Modeling Cell Division

Is molecular biology ready to use modeling to inform experimental work?

The cell cycle is a holy grail of biology. The first mutants that dissected some of its regulatory apparatus, the CDC mutants of Saccharomyces cerevisiae (yeast), electrified the field and led to a Nobel prize. These were temperature sensitive mutants, making only small changes to the protein sequence that rendered that protein inactive at high temperature (thus inducing a cell cycle arrest phenotype), while allowing wild-type growth at normal temperatures. In the fifty years since, a great deal of the circuitry has been worked out, with the result that it is now possible, as a recent paper describes, to make a detailed mathematical model of the process that claims to be useful in the sense of explaining existing findings in a unified model and making predictions of places to look for additional actors.

At the center of this regulatory scheme are transcription activators, SBF/MBF, that are partly controlled by, and in turn control the synthesis of, a series of cyclins. Cyclins are proteins that were observed (another Nobel prize) to have striking variations in abundance during the cell cycle. There are characteristic cyclins for each phase of the cell cycle, which goes from G1, a resting phase, to S, which is DNA replication, to G2, a second resting phase, and then M, which is mitosis, which brings us back to G1. Cyclins work by binding to a central protein kinase, Cdc28, which, as regulated by each distinct cyclin, phosphorylates and thus regulates distinct sets of target proteins. The key decision a cell has to make is whether to commit to DNA replication, i.e. S phase. No cell wants to run out of energy during this process, so its size and metabolic state needs to be carefully monitored. That is done by Cyclin 3 (Cln3), Whi5, and Bck2, which each influence whether the SBF/MBF regulators are active. 

Some highly simplified elements of the yeast cell cycle. Cyclins (Cln and Clb) are regulators of a central protein kinase, Cdc28, that direct it to regulate appropriate targets at each stage of the cell cycle. Cyclins themselves are regulated by transcriptional control (here, the activators SBF and MBF), and then destroyed at appropriate times by proteolysis, rendering them abundant only at specific times during the cell cycle. Focusing on the "START" process that starts the process from rest (G1 phase) to new bud formation and DNA replication (S phase), Cln3 and Bck2 respond to upstream nutritional and size cues, and each activate the SBF/MBF transcription activator.

As outlined in the figure above, Cyclin 3 is the G1 cyclin, which, in complex with Cdc28 phosphorylates Whi5, turning it off. Whi5 is an inhibitor that binds to SBF/MBF, so the Cyclin 3 activation turns these regulators on, and thus starts off the cell cycle under the proper conditions. Incidentally, the mammalian version of Whi5, Rb (for retinoblastoma), is a notorious oncogene, that, when mutated, releases cells from regulatory control over cell division. SBF and MBF bind to genes for the next series of cyclins, Cln1, Cln2, Clb5, Clb6. The first two are further G1 cyclins that orchestrate the end of G1. They induce phosphorylation and inactivation of Sic1 and Cdc6, which are inhibitors of Clb5 and Clb6. These latter two are then the initiators of S phase and DNA replication. Meanwhile, Cln3 stays around till M phase, but is then degraded in definitive fashion by the proteases that end M phase. Starvation conditions lead to rapid degradation of Cln3 at all times, and thus to no chance of starting a new cell cycle.

Charts of the abundance of some cyclins through the cell cycle. Each one has its time to shine, after which it is ubiquitinated and sent off to the recycling center / proteasome.

Bck2 is another activator of SBF/MBF that is unrelated to the Cln3/Whi5 system, but also integrates cell size and metabolic status information. Null mutants of Cln3 (or Bck2) are viable, if altered in cell cycle, while double null mutants of Cln3 and Bck2 are dead, indicating that these regulators are each important, in a complementary way, in cell cycle control. Given that little is known about Bck2, the modelers in this paper assume various properties and hope for the best down the line, predicting that cell size (at the key transition to S phase) is more affected in the Cln3 null mutant than in the Bck2 null mutant, since in the former, excess active Whi5 soaks up most of the available SBF/MBF, and requiring extra-high and active levels of Bck2 to overcome this barrier and activate the G1 cyclins and other genes.

The modelers are working from the accumulated, mostly genetic data, and in turn validate their models against the same genetic data, plus a few extra mutants they or others have made. The models are mathematical representations of how each node (i.e protein, or gene) in the system responds to the others, but since there are a multitude of unknowns, (such as what really regulates Bck2 from upstream, to cite just one example), the system is not really able to make predictions, but rather fine-tunes/reconciles what knowledge there is, and, at best, points to gaps in knowledge. It is a bit like AI, which magically recombines and regurgitates material from a vast corpus based on piece-wise cues, but is not going to find new data, other than through its notorious hallucinations.

For example, a new paper came out after this modeling, which finds that Cln3 affects Cln2 abundance by mechanisms quite apart from its SBF/MBF transcriptional control, and that it regulates cell size in large part at M phase, not through its G1/S gating. All this comes from new experimental work, unanticipated by the modeling. So, in the end, experimental work always trumps modeling, which is a bit different than how things are in, say, physics, where sometimes the modeling can be so strong that it predicts new particles, forces, and other phenomena, to be validated later experimentally. Biology may have its master predictive model in the theory of evolution, but genetics and molecular biology remain much more of an empirical slog through the resulting glorious mess.

• Bitcoin isn't a currency, but rather just another asset class, one without any fundamental or socially positive value. A little like gold, actually, except without gold's resilience against social / technological disruption.
• The disastrous post-Soviet economic transition, on our advice.
• The enormous labor drain, and resource drain, from global South to North.

Preliminary Pieces of AI

We already live in an AI world, and really, it isn't so bad.

It is odd to hear about all the hyperventilating about artificial intelligence of late. One would think it is a new thing, or some science-fiction-y entity. Then there are fears about the singularity and loss of control by humans. Count me a skeptic on all fronts. Man is, and remains, wolf to man. To take one example, we are contemplating the election of perhaps the dummbest person ever to hold the office of president. For the second time. How an intelligence, artificial or otherwise, is supposed to worm its way into power over us is not easy to understand, looking at nature of humans and of power. 

So let's take a step back and figure out what is going on, and where it is likely to take us. AI has become a catch-all for a diversity of computer methods, mostly characterized by being slightly better at doing things we have long wanted computers to do, like interpreting text, speech, and images. But I would offer that it should include much more- all the things we have computers do to manage information. In that sense, we have been living among shards of artificial intelligence for a very long time. We have become utterly dependent on databases, for instance, for our memory functions. Imagine having to chase down a bank balance or a news story, without access to the searchable memories that modern databases provide. They are breathtakingly superior to our own intelligence when it comes to the amount of things they can remember, the accuracy they can remember them, and the speed with which they can find them. The same goes for calculations of all sorts, and more recently, complex scientific math like solving atomic structures, creating wonderful CGI graphics, or predicting the weather. 

