Elements of biological computation & stochastic thermodynamics of life

This week, I was visiting the Santa Fe Institute for a workshop organized by Albert Kao, Jessica Flack, and David Wolpert on “What is biological computation?” (11 – 13 September 2019). It was an ambitious question and I don’t think that we were able to answer it in just three days of discussion, but I think that we all certainly learnt a lot.

At least, I know that I learned a lot of new things.

The workshop had around 34 attendees from across the world, but from the reaction on twitter it seems like many more would have been eager to attend also. Hence, both to help synchronize the memory networks of all the participants and to share with those who couldn’t attend, I want to use this series of blog post to jot down some of the topics that were discussed at the meeting.

During the conference, I was live tweeting. So if you prefer my completely raw, unedited impressions in tweet form then you can take a look at those threads for Wednesday (14 tweets), Thursday (15 tweets), and Friday (31 tweets). The workshop itself was organized around discussion, and the presentations were only seeds. Unfortunately, my live tweeting and this post are primarily limited to just the presentations. But I will follow up with some synthesis and reflection in the future.

Due to the vast amount discussed during the workshop, I will focus this post on just the first day. I’ll follow with posts on the other days later.

It is also important to note that this is the workshop through my eyes. And thus this retelling is subject to the limits of my understanding, notes, and recollection. In particular, I wasn’t able to follow the stochastic thermodynamics that dominated the afternoon of the first day. And although I do provide some retelling, I hope that I can convince one of the experts to provide a more careful blog post on the topic.

Read more of this post


Rationality, the Bayesian mind and their limits

Bayesianism is one of the more popular frameworks in cognitive science. Alongside other similar probalistic models of cognition, it is highly encouraged in the cognitive sciences (Chater, Tenenbaum, & Yuille, 2006). To summarize Bayesianism far too succinctly: it views the human mind as full of beliefs that we view as true with some subjective probability. We then act on these beliefs to maximize expected return (or maybe just satisfice) and update the beliefs according to Bayes’ law. For a better overview, I would recommend the foundations work of Tom Griffiths (in particular, see Griffiths & Yuille, 2008; Perfors et al., 2011).

This use of Bayes’ law has lead to a widespread association of Bayesianism with rationality, especially across the internet in places like LessWrong — Kat Soja has written a good overview of Bayesianism there. I’ve already written a number of posts about the dangers of fetishizing rationality and some approaches to addressing them; including bounded rationality, Baldwin effect, and interface theory. I some of these, I’ve touched on Bayesianism. I’ve also written about how to design Baysian agents for simulations in cognitive science and evolutionary game theory, and even connected it to quasi-magical thinking and Hofstadter’s superrationality for Kaznatcheev, Montrey & Shultz (2010; see also Masel, 2007).

But I haven’t written about Bayesianism itself.

In this post, I want to focus on some of the challenges faced by Bayesianism and the associated view of rationality. And maybe point to some approach to resolving them. This is based in part of three old questions from the Cognitive Sciences StackExhange: What are some of the drawbacks to probabilistic models of cognition?; What tasks does Bayesian decision-making model poorly?; and What are popular rationalist responses to Tversky & Shafir?

Read more of this post

Web of C-lief: conjectures vs. model assumptions vs. scientific beliefs

Web of C-lief with the non-contradiction spider

A sketch of the theoretical computer science Web of C-lief weaved by the non-contradiction spider.

In his 1951 paper on the “Two Dogmas of Empiricism”, W.V.O Quine introduced the Web of Belief as a metaphor for his holistic epistemology of scientific knowledge. With this metaphor, Quine aimed to give an alternative to the reductive atomising epistemology of the logical empiricists. For Quine, no “fact” is an island and no experiment can be focused in to resole just one hypothesis. Instead, each of our beliefs forms part of an interconnected web and when a new belief conflicts with an existing one then this is a signal for us to refine some belief. But this signal does not unambiguously single out a specific belief that we should refine. Just a set of beliefs that are incompatible with out new one, or that if refined could bring our belief system back into coherence. We then use alternative mechanisms like simplicity or minimality (or some aesthetic consideration) to choose which belief to update. Usually, we are more willing to give up beliefs that are peripheral to the web — that are connected to or change fewer other beliefs — than the beliefs that are central to our web.

In this post, I want to play with Quine’s web of belief metaphor in the context of science. This will force us to restrict it to specific domains instead of the grand theory that Quine intended. From this, I can then adapt the metaphor from belief in science to c-liefs in mathematics. This will let me discuss how complexity class seperation conjectures are structured in theoretical computer science and why this is fundamentally different from model assumptions in natural science.

So let’s start with a return to the relevant philosophy.

