Cross-species RSA: same learning rules (BP, PC, STDP, FA) tested against both human fMRI and macaque electrophysiology [P]
Our take
The recent study on cross-species Representational Similarity Analysis (RSA) presents a compelling exploration of learning rules applied to both human fMRI data and macaque electrophysiological measurements. By testing five learning paradigms—Backpropagation (BP), Predictive Coding (PC), Spike-Timing-Dependent Plasticity (STDP), and others—the research offers a nuanced understanding of how early visual processing aligns across species. Notably, the findings suggest that certain learning rules, particularly STDP and PC, maintain a qualitative consistency in macaque visual areas V1/V2, echoing patterns observed in human brains. This alignment is significant because it challenges the notion that findings from fMRI studies are artifacts, reinforcing the validity of using diverse methodologies in neuroscience.
Understanding these cross-species similarities not only enhances our grasp of basic visual processing but also has broader implications for the development of AI systems that mimic human learning. As we explore frameworks for building more intelligent machines, insights from studies like this could inform the design of algorithms that better emulate human-like learning processes. For instance, the exploration of learning rules in neural networks could benefit from findings that emphasize the effectiveness of STDP and PC in information processing. The ongoing dialogue surrounding AI development is further enriched by related discussions, such as those presented in Profiling PyTorch training without accidentally stalling the GPU and Presentation: Designing AI Platforms for Reliability: Tools for Certainty, Agents for Discovery.
However, the findings also raise important questions about the limitations of the study. The discrepancies observed in untrained baseline results between human fMRI and macaque electrophysiology highlight the potential confounding effects of different stimulus sets used in the experiments. The noted inversion in IT rankings suggests that our understanding of visual processing may be influenced more by the nature of stimuli than by the learning rules themselves. This emphasizes the need for caution in interpreting cross-species results and reminds us that further exploration is required to disentangle these complexities.
As we move forward, it is essential to consider how these insights can be applied not just in neuroscience but also in practical AI applications. The implications of this research extend beyond academic curiosity; they touch on the very fabric of how we develop intelligent systems capable of nuanced understanding. The challenge will be to integrate these findings into a framework that allows for the effective training of AI models, ensuring they are not only capable of processing information but also of learning in ways that reflect human-like adaptability. The future of AI development is bright, but it hinges on our ability to translate biological insights into practical tools. How will we take these foundational findings and shape the next generation of intelligent systems? That remains a question worth watching in the evolving landscape of AI and neuroscience.
Follow-up to my earlier post on learning rules vs. human fMRI. Same five conditions (BP, FA, PC, STDP, untrained), same model weights, now evaluated against macaque V1/V2 (FreemanZiemba2013, single-unit) and macaque V4/IT (MajajHong2015, multi-electrode).
Main findings:
- Early visual alignment is qualitatively conserved across species. STDP (ρ ≈ 0.30) and PC (ρ ≈ 0.28) lead at macaque V1/V2, consistent with their position in human V1. The pattern isn't an fMRI artifact.
- The untrained baseline result doesn't replicate cleanly. In human fMRI, Random ≥ BP at V1. In macaque, STDP and PC pull ahead of Random (electrophysiology has enough SNR to resolve the difference fMRI can't).
- IT alignment scales with capacity, not learning rule. ResNet-50 (pretrained, ImageNet): ρ ≈ 0.25 at macaque IT. Custom 3-conv CNN across all learning rules: ρ = 0.07–0.14. The IT convergence from the companion paper looks like a capacity floor.
- Cross-species IT rankings: Kendall's τ = 0.00 (p = 1.00) but n = 5 only has power at τ = ±1.0, so this is uninformative rather than evidence of non-conservation.
Limitations worth noting:
- V1/V2 and V4/IT come from different macaque datasets with different stimulus sets (textures vs. objects): the V2→V4 drop is confounded by this switch
- Stimulus control shows IT rankings are weakly inverted across stimulus sets (τ = −0.40), so cross-species IT differences may be partially stimulus-driven
Companion paper: arxiv.org/abs/2604.16875
Cross-species paper: https://arxiv.org/abs/2605.22401
Code: github.com/nilsleut/cross-species-rsa
Happy to discuss the stimulus confound issue or the capacity control in more detail.
[link] [comments]
Read on the original site
Open the publisher's page for the full experience