We should view AI as a cabinet filled with many different tools, just as our own bodies and minds are filled with many kinds of intelligence. The integration of our minds into a single consciousness tends to blind us to the diversity of what happens under the hood. While we may want gross measurements like "general intelligence", we also know increasingly that it (whatever "it" is, and whatever it fails to encompass of our many facets and talents) is composed of many functions that several decades of work in AI, computer science, and neuroscience have shown are far more complicated and difficult to replicate than the early AI pioneers imagined, once they got hold of their Turing machine with its infinite potential. 

Originally, we tended to imagine artificial intelligence as a robot- humanoid, slightly odd looking, but just like us in form and abilities. That was a natural consequence of our archetypes and narcissism. But AI is nothing like that, because full-on humanoid consciousness is an impossibly high bar, at least for the foreseeable future, and requires innumerable capabilities and forms of intelligence to be developed first. 

The autonomous car drama is a good example of this. It has taken every ounce of ingenuity and high tech to get to a reasonably passable vehicle, which is able to "understand" key components of the world around it. That a blob in front is a person, instead of a motorcycle, or that a light is a traffic light instead of a reflection of the sun. Just as our brain has a stepwise hierarchy of visual processing, we have recapitulated that evolution here by harnessing cameras in these cars (and lasers, etc.) to not just take in a flat visual scene, which by itself is meaningless, but to parse it into understandable units like ... other cars, crosswalks, buildings, bicylists, etc.. Visual scenes are very rich, and figuring out what is in them is a huge accomplishment. 

But is it intelligence? Yes, it certainly is a fragment of intelligence, but it isn't consciousness. Imagine how effortless this process is for us, and how effortful and constricted it is for an autonomous vehicle. We understand everything in a scene within a much wider model of the world, where everything relates to everything else. We evaluate and understand innumerable levels of our environment, from its chemical makeup to its social and political implications. Traffic cones do not freak us out. The bare obstacle course of getting around, such as in a vehicle, is a minor aspect, really, of this consciousness, and of our intelligence. Autonomous cars are barely up to the level of cockroaches, on balance, in overall intelligence.

The AI of text and language handling is similarly primitive. Despite the vast improvements in text translation and interpretation, the underlying models these mechanisms draw on are limited. Translation can be done without understanding text at all, merely by matching patterns from pre-digested pools of pre-translated text, regurgitated as cued by the input text. Siri-like spoken responses, on the other hand, do require some parsing of meaning out of the input, to decide what the topic and the question are. But the scope of these tools tend to be very limited, and the wider scope they are allowed, the more embarrassing their performance, since they are essentially scraping web sites and text pools for question-response patterns, instead of truly understanding the user's request or any field of knowledge.

Lastly, there are the generative ChatGPT style engines, which also regurgitate text patterns reformatted from public sources in response to topical requests. The ability to re-write a Wikipedia entry through a Shakespeare filter is amazing, but it is really the search / input functions that are most impressive- being able, like the Siri system, to parse through the user's request for all its key points. This betokens some degree of understanding, in the sense that the world of the machine (i.e. its database) is parceled up into topics that can be separately gathered and reshuffled into a response. This requires a pretty broad and structured ontological / classification system, which is one important part of intelligence.

Not only is there a diversity of forms of intelligence to be considered, but there is a vast diversity of expertise and knowledge to be learned. There are millions of jobs and professions, each with their own forms of knowledge. Back the early days of AI, we thought that expert systems could be instructed by experts, formalizing their expertise. But that turned out to be not just impractical, but impossible, since much of that expertise, formed out of years of study and experience, is implicit and unconscious. That is why apprenticeship among humans is so powerful, offering a combination of learning by watching and learning by doing. Can AI do that? Only if it gains several more kinds of intelligence including an ability to learn in very un-computer-ish ways.

This analysis has emphasized the diverse nature of intelligences, and the uneven, evolutionary development they have undergone. How close are we to a social intelligence that could understand people's motivations and empathise with them? Not very close at all. How close are we to a scientific intelligence that could find holes in the scholarly literature and manage a research enterprise to fill them? Not very close at all. So it is very early days in terms of anything that could properly be called artificial intelligence, even while bits and pieces have been with us for a long time. We may be in for fifty to a hundred more years of hearing every advance in computer science being billed as artificial intelligence.

Uneven development is going to continue to be the pattern, as we seize upon topics that seem interesting or economically rewarding, and do whatever the current technological frontier allows. Memory and calculation were the first to fall, being easily formalizable. Communication network management is similarly positioned. Game learning was next, followed by the Siri / Watson systems for question answering. Then came a frontal assault on language understanding, using the neural network systems, which discard the former expert system's obsession with grammar and rules, for much simpler statistical learning from large pools of text. This is where we are, far from fully understanding language, but highly capable in restricted areas. And the need for better AI is acute. There are great frontiers to realize in medical diagnosis and in the modeling of biological systems, to only name two fields close at hand that could benefit from a thoroughly systematic and capable artificial intelligence.

The problem is that world modeling, which is what languages implicitly stand for, is very complicated. We do not even know how to do this properly in principle, let alone having the mechanisms and scale to implement it. What we have in terms of expert systems and databases do not have the kind of richness or accessibility needed for a fluid and wide-ranging consciousness. Will neural nets get us there? Or ontological systems / databases? Or some combination? However it is done, full world modeling with the ability to learn continuously into those models are key capabilities needed for significant artificial intelligence.

After world modeling come other forms of intelligence like social / emotional intelligence and agency / management intelligence with motivation. I have no doubt that we will get to full machine consciousness at some point. The mechanisms of biological brains are just not sufficiently mysterious to think that they can not be replicated or improved upon. But we are nowhere near that yet, despite bandying about the word artificial intelligence. When we get there, we will have to pay special attention to the forms of motivation we implant, to mitigate the dangers of making beings who are even more malevolent than those that already exist... us.

Would that constitute some kind of "singularity"? I doubt it. Among humans there are already plenty of smart people and diversity, which result in niches for everyone having something useful to do. Technology has been replacing human labor forever, and will continue moving up the chain of capability. And when machines exceed the level of human intelligence, in some general sense, they will get all the difficult jobs. But the job of president? That will still go to a dolt, I have no doubt. Selection for some jobs is by criteria that artificial intelligence, no matter how astute, is not going to fulfill.

Risks? In the current environment, there are a plenty of risks, which are typically cases where technology has outrun our will to regulate its social harm. Fake information, thanks to the chatbots and image makers, can now flood the zone. But this is hardly a new phenomenon, and perhaps we need to get back to a position where we do not believe everything we read, in the National Enquirer or on the internet. The quality of our sources may become once again an important consideration, as they always should have been.

Another current risk is that the automation risks chaos. For example in the financial markets, the new technologies seem to calm the markets most of the time, arbitraging with relentless precision. But when things go out of bounds, flash breakdowns can happen, very destructively. The SEC has sifted through some past events of this kind and set up regulatory guard rails. But they will probably be perpetually behind the curve. Militaries are itching to use robots instead of pilots and soldiers, and to automate killing from afar. But ultimately, control of the military comes down to social power, which comes down to people of not necessarily great intelligence. 

