There’s this paper that keeps cropping up in my little corner. It’s used as evidence that systems which are not typically regarded as agentic or intelligent can nonetheless exhibit cognitive properties (which is a bit surprising). The systems in question here are gene regulatory networks (GRNs) and this paper claims that they can exhibit associative memory (i.e., Pavlovian conditioning).
Gene regulatory networks are, well, networks which map out interactions between genes. They look sort of like this:
At least, this is what gene regulatory networks look like when they’re modeled as Boolean networks (which is a common way to model them). Basically, you have a set of genes (blue bubbles) which are either in state ON (1) or OFF (0) (i.e., actively being expressed or not), and interaction patterns which act on those genes (e.g., if gene X is ON then turn gene Y OFF). Here all that’s shown is inhibitory and excitatory actions, but in the paper they allow any Boolean function (AND, OR, NOT).
Biologists use these models to try to target, e.g., drug discovery. You can ask questions like, if I turn this gene off, what happens to the rest of the network? Does it stop producing, e.g., cortisol? In other words, they typically approach it from the bottom-up: trying to tweak the nodes or the connections between them to achieve large-scale outcomes.
This is hard when networks are very big (I’m not sure exactly how big they can get, but e.g., the human genome has ~20,000 genes, so presumably they can be pretty unwieldy). It presents what Levin (an author on this paper) has coined the inverse problem: “inferring how to reach desired system level states by manipulating individual node relationships.” In other words, in biological systems we want to make large-scale changes (e.g., curing cancer), but we often try to address this with specific, low-level changes. This sometimes works, but it’s difficult, especially as the system grows in size and complexity.
This paper proposes an alternative: instead of adjusting individual connections, why not figure out the right input to give to the network, and let it figure out the rest? For instance, if you could get GRNs to do associative learning, then you might be able to take an existing relationship (an unconditioned but toxic stimulus which causes a desired outcome) and train it to instead associate the toxic stimulus with a non-toxic one, thereby producing the desired outcome in an innocuous way.
So, can we? This paper claims that, yes, GRNs are capable of associative memory. I believe them, but I also… I don’t know, I feel like it’s not as interesting of a claim as they make it out to be. Let’s first investigate why they make the claim, though.
Step one is taking a bunch of already described GRNs from a publicly available dataset. They only choose networks which have 25 or fewer nodes. Then they find all possible UCS (unconditioned stimulus), R (response), and NS (neutral stimulus) combinations. They do this by starting off each GRN with all inputs set to 0, and then let it evolve into a steady state. Then they set the UCS to 1. If R also becomes 1, then this is a candidate. Then they set NS to 1. If R stays at 0, then this whole set of UCS, NS, R is considered a possible memory configuration.
They take all of these, and then “train” the networks. To do this they set UCS and NS to 1 at the same time. Then they relax the network (stop forcing these nodes to be on), and let the network evolve. Then they set NS to 1 (UCS is not altered). If this causes R to be set to 1, they consider it to have learned the association (i.e., to have paired NS to UCS). If this persists (i.e., if they relax the network, and then set NS to 1 again, and it still results in R=1), then they consider it long-term memory.
There were several other memory types the paper addressed, but I find them all pretty uninteresting relative to associative learning. For instance, they consider any UCS causing R to turn on (and to stay on even after the UCS has disappeared) to be a form of memory. This seems… true, but just way less cool than associative learning.
They also go on to show things like how knowing that a GRN is from a vertebrate or multicellular organism predicts its ability to have memory. And how biological networks exhibit more capacity for memory than randomly constructed Boolean networks of the same size (i.e., they only differ in the Boolean functions acting on the nodes). I thought this was sort of interesting.
But why do I not feel that excited? I’m not totally sure but here are some impressions. For one, figuring out what inputs to put into a system seems almost as hard to me as tweaking connections? Like, you’re still running into combinatorial explosions with bigger networks, so I don’t see where the gain is coming from.
Secondly, it definitely seems cool that GRNs exhibit Pavlovian conditioning, I think, but I’m left without really any understanding of what the implications are. It seems like most GRNs are capable of it, but does that mean they’re actually doing it? Why would they? What’s the advantage?
Finally, I think something about the way it was written read to me as sort of disorienting and confusing. It’s one of those papers where you really only get it if you dig into the supplemental material, and the claims seemed to outpace the empirical support, and they did all of these tests which to me read like “beefing up the paper and making it seem thorough and quantitative” without really adding much value. I wanted a much deeper dive into associative learning and why that was interesting and important relative to, e.g., how prevalent all the different so-called memory types are in different taxa.
I think one implication that Levin takes from this paper is that figuring out the cognitive/agentic capacity of systems is an empirical question that can be addressed by tests like this, and that we shouldn’t jump to conclusions before we try. I’m sympathetic to this, but I also… I don’t know, despite believing the results, I somehow don’t feel robustly convinced of the overall point that GRNs are cognitive. I feel like the thing that’s lacking for me here is any tie to the “real world,” like alright—you can train a Boolean network to do associative learning. Does that mean that GRNs are actually doing that? What does that look like in, e.g., a human being? Can we do anything interesting with it?
If something like this were implemented in real tissue, or something like it, I’d be much more excited. I think at this point the paper seems more like “you can squeeze whatever you want out of a model if you try hard enough”—not that they aimed to do that or anything, just that they made up somewhat reasonable but also sort of arbitrary decisions in their algorithm, and chose an algorithm which has already been shown to have this sort of memory, and maybe Boolean networks are good models of genes, but maybe they aren’t, and anyways, there seems like many steps between this and feeling more robustly convinced.