An Existential Crisis in Neuroscience
m.nautil.us
m.nautil.us
I find that as we gain new tools to study the nervous system more specifically, both data and models of how neurons are organized at the circuit level become more important. To advance on an analogy in the article, it's like trying to explore the dynamics of NYC without a map. For instance, it's hard to tell how/why people interact with central park if you don't even know where they live. The more specifically you are able to pin down people, the more it matters where exactly they live to understand.
Granted, the fly is much simpler than humans or even mice, and it will likely take decades and new tools for us to study humans in this way. However, when we get there, mapping out the brain connections will be crucial to make sense of it all.
This made me realize we may one day willing allow human brain controls to get us to do things we don't want - Work, Exercise, etc.
We already have that: coffee.
Similar tricks can be played with the human brain, things we have been able to do for decades, while people are undergoing brain surgery, and now later, with TMS. However, being able to elicit limb movements or bits of speech, or even emotional qualia is different from having a dynamic understanding of the brain in vivo in everyday life.
Certainly having an understanding of detailed circuitry is interesting and important, but to me there's a forest for the trees problem.
I really hope we can realize how the whole thing works by looking at its parts. But I doubt it will bring the breakthrough. On the other hand maybe there is this one mechanism that we have to discover to make sense of all the parts. Then those efforts will form the groundwork for an explosive understanding.
My impression is that there's a lot of very oversimplified assumptions being made all the time in these fields that get glossed over in very arrogant (or naive?) ways. It's really astonishing to me, not just because of how oversimplified the assumptions are but because researchers are then surprised things don't work out.
To be fair, this is true of other fields as well. I'm more familiar with molecular genetics and genomics, and the same things happen there. There seems to be a certain hubris that goes unquestioned, and it always amazes me, the sci-fi fantasy narrative being accepted as fact.
Just to take one thing for example: there's huge anatomical differences between people's brains even at the macroscopic level, that just get glossed over in discussion. Those fMRI images you see? They're often done by aligning different images to a common map, just assuming individuals' brains are carbon copies of one another. Now you're going to try to delineate a connectome at the neural level, as if there is one connectome at that level?
When will everyone learn? Where's the public skepticism?
Neuroscience usually focuses on precision details, but doesn't aim to tell big picture stories. There are a few exceptions, however, like Karl Friston's free energy reduction model.
Proven to be wrong when you watch people with serious brain injuries re learn skills, but shown to have value by how it can predict what tumors or injuries will do to someone's ability.
The ancients knew a fair bit.
We need a story that explains how the brain uses rhythms (oscillations) for computation.
What would be really helpful is a model that can add to things we already know. Saying "studying for an exam engages the X part of the brain and uses the Y neurotransmitter" adds literally nothing to your understanding of studying (you can find out much more about studying by talking to people that are good at it and who have done a lot of studying themselves), it's just taking an everyday activity and identifying the small but still vastly complex portion of the brain that is activated more than others. Imagine being told that a particular bug in a 1-billion-lines-of-code codebase is due to some code within a 10-million-lines-of-code portion of it: that's great but how helpful is it really?
https://www.oxfordscholarship.com/mobile/view/10.1093/acprof...
https://aaai.org/ojs/index.php/aimagazine/article/view/2744
"A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics"
Daniel Coleman’s “Emotional Intelligence” offers a useful daily model for reasoning around it.
They exist but realize they’re going to “feel” different than physics models. With general physics knowledge one can literally implement tests and build on that independently.
With NS, one can read a model but without imaging machines and chemical testing... shrug... it’s harder to build a muscle memory.
Which neuroscience research argues is super important to connecting details into a composite one can intuit around competently.
Reading AND writing are important to learning English. Same with everything else.
TLDR the practical value of NS is already known: practice learning as we do, don’t be a dick.
The important question to ask is: is the deep learning abstraction any good?
There's a very strong case to be made that the answer is yes: deep learning systems can perform many (of course, not all, at least not yet) tasks that involve perception (computer vision/speech recognition), motor control (the recent openai robot), language understanding (machine translation/BERT/GPT), planning (alphago/dota/the deepmind protein folding), and even some symbolic reasoning (the recent work from facebook on symbolic integration https://ai.facebook.com/blog/using-neural-networks-to-solve-...). Some of these tasks are performed at such a high level that they become commercially useful, and in some cases, surpass "human level".
So here we have a "model family" -- deep learning -- with a set of principles so simple that it can be studied with intense mathematical rigor (for example, https://arxiv.org/pdf/1904.11955.pdf or https://papers.nips.cc/paper/9030-which-algorithmic-choices-...), and that produces many of the behaviors we want out of brains (and not just behavioral: see, e.g., https://arxiv.org/abs/1805.10734: " Interestingly, recent work has shown that deep convolutional neural networks (CNNs) trained on large-scale image recognition tasks can serve as strikingly good models for predicting the responses of neurons in visual cortex to visual stimuli, suggesting that analogies between artificial and biological neural networks may be more than superficial." -- this is just one of many papers that show that even under the hood, trained deep learning systems exhibit many properties of biological neural networks).
These reasons strongly suggest (imho) that deep learning is in fact the newtonian theory of neuroscience. More strongly, no other theory comes remotely close in its simplicity and explanatory power.
Self driving cars can't leave an enclosed environment and might never do so safely.
Richard Dawkins spoke very highly of the brains ability to do some kind of natural calculus for the sake of tracking a ball in flight, but most animals run on simple tricks and reference points.
Deep learning might be the "good think" for the next ten years, some of us are not going to let go of the transcendent truth that the brain is not defined by what we think it is. I see limited reason to see deep learning as more likely than some emergent behaviour from a vast number of simple rules. Like animals flocking together in a boid sim.
The problem is for all this power people still play chess,go and starcraft and we don't know how their brain works.
Is this the same mistake as in "The Relativity of Wrong" [1]?
> people have thought they understood the Universe at last, and in every century they were proven to be wrong. It follows that the one thing we can say about out modern "knowledge" is that it is wrong.
> [...]
> My answer to him was, "John, when people thought the Earth was flat, they were wrong. When people thought the Earth was spherical, they were wrong. But if you think that thinking the Earth is spherical is just as wrong as thinking the Earth is flat, then your view is wronger than both of them put together."
Modelling the brain as a bunch of pistons or as a complicated machine or clockwork thing is a lot better than as a magical clay golem or opaque soul. Modelling it as a computer is even better than that. Not a computer in the sense of an x86 desktop exactly, of course, but the concept of computation is clearly fundamental to understanding the system. Similarly, the brain is not ResNet but concepts like backpropagation are probably useful.
So, sure, maybe people have been using the latest fad to explain the brain forever. But that's only bad to the extent that the latest fad is getting further away instead of closer.
1: http://hermiene.net/essays-trans/relativity_of_wrong.html
Ancients used to think that thinking happened in the gut and recently the microbiome pathway for describing thought has re-emerged. Both the gut pathway and the computational stream could be wrong.
The outcome of seeing the brain as a computational device will run out of juice like revelation has.
The approach was focused on presymbolic processing -- and tried to optimize harmony. Harmony was, interestingly, the first mathematical model of the mind (by Pythagoreans/platonists in ancient Greece). It has a lot going for it these days, too, to understand oscillatory coupling in neural circuits. I learned recently that brain waves are harmonics (frequency doublings), which somehow I missed before!
Consonance results in greater periodicity, meaning that the action potentials are more likely to line up, whereas dissonance has less periodicity, so action potentials don't line up. It feels better (there is pleasure) when the action potentials align because of hebbian reinforcement (synchronous firing). This assumes that reinforcement would be pleasurable, but pleasure is the main reinforcer at a cognitive level.
This seems like an odd take on Newton to me. What made his contributions important is that they were correct up to the precision we could measure for centuries.
We are nowhere near that for a subject like neuroscience.
There have been models adopted by scientests who at the time knew they were wrong and incomplete. For example ancient astronomy or medicine or logic. but their adherents tended to hold back science when new discoveries were made. So they are double edged swords.
Ok, maybe not quite to that level. But consider this paper, by great neuroscientists and cited over 1000 times, that models brain wave bands as harmonics, with band widths as the golden mean. [1, see figure 4]
My money is on some synthesis of the many theories in oscillatory neurodynamics. Neural resonance, dissonance, harmonics and entrainment... So many of the theories* have borne out empirically, but there has hardly been an attempt at synthesis.