The biggest risk from these machines is that of security. If we have our financial markets run by machine, or our medical system run by super-knowledgeable artificial intelligences, or our military by some panopticon neural net, or even just our electrical grid run by super-computers, the problem is not that they will turn against us of their own volition, but that some hacker somewhere will turn them against us. Countless hospitals have already faced ransomware attacks. This is a real problem, growing as machines become more capable and indispensable. If and when we make artificial people, we will need the same kind of surveillance and social control mechanisms over them that we do over everyone else, but with the added option of changing their programming. Again, powerful intelligences made for military purposes to kill our enemies are, by the reflexive property of all technology, prime targets for being turned against us. So just as we have locked up our nuclear weapons and managed to not have them fall into enemy hands (so far), similar safeguards would need to be put on similar powers arising from these newer technologies.

We may have been misled by the many AI and super-beings of science fiction, Nietzsche's Übermensch, and similar archetypes. The point of Nietzsche's construction is moral, not intellectual or physical- a person who has thrown off all the moral boundaries of civilization, expecially Christian civilization. But that is a phantasm. The point of most societies is to allow the weak to band together to control the strong and malevolent. A society where the strong band together to enslave the weak.. well, that is surely a nightmare, and more unrealistic the more concentrated the power. We must simply hope that, given the ample time we have before truly comprehensive and superior artificial intelligent beings exist, we have exercised sufficient care in their construction, and in the health of our political institutions, to control them as we have many other potentially malevolent agents.

• AI in chemistry.
• AI to recognize cells in images.
• Ayaan Hirsi Ali becomes Christian. "I ultimately found life without any spiritual solace unendurable."
• The racism runs very deep.
• An appreciation of Stephen J. Gould.
• Forced arbitration against customers and employees is OK, but fines against frauds... not so much?
• Oil production still going up.

Many Ways There are to Read a Genome

New methods to unravel the code of transcriptional regulators.

When we deciphered the human genome, we came up with three billion letters of its linear code- nice and tidy. But that is not how it is read inside our cells. Sure, it is replicated linearly, but the DNA polymerases don't care about the sequence- they are not "reading" the book, they are merely copying machines trying to get it to the next generation with as few errors as possible. The book is read in an entirely different way, by a herd of proteins that recognize specific sequences of the DNA- the transcription regulators (also commonly called transcription factors [TF], in the classic scientific sense of some "factor" that one is looking for). These regulators- and there are, by one recent estimate, 1,639 of them encoded in the human genome- constitute an enormously complex network of proteins and RNAs that regulate each other, and regulate "downstream" genes that encode everything else in the cell. They are made in various proportions to specify each cell type, to conduct every step in development, and to respond to every eventuality that evolution has met and mastered over the eons.

Loops occur in the DNA between site of regulator binding, in order to turn genes on (enhancer, E, and transcription regulator/factor, TF).

Once sufficient transcription regulators bind to a given gene, it assembles a transcription complex at its start site, including the RNA polymerase that then generates an RNA copy that can float off to be made into a protein, (such as a transcription regulator), or perhaps function in its RNA form as part of zoo of rRNA, tRNA, miRNA, piRNA, and many more that also help run the cell. Some regulators can repress transcription, and many cooperate with each other. There are also diverse regions of control for any given target gene in its nearby non-coding DNA- cassettes (called enhancers) that can be bound by different regulators and thus activated at different stages for different reasons. 

These binding sites in the DNA that transcription regulators bind to are typically quite small. A classic regulator SP1 (itself 785 amino acids long and bearing three consecutive DNA binding motifs coordinated by a zinc ions) binds to a sequence resembling (G/T)GGGCGG(G/A)(G/A)(C/T). So only ten bases are specified at all, and four of those positions are degenerate. By chance, a genome of three billion bases will have such a sequence about 45,769 times. So this kind of binding is not very strictly specified, and such sites tend to appear and disappear frequently in evolution. That is one of the big secrets of evolution- while some changes are hard, others are easy, and it there is constant variation and selection going on in the regulatory regions of genes, refining and defining where / when they are expressed.

Anyhow, researchers naturally have the question- what is the regulatory landscape of a given gene under some conditions of interest, or of an entire genome? What regulators bind, and which ones are most important? Can we understand, given our technical means, what is going on in a cell from our knowledge of transcription regulators? Can we read the genome like the cell itself does? Well the answer to that is, obviously no and not yet. But there are some remarkable technical capabilities. For example, for any given regulator, scientists can determine where it binds all over the genome in any given cell, by chemical crosslinking methods. The prediction of binding sites for all known regulators has been a long-standing hobby as well, though given the sparseness of this code and the lability of the proteins/sites, one that gives only statistical, which is to say approximate, results. Also, scientists can determine across whole genomes where genes are "open" and active, vs where they are closed. Chromatin (DNA bound with histones in the nucleus) tends to be closed up on repressed and inactive genes, while transcription regulators start their work by opening chromatin to make it accessible to other regulators, on active genes.

This last method offers the prospect of truly global analysis, and was the focus of a recent paper. The idea was to merge a detailed library predicted binding sites for all known regulators all over the genome, with experimental mapping of open chromatin regions in a particular cell or tissue of interest. And then combine all that with existing knowledge about what each of the target genes near the predicted binding sites do. The researchers clustered the putative regulators binding across all open regions by this functional gene annotation to come up with statistically over-represented transcription regulators and functions. This is part of a movement across bioinformatics to fold in more sources of data to improve predictions when individual methods each produce sketchy, unsatisfying results.

In this case, mapping open chromatin by itself is not very helpful, but becomes much more helpful when combined with assessments of which genes these open regions are close to, and what those genes do. This kind of analysis can quickly determine whether you are looking at an immune cell or a neuron, as the open chromatin is a snapshot of all the active genes at a particular moment. In this recent work, the analysis was extended to say that if some regulators are consistently bound near genes participating in some key cellular function, then we can surmise that that regulator may be causal for that cell type, or at least part of the program specific to that cell. The point for these researchers is that this multi-source analysis performs better in finding cell-type specific, and function-specific, regulators than is the more common approach of just adding up the prevalence of regulators occupying open chromatin all over a given genome, regardless of the local gene functions. That kind of approach tends to yield common regulators, rather than cell-type specific ones. 

To validate, they do rather half-hearted comparisons with other pre-existing techniques, without blinding, and with validation of only their own results. So it is hardly a fair comparison. They look at the condition systemic lupus (SLE), and find different predictions coming from their current technique (called WhichTF) vs one prior method (MEME-ChIP).  MEME-ChIP just finds predicted regulator binding sites for genomic regions (i.e. open chromatin regions) given by the experimenter, and will do a statistical analysis for prevalence, regardless of the functions of either the regulator or the genes it binds to. So you get absolute prevalence of each regulator in open (active) regions vs the genome as a whole. 

Different regulators are identified from the same data by different statistical methods. But both sets are relevant.

What to make of these results? The MEME-ChIP method finds regulators like SP1, SP2, SP4, and ZFX/Y. SP1 et al. are very common regulators, but that doesn't mean they are unimportant, or not involved in disease processes. SP1 has been observed as likely to be involved in autoimmune encephalitis in mice, a model of multiple sclerosis, and naturally not so far from lupus in pathology. ZFX is also a prominent regulator in the progenitor cells of the immune system. So while these authors think little of the competing methods, those methods seem to do a very good job of identifying significant regulators, as do their own methods. 

