When I talk about "causal salad", this is what I mean: no consideration of how the covariates relate to one another or the treatment.

Awesome finale that @JTKerby's image made the @AnimalEcology cover in support of our method using drones to study animals in their natural social and physical landscapes! @adwait_d @jgraving @BlairRCostelloe @icouzin great team!

I'm really proud of this work published today in Proc B with @SmithBeeLab & @ben_koger on the importance of the 3D nest structure and building strategies in developing honeybee colonies https://t.co/fAVEUUw83B We observe, manipulate, and model 3D nest construction - see below.

The #ImagingHangar @UniKonstanz is abuzz with the sound of 60k #locust feet. More than 4k locusts have been tagged with reflective markers for tracking with the Motion Capture System. Researchers from @CBehav and @MPI_animalbehav aim to understand the behaviour of locust swarms.

Out now in @AnimalEcology! A general approach for using drones to study animal behavior in the wild. Record the location and posture of many animals simultaneously at sub-second sub-meter resolution, plus reconstruct their 3D landscape: https://t.co/2u4jtEzha2

Our drone-based method for tracking the (geo-referenced) location and body-postures of free-roaming animals, including 3D landscape models and social context - out now! https://t.co/Hcb3E46Bm5 @ben_koger @BlairRCostelloe @CollectiveBehav @MPI_animalbehav @UniKonstanz @CBehav

New blog post: Collective Intelligence for Deep Learning Recently, @yujin_tang and I published a paper about how ideas like swarm behavior, self-organization, emergence are gaining traction in deep learning. I wrote a blog post summarizing the key ideas: https://t.co/S644KjM20e

The #CASCB is super excited: Currently we are running an experiment on locust swarms in the #ImagingHangar @UniKonstanz. Normally #locusts are studied in the lab in small groups of 200 animals in small arenas despite swarming in groups of millions of individuals in the wild.

Come join our team as one of 12 PhD students in the new #WildDrone network! Visit https://t.co/UZVQLrJeNr for project descriptions and application information

Happy to announce the final release of seaborn 0.12.0, a major update with new features that I'm really excited about. Check out the highlights: https://t.co/RMoE2k7rkH Read the full release notes: https://t.co/KUe1j1okNw pip install seaborn==0.12.0 I hope you find it useful!

If you have behavioral videos and want to try out keypoint discovery, we've open-sourced B-KinD: https://t.co/iL7CjcSACr You can train and run B-KinD on videos without human annotations! Thanks to Serim Ryou for working with me on the code😊 Let us know if you have questions!

Happy to announce that our Python package for active inference, pymdp, is now published in JOSS @JOSS_TheOJ: "pymdp: A Python library for active inference in discrete state spaces" Paper: https://t.co/iZg8fbevln Code: https://t.co/Hhzsh1wOv5 Docs:

*should think carefully

To summarize, off-the-shelf tools typically make bad assumptions about behavioral data, but methods exist to avoid falling into *some* but not *all* of these logical traps. Science is incredibly hard, but we think carefully about the models we're using to make inferences

So, then what's the solution? To make worthwhile inferences, we need methods that take causal relationships into account. For example, Shapley values, the explainability tool used in the above paper has already been adapted for causal inference:

Now you might be thinking "Well, that model is still missing many important features. It is also wrong." And you would be exactly right! But causal inference is not so much a method for finding the exact "true" model, but a process for eliminating *obviously wrong* models.

Or written out as a distribution: p(behavior | speed, turning, length, weight) p(speed | turning, length, weight) p(turning | speed) p(length) p(weight | body length)

We know, in fact, these variables are not just correlated but are actually *conditional* on one another in the sense that one variable may be *causing* the value of another. If we were to take these relationships into account, our DAG might look something more like:

Based on our prior knowledge of the real world, we know this is a *bad* assumption, which can quickly lead to equally bad inferences. A situation @rlmcelreath refers to as "causal salad" (also, everyone should read his Rethinking v2 book)

This assumption is also baked into the SHAP method proposed in the paper above. Which the developers discuss here: