iiinova — where your digital self keeps evolving

Rui Ponte Costa @somnirons.bsky.social · 03/08/2026

While I am sad to miss #CCN2026 our group will showing the cool stuff that we have been doing in the model-brain alignment space, including a selected talk by Angeliki! See list, including our fun gaming collaboration with the groups of @marcelomattar.bsky.social and @momchiltomov.bsky.social

0101

Rui Ponte Costa @somnirons.bsky.social · 18/06/2026

7/9 The theory makes a further prediction that interneurons should constrain the dimensionality (rank) of error-related feedback: frontal regions need more SST interneurons! Offering a functional rationale for cortex-wide gradients in interneuron density (Kim et al. 2017).

140

Rui Ponte Costa @somnirons.bsky.social · 18/06/2026

6/9 Importantly, it unifies: interneuron-specific modulation of synaptic plasticity (Williams and Holtmaat 2019), differential learning in interneurons (Chevy et al. bioRxiv), and the very recent findings of neuron-specific dendritic error signals in a BCI task (Francioni et al. 2026).

120

Rui Ponte Costa @somnirons.bsky.social · 18/06/2026

5/9 Next, by mapping these principles to a Dalean network it predicts interneuron-specific event decoding, with SST burst decoding at its heart.

120

Rui Ponte Costa @somnirons.bsky.social · 18/06/2026

4/9 As expected, our model also approximates well ANN performance in more complex tasks such as dynamic naturalistic tasks (without labels), CIFAR-10/Imagenet and deep reinforcement learning.

120

Rui Ponte Costa @somnirons.bsky.social · 18/06/2026

3/9 For exact EI balance our model (cian) gives a near-exact approximation of ANN backprop with relaxations requiring periods of exc-inh rebalancing (purple and blue), through inhibitory-like plasticity (in line w/ Froemke, Vogels, etc.).

121

Rui Ponte Costa @somnirons.bsky.social · 18/06/2026

2/9 We have previously shown that these principles enable single-phase deep learning (Greedy et al. 2022). Here we further show that our theory predicts the need for a tight balance between QY feedback pathways (exc-inh balance at distal dendrites).

130

Rui Ponte Costa @somnirons.bsky.social · 18/06/2026

1/9 We propose differential cortical feedback pathways: one for decoding any spiking events (blue) and another for burst spike events (orange). A new task will generate errors that cause changes in bursts events that are then “back propagate” across the cortex to drive synaptic plasticity.

131

Rui Ponte Costa @somnirons.bsky.social · 29/11/2025

Cool work! I should clarify that our cortical SSL model (Nejad 2025) also predicts in the latent space like JEPA. Indeed the mapping to JEPA is relative straigtforward (L4 and L5 are encoders and L2/3 is the predictive network).

170

Rui Ponte Costa @somnirons.bsky.social · 12/08/2025

Sorry to be missing CCN'25!! But if you are around come and check-out some of our work in the cogcompneuro space! @cogcompneuro.bsky.social @achterbrain.bsky.social @Michal, @Angeliki and @Austin

1102

Rui Ponte Costa @somnirons.bsky.social · 11/07/2025

Sharing this very cool drawing that Kevin Nejad generated using AI systems that have themselves been trained via self-supervision, which seems very fitting :) www.nature.com/articles/s41...

180

Rui Ponte Costa @somnirons.bsky.social · 28/04/2025

Reminder that deadline for applying for our Encode: AI for (neuro)Science fellowship is coming up (30th!) encode.pillar.vc/apply

063

Rui Ponte Costa @somnirons.bsky.social · 24/03/2025

Looking for an exciting fellowship in neuroAI, with competitive salary (~£100k)? Our group is hiring for a project on AI for Systems-Behavioral Neuroscience, more info here: encode.pillar.vc/projects/beh... General info: encode.pillar.vc 🧪 #compneuro #neuroai #neuroscience #sciencejobs

0307

Rui Ponte Costa @somnirons.bsky.social · 13/01/2025

Ever wondered what different layer-5 PC types do for learning? Our work suggests that one (IT PCs) does representational learning whereas the other (ET PCs) encodes representational value! A great exp-theory collaboration with the Larkum and Takahashi's labs!

14016

Rui Ponte Costa @somnirons.bsky.social · 02/01/2025

🚨Our paper on how the cerebellum learns to drive cortical dynamics for rapid task learning and switching, which we propose can then be consolidated in the cortex @naturecomms.bsky.social nature.com/articles/s41... 🧠 #compneuro

14915