There is another problem with the author's WhatTF method, which is that gene annotation is in its infancy. Users are unlikely to find new functions using existing annotations. Many genes have no known function yet, and new functions are being found all the time for those already assigned functions. So if one's goal is classification of a cell or of transcription regulators according to existing schemes, this method is fine. But if one has a research goal to find new cell types, or new processes, this method will channel you into existing ones instead.

This kind of statistical refinement is unlikely to give us what we seek in any case- a strong predictive model of how the human genome is read and activatated by the herd of gene regulators. For that, we will need new methods for specific interaction detection, with a better appreciation for complexes between different regulators, (which will be afforded by the new AI-driven structural techniques), and more appreciation for the many other operators on chromatin, like the various histone modifying enzymes that generate another whole code of locks and keys that do the detailed regulation of chromatin accessibility. Reading the genome is likely to be a somewhat stochastic process, but we have not yet arrived at the right level of detail, or the right statistics, to do it justice.

• Unconscious messaging and control. How the dark side operates.
• Solzhenitsyn on evil.
• Come watch a little Russian TV.
• "Ruthless beekeeping practices"
• The medical literature is a disaster.

A Gene is Born

Yes, genes do develop out of nothing.

The "intelligent" design movement has long made a fetish of information. As science has found, life relies on encoded information for its genetic inheritance and the reliable expression of its physical manifestations. The ID proposition is, quite simply, that all this information could not have developed out of a mindless process, but only through "design" by a conscious being. Evidently, Darwinian natural selection still sticks on some people's craw. Michael Behe even developed a pseudo-mathematical theory about how, yes, genes could be copied mindlessly, but new genes could never be conjured out of nothing, due to ... information.

My understanding of information science equates information to loss of entropy, and expresses a minimal cost of the energy needed to create, compute or transmit information- that is, the Shannon limits. A quite different concept comes from physics, in the form of information conservation in places like black holes. This form of information is really the implicit information of the wave functions and states of physical matter, not anything encoded or transmitted in the sense of biology or communication. Physical state information may be indestructable (and un-create-able) on this principle, but coded information is an entirely different matter.

In a parody of scientific discussion, intelligent design proponents are hosted by the once-respectable Hoover Institution for a discussion about, well, god.

So the fecundity that life shows in creating new genes out of existing genes, (duplications), and even making whole-chromosome or whole-genome duplications, has long been a problem for creationists. Energetically, it is easy to explain as a mere side-effect of having plenty of energy to work with, combined with error-prone methods of replication. But creationistically, god must come into play somewhere, right? Perhaps it comes into play in the creation of really new genes, like those that arise from nothing, such as at the origin of life?

A recent paper discussed genes in humans that have over our recent evolutionary history arisen from essentially nothing. It drew on prior work in yeast that elegantly laid out a spectrum or life cycle of genes, from birth to death. It turns out that there is an active literature on the birth of genes, which shows that, just like duplication processes, it is entirely natural for genes to develop out of humble, junky precursors. And no information theory needs to be wheeled in to show that this is possible.

Yeast provides the tools to study novel genes in some detail, with rich genetics and lots of sequenced relatives, near and far. Here is portrayed a general life cycle of a gene, from birth out of non-gene DNA sequences (left) into the key step of translation, and on to a subject of normal natural selection ("Exposed") for some function. But if that function decays or is replaced, the gene may also die, by mutation, becoming a pseudogene, and eventually just some more genomic junk.

The death of genes is quite well understood. The databases are full of "pseudogenes" that are very similar to active genes, but are disabled for some reason, such as a truncation somewhere or loss of reading frame due to a point mutation or splicing mutation. Their annotation status is dynamic, as they are sometimes later found to be active after all, under obscure conditions or to some low level. Our genomes are also full of transposons and retroviruses that have died in this fashion, by mutation.

Duplications are also well-understood, some of which have over evolutionary time given rise to huge families of related proteins, such as kinases, odorant receptors, or zinc-finger transcription factors. But the hunt for genes that have developed out of non-gene materials is a relatively new area, due to its technical difficulty. Genome annotators were originally content to pay attention to genes that coded for a hundred amino acids or more, and ignore everything else. That became untenable when a huge variety of non-coding RNAs came on the scene. Also, occasional cases of very small genes that encoded proteins came up from work that found them by their functional effects.

As genome annotation progressed, it became apparent that, while a huge proportion of genes are conserved between species, (or members of families of related proteins), other genes had no relatives at all, and would never provide information by this highly convenient route of computer analysis. They are orphans, and must have either been so heavily mutated since divergence that their relationships have become unrecognizable, or have arisen recently (that is, since their evolutionary divergence from related species that are used for sequence comparison) from novel sources that provide no clue about their function. Finer analysis of ever more closely related species is often informative in these cases.

The recent paper on human novel genes makes the finer point that splicing and export from the nucleus constitute the major threshold between junk genes and "real" genes. Once an RNA gets out of the nucleus, any reading frame it may have will be translated and exposed to selection. So the acquisition of splicing signals is a key step, in their argument, to get a randomly expressed bit of RNA over the threshold.

A recent paper provided a remarkable example of novel gene origination. It uncovered a series of 74 human genes that are not shared with macaque, (which they took as their reference), have a clear path of origin from non-coding precursors, and some of which have significant biological effects on human development. They point to a gradual process whereby promiscuous transcription from the genome gave rise by chance to RNAs that acquired splice sites, which piped them into the nuclear export machinery and out to the cytoplasm. Once there, they could be translated, over whatever small coding region they might possess, after which selection could operate on their small protein products. A few appear to have gained enough function to encourage expansion of the coding region, resulting in growth of the gene and entrenchment as part of the developmental program.

Brain "organoids" grown from genetically manipulated human stem cells. On left is the control, in middle is where ENSG00000205704 was deleted, and on the right is where ENSG00000205704 is over-expressed. The result is very striking, as an evolutionarily momentous effect of a tiny and novel gene.

One gene, "ENSG00000205704" is shown as an example. Where in macaque, the genomic region corresponding to this gene encodes at best a non-coding RNA that is not exported from the nucleus, in humans it encodes a spliced and exported mRNA that encodes a protein of 107 amino acids. In humans it is also highly expressed in the brain, and when the researchers deleted it in embryonic stem cells and used those cells to grow "organoids", or clumps of brain-like tissue, the growth was significantly reduced by the knockout, and increased by the over-expression of this gene. What this gene does is completely unknown. Its sequence, not being related to anything else in human or other species, gives no clue. But it is a classic example of gene that arose from nothing to have what looks like a significant effect on human evolution. Does that somehow violate physics or math? Nothing could be farther from the truth.

• Will nuclear power get there?
• What the heck happened to Amazon shopping?

Tangling With the Network

Molecular biology needs better modeling.

