Thanks, Joao, for the shout out!

We do now have another working memory task for mice, which is a 2-step sound categorisation -- for WM intervals beyond 1.5sec, we see sound to category transfer where they hold a noisy memory of category in mind, dissociated from future action.

Check out the preprint for a lot more results.

A big shout out to co-first authors: @elenameni.bsky.social, Quentin Pajot-Moric, Ryan Low (listed alphabetically based on first name) & Amirali Pourdehghan, Victor Pedrosa, Peter Vincent, Lillianne Teachen & Liang Zhou.
www.biorxiv.org/content/10.1...

We believe this work provides the first systematic, cross-species measure of generative & discriminative learning & establishes a general framework for linking normative theory, learning algorithms, & behaviour across species.

Animals may rely on rich generative models in natural behaviour, but our results show that in lab tasks they can reach the same optimal outcome through v diff algorithms. Recognising this diversity is crucial for interpreting neural data & for understanding what diff brain areas actually compute

Cross-validated model comparison shows that humans predominantly relied on generative representations, mice on discriminative boundary-tracking, and rats spanned both regimes. Moreover, all species achieved statistical adaptation with similar efficiency.

Both models can produce similar final behaviour & match a normative observer at steady state. But they make different predictions for trial-by-trial choice updates.

We contrasted two models:
• Stimulus–Category (generative): agents update beliefs about the distribution of sensory inputs within each category.
• Boundary–Estimation (discriminative): agents update only their belief about the category boundary location, ignoring within-category structure.

The key insight: end-point performance can’t reveal the learning algorithm; every species looked near-optimal at steady state.
To uncover the true strategy, we analysed learning dynamics: how each choice shifted trial-by-trial after feedback. These temporal signatures turned out to be diagnostic.

All species adapted behaviour near-optimally, consistent with a normative observer constrained by sensory and decision noise.
Look only at final performance, and all species seem to behave the same.

We trained humans, rats, & mice on the same auditory categorisation task, holding category boundaries and rewards constant while varying sensory statistics (ie sound distribution within each category).
This isolates how each species learns from sensory structure, not from reward or decision biases.

Key distinction:
Generative learners build internal models of sensory world, capturing input distributions, supporting richer inferences (if it’s a cat, it likely has claws).
Discriminative learners learn only the stimulus-to-choice mapping, effective for classification but blind to deeper structure

Oh nice!! So happy to have you here in London with us! Welcome to UCL!! ❤️🥳

Thanks, Matteo! Will do.

Folks in the UK who want to help may want to write to their MP. The list is at members.parliament.uk/constituencies. (I wrote to mine).

And read the excellent post by the amazing science communication team at @sainsburywellcome.bsky.social, explaining "Heron the progeny of Bonsai and ROS".

15/
sainsburywellcome.org/web/blog/dev...

Read George's blog post on why/how he developed Heron:
python.plainenglish.io/heron-a-hybr...

14/

The code for all repositories can be found on GitHub:
github.com/Heron-Reposi...

Access Heron’s documentation here:
heron-42ad.readthedocs.io/en/latest/

13/