I wrote as one of the major hypotheses that the System (i.e. the brain) is always active (Major Hypothseis 6, 4.5.3). I also wrote that it looks to me obvious (6.3.2.6).

1.2) At least 99% of published articles in the area are effectively based on the assumption that the activity of the brain when it does not explicitly do something is of no interest. However, this seems to change. For example, in this article (Spontaneous Activity Associated with Primary Visual Cortex: A Resting-State fMRI Study, Wang et al, Cerebral Cortex Advance Access published online on June 29, 2007 ; Full text here) they look at activity of the resting brain, and say that "This confirmation supports the perspective that brain is a system intrinsically operating on its own, and sensory information interacts with rather than determines the operation of the system."

1.3) A much stronger support to their conclusion is the observation that the activity of the "resting brain" is far larger than the changes that are normally reported, which they discuss in the paragraph following the abstract. As the references that they mention (one from 1955) show, that is an old fact. What is progress is the fact that they are actually looking at a resting brain, and interpret it as ".. operating on its own..".

1.4) They (and apparently the papers by Raichle that they quote) are still worried of "going too far". They say "Therefore, as suggested by Raichle and colleagues, in terms of overall brain functions, the ongoing intrinsic activity within various brain systems may be at least as important as the activity evoked by external stimuli (Raichle and Gusnard 2005; Raichle and Mintun 2006)." Thus the intrinsic activity is only ".. at least as important..", rather than the obvious "much more important", so we still have some distance to go. But there is a progress in the right direction.

2. Do neurons "represent" anything?

In this article (Churchland MM, Shenoy KV (2007) Temporal complexity and heterogeneity of single-neuron activity in premotor and motor cortex. Journal of Neurophysiology. 97:4235-4257doi ), they find that the neurons in the premotor and motor cortex show complex and heterogenous activity. From this, they suggest the possibility that these neurons don't represent anything.

As discussed here, neurons don't represent anything, but many cognitive scientists and neurosceintists seem to be unable to comprehend this possibility. The authors of the article above are clrearly capable to comprehending it. They present alternative view, but seem to regard their observations as showing that the neurons that they look at do not represent anything. They present the same view in other articles (list of publications, for example the one about "Reference frames for reach planning in macaque dorsal premotor cortex").

The fact that they positively think that the neurons do not represent is progress. The progress is limited, however, because it is still based on the assumption that showing correlation between neural activity and something else (behaviour, stimulus) shows representation. Therefore "representationalists" can still believe that these neurons represent something else.

3. looking at ensembles of neurons.

In this article (Natural stimuli evoke dynamic sequences of states in sensory cortical ensembles, Jones et al, PNAS | November 20, 2007 | vol. 104 | no. 47 | 18772-18777 (open accesss article)), they analyze ensembles of neurons, and show that the ensemble response correlates much better with the input than single-neuron analysis. They stress that the single-neuron analysis that is normally used loses information.

Their data is about taste in rats, and it is not obvious how it is going to generalize to other senses. The number of neurons in each "ensemble" is also pretty small (10). But it is encouraging to see researchers that look at ensembles, and explicitly state that single-neuron analysis loses information. They explicitly state that the coherent state sequences that they see do not represent sensory codes, which is also progress.

4. "The Brain Activity Map Project and the Challenge of Functional Connectomics"

That is the title of this article (doi) (A. Paul Alivisatos, Miyoung Chun, George M. Church, Ralph J. Greenspan, Michael L. Roukes, Rafael Yuste; Neuron - 21 June 2012 (Vol. 74, Issue 6, pp. 970-974)). The main point about this is that they recognize that you need to look at the network rather than individual neurons to understand what it does. For example, they start the summmary by saying: "The function of neural circuits is an emergent property that arises from the coordinated activity of large numbers of neurons."

They criticize current studies, and for example say (end of first paragraph of "Emergent Properties of Brain Circuits"):

However, neural circuits can involve millions of neurons, so it is probable that neuronal ensembles operate at a multineuronal level of organization, one that will be invisible from single neuron recordings, just as it would be pointless to view an HDTV program by looking just at one or a few pixels on a screen.

As to their suggestion, I think they are over-optimistic about our ability to measure activity in living networks. I suspect that recognizing this limit is the reason that they and other reasearchers did not advocate network research like that until now, and perfomed and supported single-neuron research even when it is clear that it is useless (in complex animals, that means almost always), and that they would not feel able to say it without the optimistic predictions.

Because of the technical issues, it is not obvious that their suggestion is the best approach. But thinking about the actual real networks rather than about single neurons or artifical networks is a large step forward.

4. "The importance of mixed selectivity in complex cognitive tasks"

See here. Looks like real progress.

5. "Small sample sizes reduce the replicability of task-based fMRI studies"

See here. Looks like real progress.

5. "In vitro neurons learn and exhibit sentience when embodied in a simulated game-world"

Here they they plate around millions neurosn on a system with many electrodes for stimualtion and some reading channels. They then teach this system to play Pong. They show that the teaching improved the performance of it in playing Pong. It should be noted that this research is done mainly by a commercial entity (Cortical Labs).

"Playing Pong" here is a very simple repsonse. It means just increasing descreasing the activity in some areas of the system, in response to the input about the state of the game. The teaching is done by giving the system random feedback when it fails, and fixed (i.e. the same each time) input when it succeed. They call the fixed input "predictable", but since their system clearly doesn't predict anything, the predictability is irrelevant.

The improvement in performance is quite unimpresive (see their figure 5), but it does look significant. It is diffcult to say how real it is.

The main progress in this article is probably the demonstartion of the technical fit of working with large number of neurons over a long period of time.

On the theoretical side, it is a large progress that they are trying to do "computation" with neurons themselves, rather than using models of neurons. As a result, they cannot use any unrealistic feature, which is on its own a significant progress.

The feedback thay they give (random for failure vs. fixed for success) can be compared to the "Cognitive Positive Outcome" which I hypothesized as the driver for learning (See Major Hypothesis 12 in my modle). The negative outcome in my model is speeding up the random activation of micornodes, while in this paper it is random input. That is essentailly the same.

The positive output on my model is simply slowing down random activity, while in this paper they give some fixed feedback. That is not obviously the same, but if we assume that the fixed feedback cause their system to to tend to stay in a specifc place in the "activity space" all the time (not only when it is on), the fixed feedback will tend to not cause their system large disturbance as the random input, so in this sense it is similar to my idea. They didn't actually check if the feedback has a large or small effect on the activity in the sytsem, which would have given us a hint if this assumption is correct.

[26 Oct 2022] I sent an email to the first author asking whether they measured the effect of the freedback on the activity, or maybe can compute it. He replied that they didn't compute it, and it would be difficult to do it now, but it is something they may do in the future. Apparently they haven't thought about it.

That is on its own a quite interesting observation, because the feedback affects the behaviour of the system by the way it changes the activity of it. So it looks obvious that it is something that it would be useful to measure, but apparently it is not obvious, and these people haven't considered it.

In their theoretical discussion they try to wrap the results in theoretical ideas of entropy and surprise, which are just rubbish (second and third paragraphs of the Discussion). But they seem to reelize that they don't actually have a proper model.

It is not obvious how much impact this kind of research will have, which will depend on how attractive researchers will find it. On one hand, it shows that you can try to look at computations in actual neurons, shich is attractive. But the other hand, you can do much less this way than you can do with computer models (at least in the near future), which makes it less attractive.