Molecular biologists think in cartoons. It takes a great deal of work to establish the simplest points, like that two identifiable proteins interact with each other, or that one phosphorylates the other, which has some sort of activating effect. So biologists have been satsified to achieve such critical identifications, and move on to other parts of the network. With 20,000 genes in humans, expressed in hundreds of cell types, regulated states and disease settings, work at this level has plenty of scope to fill years of research.

But the last few decades have brought larger scale experimentation, such as chips that can determine the levels of all proteins or mRNAs in a tissue, or the sequences of all the mRNAs expressed in a cell. And more importantly, the recognition has grown that any scientific field that claims to understand its topic needs to be able to model it, in comprehensive detail. We are not at that point in molecular biology, at all. Our experiments, even those done at large scale and with the latest technology, are in essence qualitative, not quantitative. They are also crudely interventionistic, maybe knocking out a gene entirely to see what happens in response. For a system as densely networked as the eukaryotic cell, it will take a lot more to understand and model it.

One might imagine that this is a highly detailed model of cellular responses to outside stimuli. But it is not. Some of the connections are much less important than others. Some may take hours to have the indicated effect, while others happen within seconds or less. Some labels hide vast sub-systems with their own dynamics. Important items may still be missing, or assumed into the background. Some connections may be contingent on (or even reversed by) other conditions that are not shown. This kind of cartoon is merely a suggestive gloss and far from a usable computational (or true) model of how a biological regulatory system works.

The field of biological modeling has grown communities interested in detailed modeling of metabolic networks, up to whole cells. But these remain niche activities, mostly because of a lack of data. Experiments remain steadfastly qualitative, given the difficulty of performing them at all, and the vagaries of the subjects being interrogated. So we end up with cartoons, which lack not only quantitative detail on the relative levels of each molecule, but also critical dynamics of how each relationship develops in time, whether in a time scale of seconds or milliseconds, as might be possible for phosphorylation cascades (which enable our vision, for example), or a time scale of minutes, hours, or days- the scale of changes in gene expression and longer-term developmental changes in cell fate.

These time and abundance variables are naturally critical to developing dynamic and accurate models of cellular activities. But how to get them? One approach is to work with simple systems- perhaps a bacterial cell rather than a human cell, or a stripped down minimal bacterial cell rather than the E. coli standard, or a modular metabolic sub-network. Many groups have labored for years to nail down all the parameters of such systems, work which remains only partially successful at the organismal scale.

Another approach is to assume that co-expressed genes are yoked together in expression modules, or regulated by the same upstream circuitry. This is one of the earliest forms of analysis for large scale experiments, but it ignores all the complexity of the network being observed, indeed hardly counts as modeling at all. All the activated genes are lumped together into one side, and all the down-regulated genes on the other side, perhaps filtered by biggest effect. The resulting collections are clustered by some annotation of those gene's functions, thereby helping the user infer what general cell function was being regulated in her experiment / perturbation. This could be regarded perhaps as the first step on a long road from correlation analysis of gene activities to a true modeling analysis that operates with awareness of how individual genes and their products interact throughout a network.

Another approach is to resort to a lot of fudge factors, while attempting to make a detailed model of the cell /components. Assume a stable network, and fill in all the values that could get you there, given the initial cartoon version of molecule interactions. Simple models thus become heuristic tools to hunt for missing factors that affect the system, which are then progressively filled in, hopefully by doing new experiments. Such factors could be new components, or could be unsuspected dynamics or unknown parameters of those already known. This is, incidentally, of intense interest to drug makers, whose drugs are intended to tweek just the right part of the system in order to send it to a new state- say, from cancerous back to normal, well-behaved quiescence.

A recent paper offered a version of this approach, modular response analysis (MRA). The authors use perturbation data from other labs, such as the inhibition of 1000 different genes in separately assayed cells, combined with a tentative model of the components of the network, and then deploy mathematical techniques to infer / model the dynamics of how that cellular system works in the normal case. What is observed in either case- the perturbed version, or the wild-type version- is typically a system (cell) at steady state, especially if the perturbation is something like knocking out a gene or stably expressing an inhibitor of its mRNA message. Thus, figuring out the (hidden) dynamic in between- how one stable state gets to another one after a discrete change in one or more components- is the object of this quest. Molecular biologists and geneticists have been doing this kind of thing off-the-cuff forever (with mutations, for instance, or drugs). But now we have technologies (like siRNA silencing) to do this at large scale, altering many components at will and reading off the results.

This paper extends one of the relevant mathematical methods (modular response analysis, MRA) to this large scale, and finds that, with a bit of extra data and some simplifications, it is competitive with other methods (mutual information) in creating dynamic models of cellular activities, at the scale of a thousand components, which is apparently unprecedented. At the heart of MRA are, as its name implies, modules, which break down the problem into manageable portions and allow variable amounts of detail / resolution. For their interaction model, they use a database of protein interactions, which is a reasonably comprehensive, though simplistic, place to start.

What they find is that they can assemble an effective system that handles both real and simulated data, creating quantitative networks from their inputs of gene expression changes upon inhibition of large numbers of individual components, plus a basic database of protein relationships. And they can do so at reasonable scale, though that is dependent on the ability to modularize the interaction network, which is dangerous, as it may ignore important interactions. As a state of the art molecular biology inference system, it is hardly at the point of whole cell modeling, but is definitely a few steps ahead of the cartoons we typically work with.

The authors offer this as one result of their labors. Grey nodes are proteins, colored lines (edges) are activating or inhibiting interactions. Compared to the drawing above, it is decidedly more quantitative, with strengths of interactions shown. But timing remains a mystery, as do many other details, such as the mechanisms of the interactions

• Fiscal contraction + interest rate increase + trade deficit = recession.
• The lies come back to roost.
• Status of carbon removal.
• A few notes on stuttering.
• A pious person, on shades of abortion.
• Discussion on the rise of China.

Computer Science Meets Neurobiology in the Hippocampus

Review of Whittington, et al. - a theoretical paper on the generalized mapping and learning capabilities of the entorhinal/hippocampal complex that separates memories from a graph-mapping grid.

These are exciting times in neurobiology, as a grand convergence is in the offing with computational artificial intelligence. AI has been gaining powers at a rapid clip, in large part due to the technological evolution of neural networks. But computational theory has also been advancing, on questions of how concepts and relations can be gleaned from data- very basic questions of interest to both data scientists and neuroscientists. On the other hand, neurobiology has benefited from technical advancements as well, if far more modestly, and from the relentless accumulation of experimental and clinical observations. Which is to say, normal science. 

One of the hottest areas of neuroscience has been the hippocampus and the closely connected entorhinal cortex, seat of at least recent memory and of navigation maps and other relational knowledge. A recent paper extends this role to a general theory of relational computation in the brain. The basic ingredients of thought are objects and relations. Computer scientists typically represent these as a graph, where the objects are nodes, and the relations are the connecting lines, or edges. Nodes can have a rich set of descriptors (or relations to property nodes that express these descriptions). A key element to get all this off the ground is the ability to chunk, (or abstract, or generalize, or factorize) observations into discrete entities, which then serve as the objects of the relational graph. The ability to say that what you are seeing, in its whirling and colorful reality, is a dog .. is a very significant opening step to conceptualization, and the manipulation of those concepts in useful ways, such as understanding past events and predicting future ones.

