kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026The paper appeared today in @natrevneuro.nature.com . Full text: rdcu.be/ah9QhWNMqG1ardcu.be 010
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026The paper just appeared in @natrevneuro.nature.com . Full text: rdcu.be/ah9QhWNMqG1ardcu.be 010
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026Making models disagree is a key challenge of computational neuroscience & takes on a new quality now that we are expressing our theories in NN models. “True” models may always be elusive, but we can make progress by comparing imperfect models with carefully designed experiments. 120
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026The small number of studies that have used optimized stimulus sets to adjudicate between models have taken heuristic approaches, which we also review in the paper. @talgolanneuro.bsky.social tweeted a summary... bsky.app/profile/talg... 220
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026A key consideration is what degree of certainty to assume about the parameters of our models. If we compare models with fixed parameters, model adjudication is easy, but we may reject models that are salvageable by minor reweighting of their representations. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026Explicitly or implicitly, our choices define a prior over stimuli that constrains the explanatory scope of model evaluation. Since our models are far from offering unified explanations, it’s important to adjust the scope of our explanatory ambition to make proximal progress. 120
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026The paper also explains how to make NN models disagree in practice. Beyond stimulus-computable models, we need to pick an optimization objective that quantifies how controversial a stimulus set is among the models, a stimulus search space and an optimization procedure. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026If we choose a stimulus set that achieves good discriminability even in the worst case, we will be able to discriminate all models no matter how confusable their parameters render them. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026For a given stimulus set, we can estimate the worst-case confusability for all pairs of models and corresponding parameter vectors. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026The frequentist perspective on optimal experimental design suggests a third approach: directly maximizing the discrepancy between the distributions predicted by different models. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026This motivates maximizing the model-recovery accuracy: The chance of finding the true model given the actual inferential method we will use for model comparison. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026It is often already too hard to do Bayesian inference about our models given a particular data set! If we won’t use Bayesian inference in analysis, designing our experiment to maximize the EIG is not well motivated: It doesn’t optimize our chances of finding the true model. 120
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026In practice, the Bayesian normative EIG is difficult to estimate because it requires an expectation over models, their parameters, and all possible data sets that our candidate stimulus set might generate. And that’s just for evaluating one candidate experiment. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026However, the outcome y is unknown and depends on the true model m and its parameters θm. The expected information gain (EIG) is therefore the expectation of u(𝛏, y) when the model, its parameters, and the resulting experimental outcome are drawn from our prior: 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026For a given stimulus set 𝛏 and experimental outcome y, the information gain is the entropy reduction of our belief distribution over models M: 120
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026If we have a prior over all possible computational models (model index m), we can maximize the reduction in uncertainty (= information gain) afforded by an experiment with stimulus set 𝛏. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026We approach the problem from first principles: the normative perspective of Bayesian optimal experimental design (BOED). BOED turns out not to be a realistic in practice in this context, but it serves as our north star. Let’s consider what would be required… 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026When we optimize a stimulus set to make the models disagree in their predictions, we can think of the stimulus set as controversial among the models. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026Since our goal is to compare competing models, we need experiments that elicit contrasting predictions from different models: critical experiments. This is a classical problem of experimental design, which we consider in the modern context where our hypotheses are instantiated in NN models. 110
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026Despite their parametric complexity, neural network models have distinct inductive biases that can be elicited with out-of-distribution probes. 120
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 30/08/2026To learn how brains compute, we need to experimentally adjudicate among competing computational hypotheses. How do we do this in the age of complex neural network models? New review paper: “Making models disagree to learn how brains compute” with @talgolanneuro.bsky.social @heikoschuett.bsky.social 1348
Reposted by @kriegeskorte.bsky.socialTal Golan @talgolanneuro.bsky.social · 28/08/2026How can we design experiments that make computational models disagree? One section of our new @natrevneuro.nature.com Review with @kriegeskorte.bsky.social and @heikoschuett.bsky.social examines studies that used stimulus sets designed to elicit distinct predictions from competing models. 1/16 114156
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 13/08/2026Videos of amazing workshop "Toward AI with Human Level Efficiency" organized by Ilker Yildirim et al. ... cnclgithub.github.io/yale-ai-work... with T. Brooke-Wilson, @nancykanwisher.bsky.social, @lisik.bsky.social, @joshtenenbaum.bsky.social, @toddgureckis.bsky.social, McCoy, Lew, and Kaelbling.cnclgithub.github.ioYale AI Workshop · Toward AI with Human-Level Efficiency 092
kriegeskorte.bsky.social @kriegeskorte.bsky.social · 09/08/2026In this podcast interview with speed cuber Sanjay Adireddi, he and I discuss an imaginary research program on how humans solve Rubik's cube... www.youtube.com/watch?v=m1uJ...youtube.comThe Computational Basis of Vision | Dr. Nikolaus KriegeskorteYouTube video by Neurocracy 0101
Reposted by @kriegeskorte.bsky.socialTal Golan @talgolanneuro.bsky.social · 10/07/2026I find agentic coding a powerful amplifier for teaching. This term, I taught with @simonprinceai.bsky.social 's excellent book, Understanding Deep Learning. The book's interactive figures are brilliant, but some students struggle to decipher surfaces presented top-down as 2D heat maps. 1/5 141