Read more of this post

Idealization vs abstraction for mathematical models of evolution

This week I was in Turku, Finland for the annual congress of the European Society for Evolutionary Biology. I presented in the symposium on mathematical models in evolutionary biology organized by Guy Cooper, Matishalin Patel, Tom Scott, and Asher Leeks. It was a fun. It was also a big challenge given the short ten minute format. I decided to use my ten minutes to try to convince the audience that we should consider not just idealized models but also abstractions. So after my typical introduction of computational vs algorithmic biology, I switched to talking about triangles. If you would like, dear reader, then you can watch the whole session online (or grab my slides as pdf). In this post, I just want to focus on the distinction between idealized vs. abstract models.

Just as in my ESEB talk, I’ll use triangles to explain the distinction between idealized vs. abstract models.

Read more of this post

Process over state: Math is about proofs, not theorems.

A couple of days ago, Maylin and I went to pick blackberries along some trails near our house. We spent a number of hours doing it and eventually I turned all those berries into one half-litre jar of jam.

On the way to the blackberry trails, we passed a perfectly fine Waitrose — a supermarket that sells (among countless other things) jam. A supermarket I had to go to later anyways to get jamming sugar. Why didn’t we just buy the blackberries or the jam itself? It wasn’t a matter of money: several hours of our time picking berries and cooking them cost much more than a half-litre of jam, even from Waitrose.

I think that we spent time picking the berries and making the jam for the same reason that mathematicians prove theorems.

Imagine that you had a machine where you put in a statement and it replied with perfect accuracy if that statement was true or false (or maybe ill-posed). Would mathematicians welcome such a machine? It seems that Hilbert and the other formalists at the start of the 20th century certainly did. They wanted a process that could resolve any mathematical statement.

Such a hypothetical machine would be a Waitrose for theorems.

But is math just about establishing the truth of mathematical statements? More importantly, is the math that is written for other mathematicians just about establishing the truth of mathematical statements?

I don’t think so.

Math is about ideas. About techniques for thinking and proving things. Not just about the outcome of those techniques.

This is true of much of science and philosophy, as well. So although I will focus this post on the importance of process over state/outcome in pure math, I think it can also be read from the perspective of process over state in science or philosophy more broadly.

Read more of this post

Generating random power-law graphs

‘Power-law’ is one of the biggest buzzwords in complexology. Almost everything is a power-law. I’ve even used it to sell my own work. But most work that deals in power-laws tends to lack rigour. And just establishing that something is a power-law shouldn’t make us feel that it is more connected to something else that is a power-law. Cosma Shalizi — the great critic of sloppy thinking in complexology — has an insightful passage on power-laws:

[T]here turn out to be nine and sixty ways of constructing power laws, and every single one of them is right, in that it does indeed produce a power law. Power laws turn out to result from a kind of central limit theorem for multiplicative growth processes, an observation which apparently dates back to Herbert Simon, and which has been rediscovered by a number of physicists (for instance, Sornette). Reed and Hughes have established an even more deflating explanation (see below). Now, just because these simple mechanisms exist, doesn’t mean they explain any particular case, but it does mean that you can’t legitimately argue “My favorite mechanism produces a power law; there is a power law here; it is very unlikely there would be a power law if my mechanism were not at work; therefore, it is reasonable to believe my mechanism is at work here.” (Deborah Mayo would say that finding a power law does not constitute a severe test of your hypothesis.) You need to do “differential diagnosis”, by identifying other, non-power-law consequences of your mechanism, which other possible explanations don’t share. This, we hardly ever do.

The curse of this multiple-realizability comes up especially when power-laws intersect with the other great field of complexology: networks.

I used to be very interested in this intersection. I was especially excited about evolutionary games on networks. But I was worried about some of the arbitrary seeming approaches in the literature to generating random power-law graphs. So before starting any projects with them, I took a look into my options. Unfortunately, I didn’t go further with the exploration.

Recently, Raoul Wadhwa has gone much more in-depth in his thinking about graphs and networks. So I thought I’d share some of my old notes on generating random power-law graphs in the hope that they might be useful to Raoul. These notes are half-baked and outdated, but maybe still fun.

Hopefully, you will find them entertaining, too, dear reader.

Read more of this post

Closing the gap between quantum and deterministic query complexity for easy to certify total functions

Recently, trying to keep with my weekly post schedule, I’ve been a bit strapped for inspiration. As such, I’ve posted a few times on a major topic from my past life: quantum query complexity. I’ve mostly tried to describe some techniques for (lower) bounding query complexity like the negative adversary method and span programs. But I’ve never really showed how to use these methods to actually set up interesting bounds.

Since I am again short of a post, I thought I’d share this week a simple proof of a bound possible with these techniques. This is based on an old note I wrote on 19 April 2011.

One of the big conjectures in quantum query complexity — at least a half decade ago when I was worrying about this topic — is that quantum queries give you at most a quadratic speedup over deterministic queries for total functions. In symbols: D(f) = O(Q^2(f)). Since Grover’s algorithm can give us a quadratic quantum speed-up for arbitrary total functions, this conjecture basically says: you can’t do better than Grover.

In this post, I’ll prove a baby version of this conjecture.

Let’s call a Boolean total-function easy to certify if one side of the function has a constant-length certificate complexity. I’ll prove that for easy-to-certify total functions, D(f) = O(Q^2(f)).

This is not an important result, but I thought it is a cute illustration of standard techniques. And so it doesn’t get lost in my old pdf, I thought I’d finally convert it to a blog post. Think of this as a simple application of the adversary method.

Read more of this post

The gene-interaction networks of easy fitness landscapes

Since evolutionary fitness landscapes have been a recurrent theme on TheEGG, I want to return, yet again, to the question of finding local peaks in fitness landscapes. In particular, to the distinction between easy and hard fitness landscapes.

Roughly, in easy landscapes, we can find local peaks quickly and in hard ones, we cannot. But this is very vague. To be a little more precise, I have to borrow the notion of orders of growth from the asymptotic analysis standard in computer science. A family of landscapes indexed by a size n (usually corresponding to the number of genes in the landscape) is easy if a local fitness optimum can be found in the landscapes in time polynomial in n and hard otherwise. In the case of hard landscapes, we can’t guarantee to find a local fitness peak and thus can sometimes reason from a state of perpetual maladaptive disequilibrium.

In Kaznatcheev (2019), I introduced this distinction to biology. Since hard landscapes have more interesting properties which are more challenging to theoretical biologist’s intuitions, I focused more on this. This was read — perhaps rightly — as me advocating for the existence or ubiquity of hard landscapes. And that if hard landscapes don’t occur in nature then my distinction is pointless. But I don’t think this is the most useful reading.

It certainly would be fun if hard landscapes were a feature of nature since they give us a new way to approach certain puzzles like the maintenance of cooperation, the evolution of costly learning, or open-ended evolution. But this is an empirical question. What isn’t a question is that hard landscape are a feature of our mental and mathematical models of evolution. As such, all — or most, whatever that means — fitness landscapes being easy is still exciting for me. It means that the easy vs hard distinction can push us to refine our mental models such that if only easy landscapes occur in nature then our models should only be able to express easy landscapes.

In other words, using computational complexity to build upper-bounds arguments (that on certain classes of landscapes, local optima can be found efficiently) can be just as fun as lower-bounds arguments (that on certain classes of landscapes, evolution requires at least a super-polynomial effort to find any local fitness peak). However, apart from a brief mention of smooth landscapes, I did not stress the upper-bounds in Kaznatcheev (2019).

Now, together with David Cohen and Peter Jeavons, I’ve taken this next step — at least in the cstheory context, we still need to write on the biology. So in this post, I want to talk briefly about a biological framing of Kaznatcheev, Cohen & Jeavons (2019) and the kind of fitness landscapes that are easy for evolution.

Read more of this post

Description before prediction: evolutionary games in oncology

As I discussed towards the end of an old post on cross-validation and prediction: we don’t always want to have prediction as our primary goal, or metric of success. In fact, I think that if a discipline has not found a vocabulary for its basic terms, a grammar for combining those terms, and a framework for collecting, interpreting, and/or translating experimental practice into those terms then focusing on prediction can actually slow us down or push us in the wrong direction. To adapt Knuth: I suspect that premature optimization of predictive potential is the root of all evil.

We need to first have a good framework for describing and summarizing phenomena before we set out to build theories within that framework for predicting phenomena.

In this brief post, I want to ask if evolutionary games in oncology are ready for building predictive models. Or if they are still in need of establishing themselves as a good descriptive framework.

Read more of this post

Fighting about frequency and randomly generating fitness landscapes

A couple of months ago, I was in Cambridge for the Evolution Evolving conference. It was a lot of fun, and it was nice to catch up with some familiar faces and meet some new ones. My favourite talk was Karen Kovaka‘s “Fighting about frequency”. It was an extremely well-delivered talk on the philosophy of science. And it engaged with a topic that has been very important to discussions of my own recent work. Although in my case it is on a much smaller scale than the general phenomenon that Kovaka was concerned with,

Let me first set up my own teacup, before discussing the more general storm.

Recently, I’ve had a number of chances to present my work on computational complexity as an ultimate constraint on evolution. And some questions have repeated again and again after several of the presentations. I want to address one of these persistent questions in this post.

How common are hard fitness landscapes?

This question has come up during review, presentations, and emails (most recently from Jianzhi Zhang’s reading group). I’ve spent some time addressing it in the paper. But it is not a question with a clear answer. So unsurprisingly, my comments have not been clear. Hence, I want to use this post to add some clarity.

Read more of this post