Gross anatomy of the hippocampus and associated entorhinal cortex, which function together in conceptual binding and memory.

A particular function of the entorhinal/hippocampal complex is spatial navigation. Reseachers have found place cells, grid cells, and boundary cells (describing when these cells fire) as clear elements of spatial consciousness, which even replay in dreams as the rats re-run their daytime activities. It is evident that these cells are part of an abstraction mechanism that dissociates particular aspects of conceptualized sensory processing from the total scene and puts them back together again in useful ways, i.e. as various maps.

This paper is conducted at a rather abstruse level, so there is little that I can say about it in detail. Yet it and the field it contributes to is so extremely interesting that some extra effort is warranted. By the time the hippocampus is reached, visual (and other sensory data) has already been processed to the conceptual stage. Dogs have been identified, landmarks noted, people recognized. Memories are composed of fully conceptualized, if also sensorily colored, conceptual chunks. The basic idea the authors present is that key areas of the entorhinal cortex provide general and modular mapping services that allow the entorhinal/hippocampal complex to deal with all kinds of relational information and memories, not just physical navigation. Social relations, for example, are mapped similarly.

It is important to note tangentially that conceptualization is an emergent process in the brain, not dictated by pre-existing lists of entities or god-given databases of what exists in the world and beyond. No, all this arises naturally from experience in the world, and it has been of intense interest to computer scientists to figure out how to do this efficiently and accurately, on a computer. Some recent work was cited here and is interesting for its broad implications as well. It is evident that we will in due time be faced with fully conceptualizing, learning, and thinking machines.

"Structural sparsity also brings a new perspective to an old debate in cognitive science between symbolic versus emergent approaches to knowledge representation. The symbolic tradition uses classic knowledge structures including graphs, grammars, and logic, viewing these representations as the most natural route towards the richness of thought. The competing emergent tradition views these structures as epiphenomena: they are approximate characterizations that do not play an active cognitive role. Instead, cognition emerges as the cooperant consequence of simpler processes, often operating over vector spaces and distributed representations. This debate has been particularly lively with regards to conceptual organization, the domain studied here. The structural forms model has been criticized by the emergent camp for lacking the necessary flexibility for many real domains, which often stray from pristine forms. The importance of flexibility has motivated emergent alternatives, such as a connectionist network that maps animals and relations on the input side to attributes on the output side. As this model learns, an implicit tree structure emerges in its distributed representations. But those favoring explicit structure have pointed to difficulties: it becomes hard to incorporate data with direct structural implications like 'A dolphin is not a fish although it looks like one', and latent objects in the structure support the acquisition of superordinate classes such as 'primate' or 'mammal'. Structural sparsity shows how these seemingly incompatible desiderata could be satisfied within a single approach, and how rich and flexible structure can emerge from a preference for sparsity." - from Lake et al., 2017

Getting back the the hippocampus paper, the authors develop a computer model, which they dub the Tolman-Eichenbaum machine [TEM] after key workers in the field. This model implements a three-part system modeled on the physiological situation, plus their theory of how relational processing works. Medial entorhinal cells carry generalized mapping functions (grids, borders, vectors), which can be re-used for any kind of object/concept, supplying relations as originally deduced from sensory processing or possibly other abstract thought. Lateral entorhinal cells carry specific concepts or objects as abstracted from sensory processing, such as landmarks, smells, personal identities, etc. It is then the crossing of these "what" and "where" streams that allows navigation, both in reality and in imagination. This binding is proposed to happen in the hippocampus, as firing that happens when firing from the two separate entorhinal regions happen to synchronize, stating that a part of the conceptual grid or other map and an identified object have been detected in the same place, generating a bound sensory experience, which can be made into a memory, or arise from a memory, or an imaginative event, etc. This is characteristic of "place cells", hippocampal cells that fire when the organism is at a particular place, and not at other times.

"We propose TEM’s [the computational model they call a Tolman-Eichenbaum Machine] abstract location representations (g) as medial entorhinal cells, TEM’s grounded variables (p) as hippocampal cells, and TEM’s sensory input x as lateral entorhinal cells. In other words, TEM’s sensory data (the experience of a state) comes from the ‘what stream’ via lateral entorhinal cortex, and TEM’s abstract location representations are the ‘where stream’ coming from medial entorhinal cortex. TEM’s (hippocampal) conjunctive memory links ‘what’ to ‘where’, such that when we revisit ‘where’ we remember ‘what’."

Given the abstract mapping and a network of relations between each of the components, reasoning or imagining about possible events also becomes feasible, since the system can solve for any of the missing components. If a landmark is seen, a memory can be retrieved that binds the previously known location. If a location is surmised or imagined, then a landmark can be dredged up from memory to predict how that location looks. And if an unfamiliar combination of location and landmark is detected, then either a new memory can be made, or a queasy sense of unreality or hallucination would ensue if one of the two are well-known enough to make the disagreement disorienting.

As one can tell, this allows not only the experience of place, but the imagination of other places, as the generic mapping can be traversed imaginatively, even by paths that the organism has never directly experienced, to figure out what would happen if one, for instance, took a short-cut. 

The combination of conceptual abstraction / categorization with generic mapping onto relational graph forms that can model any conceptual scale provides some of the most basic apparatus for cognitive thought. While the system discussed in this paper is mostly demonstrated for spatial navigation, based on the proverbial rat maze, it is claimed, and quite plausible, that the segregation of the mapping from the object identification and binding allows crucial generalization of cognition- the tools we, and someday AI as well, rely on to make sense of the world.

• A database of superspreader events. It suggests indoor, poorly ventilated spread of small aerosols. And being in a cold locker helps as well.
• The curious implications of pardons.
• It is going to be a long four years.
• How tires kill salmon.
• Ever think that there is more to life?

High Intelligence is Highly Overrated by Highly Intelligent People

AI, the singularity, and watching way too much science fiction: Review of Superintelligence by Nic Bostrom.

How far away is the singularity? That is the point when machine intelligence exceeds human intelligence, after which it is thought that this world will no longer be ours to rule. Rick Bostrom, a philosopher at Oxford, doesn't know when this will be, but is fearful of its consequences, since, if we get it wrong, humanity's fate may not be a happy one.

The book starts strongly, with some well argued and written chapters about the role of intelligence in humanity's evolution, and the competitive landscape of technology today that is setting the stage for this momentous transition. But thereafter, the armchair philosopher takes over, with tedious chapters of hairsplitting and speculation about how fast or slow the transition might be, how collaborative among research groups, and especially, how we could pre-out-think these creations of ours, to make sure they will be well-disposed to us, aka "the control problem".