* Theories like "Communication through coherence", "binding by synchrony", "phase amplitude coupling" + "working memory", etc etc etc.
[1] Klimesch, W. (2012). Alpha-band oscillations, attention, and controlled access to stored information. Trends in cognitive sciences, 16(12), 606-617.
What would you recommennd?
More broadly, there is a crisis. Our statistical methods/understandings are not working out in these large N-dimensional data sets, at least for the researchers that were raised on excel and not numpy.
Aside: I'm surprised that the FAANGs haven't revolutionized statistics yet. When you have 'phase changes' where a LOT more data becomes available, you get to see very low probability events. It's happened in psych, in bio, in physics most famously, in politics, in economics, etc. We have a LOT more data now in for statistical use, but it's still just Poisson distributions and t-tests. What gives?
What does that even mean?
Similarly, with all this 'big data', I would have guessed that we would have come up with something kinda like that. Finally, we have all this stuff coming in. It's all fairly well captured, fairly well correlated, and fairly valuable so you can pay people to just screw around in the labs and see what drops out.
It's not like when we got a lot of experience gathering physical data we suddenly realized we were doing integrals and derivatives wrong.
Math is a brach of logic, sure, but the computer has acted like a microscope for math already. We can more easily check our conjectures and see if they hold true. Also, we've been able to communicate better, so that has helped out math as well. Even though math is 'above' us little humans, the day-to-day work of mathematicians has been helped out a lot, even in the small ways of not having to lug your mortal-coil through the library stacks.
So, with all this big data, I'd have guessed that we could have seen more edge cases with statistics. Little areas where someone said 'huh that's funny'.
But my experience in software companies has been that in a corporation you are a slave to convention. Because even if you innovate in your domain, that innovation will be invisible to the rest of the corporation. Because when they look at it, they will evaluate it by asking which conventions it follows. But inventions by definition don’t follow conventions. What people like is to see conventions followed, and the results of that. It reads as progress.
Now I think it’s primarily artists who invent. But it’s hard to much production as a solo artist. So those who succeed are artists who split their time between doing their work in secret and public ally playing the game of performing conventions in a corporate body. If you split your time that way, you will be praised for applying convention and then periodically your secret art will appear as a magic trick and delight people who are already happy with you.
Without the convention theater, the art will just strike them as confirmation that you’re out touch with the organization.
The same is true in academia really. It’s just a different set of convention theater, around publishing rather than sales.
Namely that the field seems to be doubling down on increasingly hard computational problems, with deminishing returns, rather than generating wholly new avenues of exploration or insights.
Of course, I've no idea if this is a valid complaint about either neuroscience or theoretical physics.
The point of the article is discussing how mapping the full human connectome is only going to be a small next step towards understanding what's actually going on. That doesn't mean it's not worth doing.
[1] https://journals.plos.org/ploscompbiol/article?id=10.1371/jo...
But there is such elegance in "rudimentary" DNN's giving us the ability to at all assemble this stuff.
Exemplifies:
- Data is not information. - Information is not knowledge. - Knowledge is not understanding.
We've not even left the gate of the first tier. Both exciting and intimidating, but mostly humbling. Or should be.
I think the best story for understanding circuitry in vertebrates comes from the work on hierarchical pattern generators -- where much of the work was done on lampreys.
Grillner, S. (2006). Biological pattern generation: the cellular and computational logic of networks in motion. Neuron, 52(5), 751-766.
As if there was something other than the laws of physics preventing natural selection from testing smaller structures.
Or that there’s something (other than the demands of the computation itself) constraining the architectures that were tested through natural selection.
Likewise manmade models based in mathematics and statistics have long proved more accurate in predicting outcomes than the human mind, even though we know the mind doesn't employ math.
Human-made machines have made it possible for elephants to fly. Nature never will.
There is. Anything that can't be reached by a small number of genetically small steps that either enhance or at least maintain fitness will never be reached.
As a scientist to say something you don’t fully understand as impossible really pisses me off.
That said, in the few years of experience I've had with ANNs, they really do seem like an intuitive analog for human learning, at a high level. And thinking about training problems in this way, approximately "what kind of training data would I need to train a small child to do this" can be more helpful than one might have otherwise expected.
I think we're just a handful of major breakthroughs away from true AI, assuming compute and memory continue to scale. Certainly within 100 years.