{"id":8274,"date":"2026-06-20T08:16:38","date_gmt":"2026-06-20T07:16:38","guid":{"rendered":"https:\/\/sinatootoonian.com\/?post_type=notes&#038;p=8274"},"modified":"2026-06-20T09:20:13","modified_gmt":"2026-06-20T08:20:13","slug":"alex-pouget-talk","status":"publish","type":"notes","link":"https:\/\/sinatootoonian.com\/index.php\/notes\/alex-pouget-talk\/","title":{"rendered":"Alex Pouget Talk"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li>Compositionality = parts + syntax<\/li>\n\n\n\n<li>First part: 2024 paper\n<ul class=\"wp-block-list\">\n<li>System: RNN <\/li>\n\n\n\n<li>Inputs: \n<ul class=\"wp-block-list\">\n<li>Sensory signals, <\/li>\n\n\n\n<li>Text input passed through LLM<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Outputs:\n<ul class=\"wp-block-list\">\n<li>Motor outputs: fixation, decision<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Training:\n<ul class=\"wp-block-list\">\n<li>Tell it what to do with text in various ways.<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Finding:\n<ul class=\"wp-block-list\">\n<li>Could zero-shot new tasks specified through text.<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Explanation:\n<ul class=\"wp-block-list\">\n<li>RNN learns a compositional representation\n<ul class=\"wp-block-list\">\n<li><strong>Question:<\/strong> Is that an inductive bias of the RNN?<\/li>\n\n\n\n<li><em>Observation<\/em>: Dictionary learning\/sparse coding.\n<ul class=\"wp-block-list\">\n<li><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Question:<\/strong> Is that induced by the text?\n<ul class=\"wp-block-list\">\n<li>Finding: Could train the network using a binary vector indicating the task. But networks couldn&#8217;t generalise.\n<ul class=\"wp-block-list\">\n<li>Why not?<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Modules are gated by the text.<\/li>\n\n\n\n<li>A related experiment with 3 binary variables\n<ul class=\"wp-block-list\">\n<li>Forcing system to search on the 3D hypercube of possibilities dramatically improved learning time.\n<ul class=\"wp-block-list\">\n<li><strong>Question:<\/strong> How would the system learn to do this?<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Second part: Paper with Tiago\n<ul class=\"wp-block-list\">\n<li>System: mice performing escape behaviour<\/li>\n\n\n\n<li>Inputs: \n<ul class=\"wp-block-list\">\n<li>arena with wall in the  middle, <\/li>\n\n\n\n<li>shelter point,<\/li>\n\n\n\n<li>Occasional loud noise<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Outputs: Mouse trajectories<\/li>\n\n\n\n<li>Model: Motor primitives\n<ul class=\"wp-block-list\">\n<li>Figure out which primitives to combine to build the trajectory.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Questions:<\/strong>\n<ul class=\"wp-block-list\">\n<li>Similarity to dictionary learning<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Chat afterwards:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>PPCs:\n<ul class=\"wp-block-list\">\n<li><span style=\"text-decoration: underline;\">To read<\/span>: Deneve, Latham, Pouget (2001): \n<ul class=\"wp-block-list\">\n<li>How to map PPC reps to population means.<\/li>\n\n\n\n<li>Alex: &#8220;Peter derived the equation, the most beautiful to come out of our work.&#8221;<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>vs. Sampling:\n<ul class=\"wp-block-list\">\n<li>Limited evidence of sampling beyond V1<\/li>\n\n\n\n<li>Sampling: more samples will give more information about distribution\n<ul class=\"wp-block-list\">\n<li>So <em>distribution<\/em> will be come better defined\n<ul class=\"wp-block-list\">\n<li>But will e.g. remain bimodal, if so.<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>But in experiments: given more observations, people&#8217;s decisions become more certain.\n<ul class=\"wp-block-list\">\n<li>Suggests: People summarising distributions with means, etc.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Olfaction:\n<ul class=\"wp-block-list\">\n<li>Reminder of the Mathis paper: is it inference that all, what&#8217;s the machinery for, if a glomerular readout suffices?<\/li>\n\n\n\n<li>Why does taste go through the thalamus, but smell does not?<\/li>\n\n\n\n<li>Is the granule cell system\/solution just an older, evolutionary approach, before thalamus was available?<\/li>\n\n\n\n<li>Dimensionality of olfactory vs. visual space\n<ul class=\"wp-block-list\">\n<li>What&#8217;s the equivalent of activating single cones in the visual system?<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Learning by inference \n<ul class=\"wp-block-list\">\n<li>Like dictionary learning?\n<ul class=\"wp-block-list\">\n<li>Learn primitives to reconstruct environment<\/li>\n\n\n\n<li>Sequence the primitives for a specific task.<\/li>\n\n\n\n<li>Nested EM?\n<ul class=\"wp-block-list\">\n<li>Inner loop learns primitives<\/li>\n\n\n\n<li>Outer loop sequences primitives<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">To read<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><a href=\"https:\/\/www.nature.com\/articles\/s41593-024-01607-5\">Zero-shot learning in RNNs trained through natural language<\/a> <\/li>\n\n\n\n<li>Stepwise learning in mice doing the IBL task: <a href=\"https:\/\/www.nature.com\/articles\/s41593-025-02130-x\">Infinite hidden Markov models can dissect the complexities of learning<\/a><\/li>\n\n\n\n<li>The benefits of world and inverse models: <a href=\"https:\/\/www.sciencedirect.com\/science\/article\/abs\/pii\/036402139290036T\">Rumelhart and Jordan 1992<\/a><\/li>\n\n\n\n<li>Reading out PPCs: <a href=\"https:\/\/www.nature.com\/articles\/nn0801_826\">Deneve, Latham and Pouget<\/a><\/li>\n\n\n\n<li><a href=\"https:\/\/www.cell.com\/neuron\/comments\/S0896-6273(16)30499-8\">No need for inference in the olfactory bulb.<\/a><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"featured_media":0,"template":"","class_list":["post-8274","notes","type-notes","status-publish","hentry"],"acf":[],"_links":{"self":[{"href":"https:\/\/sinatootoonian.com\/index.php\/wp-json\/wp\/v2\/notes\/8274","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/sinatootoonian.com\/index.php\/wp-json\/wp\/v2\/notes"}],"about":[{"href":"https:\/\/sinatootoonian.com\/index.php\/wp-json\/wp\/v2\/types\/notes"}],"wp:attachment":[{"href":"https:\/\/sinatootoonian.com\/index.php\/wp-json\/wp\/v2\/media?parent=8274"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}