Despite the glowing blurbs from Bill Gates and others on the jacket, I think there are fundamental flaws with this whole approach and analysis. One flaw is a failure to distinguish between intelligence and power. Our president is a moron. That should tell us something about this relationship. It is not terribly close- the people generally acknowledged as the smartest in history have rarely been the most powerful. This reflects a deeper flaw, which is, as usual, a failure to take evolution and human nature seriously. The "singularity" is supposed to furnish something out of science fiction- a general intelligence superior to human intelligence. But Bostrom and others seem to think that this means a fully formed human-like agent, and those are two utterly different things. Human intelligence takes many forms, and human nature is composed of many more things than intelligence. Evolution has strained for billions of years to form our motivations in profitable ways, so that we follow others when necessary, lead them when possible, define our groups in conventional ways that lead to warfare against outsiders, etc., etc. Our motivational and social systems are not the same as our intelligence system, and to think that anyone making an AI with general intelligence capabilities will, will want to, or even can, just reproduce the characteristics of human motivation to tack on and serve as its control system, is deeply mistaken.

The fact is that we have AI right now that far exceeds human capabilities. Any database is far better at recall than humans are, to the point that our memories are atrophying as we compulsively look up every question we have on Wikipedia or Google. And any computer is far better at calculations, even complex geometric and algebraic calculations, than we are in our heads. That has all been low-hanging fruit, but it indicates that this singularity is likely to be something of a Y2K snoozer. The capabilities of AI will expand and generalize, and transform our lives, but unless weaponized with explicit malignant intent, it has no motivation at all, let alone the motivation to put humanity into pods for its energy source, or whatever.

People-pods, from the Matrix.

The real problem, as usual, is us. The problem is the power that accrues to those who control this new technology. Take Mark Zuckerberg for example. He stands at the head of multinational megacorporation that has inserted its tentacles into the lives of billions of people, all thanks to modestly intelligent computer systems designed around a certain kind of knowledge of social (and anti-social) motivations. All in the interests of making another dollar. The motivations for all this do not come from the computers. They come from the people involved, and the social institutions (of capitalism) that they operate in. That is the real operating system that we have yet to master.

• Facebook - the problem is empowering the wrong people, not the wrong machines.
• Barriers to health care.
• What it is like to be a psychopathic moron.

We'll Keep Earth

The robots can have the rest of the universe.

The Apollo 11 aniversary is upon us, a wonderful achievement and fond memory. But it did not lead to the hopeful new-frontier future that has been peddled by science fiction for decades, for what are now obvious reasons. Saturn V rockets do not grow on trees, nor is space, once one gets there, hospitable to humans. Earth is our home, where we evolved and are destined to stay.

But a few among us have continued taking breathtaking adventures among the planets and toward other stars. They have done pirouettes around the Sun and all the planets, including Pluto. They are our eyes in the heavens- the robots. I have been reading a sober book, Nick Bostrom's Superintelligence, which works through in painstaking, if somewhat surreal, detail what artificial intelligence will become in the not too distant future. Whether there is a "singularity" in a few decades, or farther off, there will surely come a time when we can reproduce human level intelligence (and beyond) in machine form. Already, machines have far surpassed humans in memory capacity, accuracy, and recall speed, in the form of databases that we now rely on to run every bank, government, and app. It seems inescapable that we should save ourselves the clunky absurdity, vast expense, and extreme dangers of human spaceflight and colonization in favor of developing robots with increasing capabilities to do all that for us.

It is our fleet of robots that can easily withstand the radiation, weightlessness, vacuum, boredom, and other rigors of space. As they range farther, their independence increases. On the Moon, at 1.3 light seconds away, we can talk back and forth, and control things in near real time from Earth. The Mars rovers, on the other hand, needed to have some slight intelligence to avoid obstacles and carry out lengthy planned maneuvers, being roughly 15 light-minutes from Earth. Having any direct control over rovers and other probes farther afield is increasingly impossible, with Jupiter 35 minutes away, and Neptune four light hours away. Rovers or drones contemplated for Saturn's interesting moon Titan will be over a light hour away, and will need extensive autonomous intelligence to achieve anything.

These considerations strongly suggest that our space program is, or should be in large part joined with our other artificial intelligence and robotics activities. That is how we are going to be able to achieve great things in space, exploring far and wide to figure out how we came to be, what other worlds are like, and whether life arose on them as well. Robots can make themselves at home in the cosmos in a way that humans never will.

Matt Damon, accidentally marooned on Mars.

Bostrom's book naturally delves into our fate, once we have been comprehensively outclassed by our artificial creations. Will we be wiped out? Uploaded? Kept as pets? Who knows? But a reasonable deal might be that the robots get free reign to colonize the cosmos, spreading as far as their industry and inventiveness can carry them. But we'll keep earth, a home for a species that is bound to it by evolution, sentiment, and fate, and hopefully one that we can harness some of that intelligence to keep in a livable, even flourishing, condition.

• Some impressive mind reading.

Can Machines Read Yet?

Sort of, and not very well.

Reading- such a pleasure, but never time enough to read all that one would like, especially in technical fields. Scholars, even scientists, still write out their findings in prose- which is the richest form of communication, but only if someone else has the time and interest to read it. The medical literature is, at the flagship NCBI Pubmed resource, at about 30 million articles in abstract and lightly annotated form. Its partner, PMC, has 5.5 million articles in full text. This represents a vast trove of data which no one can read through, yet which tantalizes with its potential to generate novel insights, connections, and comprehensive and useful models, were we only able to harvest it in some computable form.

That is one of the motivations for natural language processing, or NLP, one of many subfields of artificial intelligence. What we learn with minimal effort as young children, machines have so far been unable to truly master, despite decades of effort and vast computational power. Recent advances in "deep learning" have made great progress in pattern parsing, and learning from large sets of known texts, resulting in the ability to translate one language to another. But does Google Translate understand what it is saying? Not at all. Understanding has taken strides in constricted areas, such as phone menu interactions, and Siri-like services. As long as the structure is simple, and has key words that tip off meaning, machines have started to get the hang of verbal communication.

But dealing with extremely complex texts is another matter entirely. NLP projects directed against the medical literature have been going on for decades, with relatively little to show, since the complexity of the corpus far outstrips the heuristics used to analyze it. These papers are, indeed, often very difficult for humans to read. They are frequently written by non-English speakers, or just bad writers. And the ideas being communicated are also complex, not just the language. The machines need to have a conceptual apparatus ready to accommodate, or better yet, learn within such a space. Recall how perception likewise needs an ever-expanding database / model of reality. Language processing is obviously a subfield of such perception. These issues raises a core question of AI- is general intelligence needed to fully achieve NLP?

I think the answer is yes- the ability to read human text with full understanding assumes a knowledge of human metaphors, general world conditions, and specific facts and relations from all areas of life which amounts to general intelligence. The whole point of NLP, as portrayed above, is not to spew audio books from written texts, (which is already accomplished, in a quite advanced way), but to understand what it is reading fully enough to elaborate conceptual models of the meaning of what those texts are about. And to do so in a way that can be communicated back to us humans in some form, perhaps diagrams, maps, and formulas, if not language.

The intensive study of NLP processing over the Pubmed corpus reached a fever pitch in the late 2000's, but has been quiescent for the last few years, generally for this reason. The techniques that were being used- language models, grammar, semantics, stemming, vocabulary databases, etc. had fully exploited the current technology, but still hit a roadblock. Precision could be pushed to ~ %80 levels for specific tasks, like picking out the interactions of known molecules, or linking diseases with genes mentioned in the texts. But general understanding was and remains well out of reach of these rather mechanical techniques. This is not to suggest any kind of vitalism in cognition, but only that we have another technical plateau to reach, characterized by the unification of learning, rich ontologies (world models), and language processing.

The new neural network methods (tensorflow, etc.) promise to provide the latter part of the equation, sensitive language parsing. But from what I can see, the kind of model we have of the world, with infinite learnability, depth, spontaneous classification capability, and related-ness, remains foreign to these methods, despite the several decades of work lavished on databases in all their fascinating iterations. That seems to be where more work is needed, to get to machine-based language understanding.

• What to do about media pollution?
• Maybe ideas will matter eventually in this campaign.
• Treason? Yes.
• Stalinist confessions weren't the only bad ones.
• Everything over-the-air ... the future of TV.

Backward and Forward... Steps to Perception

Perception takes a database, learning, and attention.

We all know by now that perception is more than simply being a camera, getting visual input from the world. Cameras see everything, but they recognize nothing, conceptualize nothing. Perception implies categorization of that input into an ontology that makes hierarchical sense of the world, full of inter-relationships that establish context and meaning. In short, a database is needed- one that is dynamically adaptive to allow learning to slice its model of reality into ever finer and more accurate categories.

How does the brain do that? The work of Karl Friston has been revolutionary in this field, though probably not well-enough appreciated and admittedly hard for me and others not practiced in mathematical statistics to understand. A landmark paper is "A theory of cortical responses", from 2005. This argues that the method of "Empirical Bayes" is the key to unlock the nature of our mental learning and processing. Bayesian statistics seems like mere common sense. The basic proposition is that the likelihood of some thing is related to our naive model (hypothesis) of its likelihood arrived at prior to any evidence or experience, combined with evidence expressed in a way that can weight or alter that model. Iterate as needed, and the model should improve with time. What makes this a statistical procedure, rather than simple common sense? If one can express the hypothesis mathematically, and the evidence likewise, in a way that relates to the hypothesis, then the evaluation and the updating from evidence can be done in a mechanical way.

Friston postulates that the brain is such a machine, which studiously models the world, engaging in what statisticians call "expectation maximization", which is to say, progressive improvements in the in detail and accuracy of its model, driven by inputs from sensory and other information. An interesting point is that sensory input functions really as feedback to the model, rather than the model functioning as an evaluator of the inputs. We live in the model, not in our senses. The overall mechanism works assiduously to reduce surprise, which is a measure of how inputs differ from the model. Surprise drives both attention and learning.

Another interesting point is the relationship between inference and learning. The model exists to perform inference- that is the bottom-up process of judging the reality and likely causes of some event based on the internal model, activated by the input-drive attention. We see a ball fall down, and are not surprised because our model is richly outfitted with calculations of gravitation, weight, etc. We infer that it has weight, and no learning is required. But suppose it is a balloon that floats up instead of falling- a novel experience? The cognitive dissonance represents surprise, which prompts higher-level processing and downward, top-down alterations to the model to allow for lighter-than-air weights. Our inferences about the causes may be incorrect. We may resort to superstition rather than physics for the higher-level inference or explanation. But in any case, the possibility of rising balls would be added to our model of reality, making us less surprised in the future.

The brain as a surprise-minimizing machine. Heading into old age, we are surprised by nothing, whether by great accumulated experience or by a closure against new experiences, and thus reach a stable / dead state. 

This brings up the physiology of what is going on in the brain, featuring specialization, integration, and recurrent networks with distinct mechanisms of bottom-up and top-down connection. Each sensory mode has its specialized processing system, with sub-modules, etc. But these only work by working together, both in parallel, cross-checking forms of integration, and by feeding into higher levels that integrate their mini-models (say for visual motion, or color assignment) into more general, global models.

"The cortical infrastructure supporting a single function may involve many specialized areas whose union is mediated by functional integration. Functional specialization and integration are not exclusive; they are complementary. Functional specialization is only meaningful in the context of functional integration and vice versa."

But the real magic happens thanks to the backward connections. Friston highlights a variety of distinctions between the forward and backward (recurrent) connections:

Forward connections serve inference, which is the primary job of the brain most of the time. They are regimented, sparsely connected, topographically organized, (like in the regular striations of the visual system). They are naturally fast, since speed counts most in making inferences. On the molecular level, forward connections use fast voltage-gated receptors, AMPA and GABA.

Backward connections, in contrast, serve learning and top-down modulation/attention. They are slow, since learning does not have to obey the rapid processing of forward signals. They tend to occupy and extend to complimentary layers of the cortex vs the forward connecting cells. They use NMDA receptors, which are roughly 1/10 as fast in response as the receptors use in forward synapses. They are diffuse and highly elaborated in their projections. And they extend widely, not as regimented as the forward connections. This allows lots of different later effects (i.e. error detection) to modulate the inference mechanism. And surprisingly, they far outnumber the forward connections:

"Furthermore, backward connections are more abundant. For example, the ratio of forward efferent connections to backward afferents in the lateral geniculate is about 1 : 10. Another distinction is that backward connections traverse a number of hierarchical levels whereas forward connections are more restricted."

Where does the backward signal come from, in principle? In the brain, error = surprise. Surprise expresses a violation of the expectation of the internal model, and is accommodated by a variety of responses. An emotional response may occur, such as motivation to investigate the problem more deeply. More simply, surprise would induce backward correction in the model that predicted wrongly, whether that is a high-level model of our social trust network, or something at a low level like reaching for a knob and missing it. Infants spend a great deal of time reaching, slowly optimizing their models of their own capabilities and the context of the surrounding world.

"Recognition is simply the process of solving an inverse problem by jointly minimizing prediction error at all levels of the cortical hierarchy. The main point of this article is that evoked cortical responses can be understood as transient expressions of prediction error, which index some recognition process. This perspective accommodates many physiological and behavioural phenomena, for example, extra classical RF [receptive field] effects and repetition suppression in unit recordings, the MMN [mismatch negativity] and P300 in ERPs [event-related potentials], priming and global precedence effects in psychophysics. Critically, many of these emerge from the same basic principles governing inference with hierarchical generative models."

This paper came up due to a citation from current work investigating this model specifically with non-invasive EEG methods. It is clear that the model cited and outlined above is very influential, if not the leading model now of general cognition and brain organization. It also has clear applications to AI, as we develop more sophisticated neural network programs that can categorize and learn, or more adventurously, develop neuromorphic chips that model neurons in a physical rather then abstract basis and show impressive computational and efficiency characteristics.

• Where is the outrage?
• Follow the money.
• And at work as well.
• Is America a full-on gerontocracy?
• Ideology and economics.
• On labor power.
• A more sanguine view of solar energy potential.
• The fascinating mechanics of opioid physiology.