- \n
- Agents are smaller than the world<\/li>\n\n\n\n
- Therefore everything they learn is an approximation.<\/li>\n\n\n\n
- Therefore the world appears non-stationary<\/li>\n\n\n\n
- Agents must be able to learn and act at run-time<\/li>\n\n\n\n
- Building in domain-dependent knowledge is ultimately harmful (the bitter lesson)<\/li>\n\n\n\n
- Agents should be capable of developing open-ended abstractions.<\/li>\n\n\n\n
- General agreement across fields on design components of intelligent agents:\n
- \n
- Perception<\/li>\n\n\n\n
- Policy<\/li>\n\n\n\n
- Value Function<\/li>\n\n\n\n
- Transition Model<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- The one-step trap of RL: planning one step ahead\n
- \n
- Errors compound<\/li>\n\n\n\n
- Exploration blows up in time.<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- Solution: higher-order transition models to plan over larger timescales.<\/li>\n<\/ul>\n\n\n\n
The OaK architecture<\/h3>\n\n\n\n
- \n
- Learning higher-order transitions by constructing reward-respecting subgoals<\/strong>.<\/li>\n<\/ul>\n\n\n\n
- \n
- Options and Knowledge\n
- \n
- Option: A tuple of policy (state $\\to$ action) and termination function (state $\\to [0,1]$ termination)<\/li>\n\n\n\n
- <\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
Concepts and References<\/h3>\n\n\n\n
- \n
- Continual learning\n
- \n
- “Catastrophic loss of plasticity”<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- “Reward-respecting sub<\/li>\n<\/ul>\n\n\n\n
Why diffusion models don’t memorize<\/h2>\n\n\n\n
- \n
- Time to generation $\\tau_\\text{gen}$ is constant.<\/li>\n\n\n\n
- Time to memorization increases with number of examples $n$.<\/li>\n\n\n\n
- Gap between is good generalization time, early stopping.<\/li>\n<\/ul>\n\n\n\n
Poster Session 1<\/h2>\n\n\n\n
Generalizable, real-time neural decoding with hybrid state-space models<\/h3>\n\n\n\n
- \n
- Model has three parts:\n
- \n
- Mapping spikes across units and sessions into a common vector state.\n
- \n
- Split time into chunks.<\/li>\n\n\n\n
- Learns unit-specific embedding of units.<\/li>\n\n\n\n
- Session-specific keys and values queried with a fixed vector.<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- Vector state is updated using a state-state model to maintain memory.<\/li>\n\n\n\n
- Reading out the predicted behaviour from the history.\n
- \n
- Queried with the session and time.<\/li>\n\n\n\n
- Use previous history as keys and values.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- What is it doing?\n
- \n
- Finding a state-space representation that can be linearly decoded to extract behaviour.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
Finding separatrices of dynamical flows with Deep Koopman Eigenfunctions<\/a><\/h3>\n\n\n\n
- \n
- Decision making as moving between different basins<\/li>\n\n\n\n
- Typical analysis is at the fixed points.<\/li>\n\n\n\n
- Found as locations where $\\|f(x)\\|_2^2 = 0$.\n
- \n
- i.e. zero of a scalar function of the dynamics.<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- What such function to find sepratrices?<\/li>\n\n\n\n
- Key idea: map dynamics to unstable scalar dynamics $\\psi(x)$<\/li>\n<\/ul>\n\n\n\n
![\"\"](\"https:\/\/sinatootoonian.com\/wp-content\/uploads\/2025\/12\/image-1024x702.png\")
<\/figure>\n\n\n\n- \n
- Zero of the function corresponds to separatrix.<\/li>\n\n\n\n
- On either side the function increases to +\/- $\\infty$: unstable dynamics.<\/li>\n\n\n\n
- Use a neural net to find such a function.<\/li>\n\n\n\n
- We want $$\\dot \\psi(x(t)) = \\nabla_x \\psi^T \\dot x = \\nabla_x \\psi^T f(x) = \\lambda \\psi,$$ so minimize $$\\EE_{x \\sim p(x)} \\|\\nabla \\psi ^T f(x) – \\lambda \\psi \\|_2^2.$$ <\/li>\n\n\n\n
- Do this with a neural net using gradient descent?<\/li>\n\n\n\n
- Once you have it, you can design optimal (smallest) perturbations that will move state to the separatrix, evoke new behaviour.<\/li>\n<\/ul>\n\n\n\n
Learning to Cluster Neuronal Function<\/a><\/h3>\n\n\n\n
- \n
- Deep learning models can learn per-neuron embeddings that predict responses.<\/li>\n\n\n\n
- These embeddings may reflect neuronal types, but don’t cluster.<\/li>\n\n\n\n
- If true neuron types cluster, then incorporating this information should improve embeddings.<\/li>\n\n\n\n
- Key idea: learn an initial embedding to predict responses, then augment the loss with a clustering promoting term:<\/li>\n<\/ul>\n\n\n\n
![\"\"](\"https:\/\/sinatootoonian.com\/wp-content\/uploads\/2025\/12\/image-1.png\")
<\/figure>\n\n\n\n- \n
- Here Q is the data distribution, modelled as a mixture of t-distributions\n
- \n
- Number of distributions is a parameter.<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- P is a target distribution based on previous work.<\/li>\n\n\n\n
- Fit using various numbers of clusters, measure quality of clustering using adjusted rand index.\n
- \n
- Optimize $\\beta$ using cross-validation?<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- Setting # clusters to whateve maximizes ARI finds the right number of clusters in marmoset.<\/li>\n\n\n\n
- No clear peak in ARI in V1, suggests no clear types but rather a continuum.<\/li>\n<\/ul>\n\n\n\n
Johnson-Lindenstrauss Lemma Beyond Euclidean Geometry<\/a><\/p>\n\n\n\n
- \n
- JL lemma: given a set of points in $n$-dimensional Euclidean space, there exists a mapping $f$ to $m$-dimensional Euclidean space that approximately preserves distances, so $(1 – \\vareps) \\|x_i – x_j\\| \\le \\|f(x_i) – f(x_j)\\| \\le (1 + \\vareps) \\|x_i – x_j\\|$, where $m$ is $O(\\vareps^{-2} \\log n).$<\/li>\n\n\n\n
- What if only a distance matrix is available, not point coordinates, and what if distance are not necessarily metric (don’t satisfy triangle inequality)?<\/li>\n\n\n\n
- Non-metric distances can be expressed as a sum of positive and negative Euclidean part.\n
- \n
- Result: Prove that JL holds for this positive-negative split.<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- Non-metric distances can also be expressed as a power distance.\n
- \n
- Result: Prove that JL holds for the power distance.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
Spectral Analysis of Representational Similarity with Limited Neurons<\/a><\/p>\n\n\n\n
Brain-like Variational Inference<\/a>**<\/p>\n\n\n\n
- \n
- Introduce FOND: Free Energy Optimization using Natural Gradient Dynamics.<\/li>\n\n\n\n
- Their idea is to derive neurally plausible algorithms by iterative minimization of the negative variational free energy.\n
- \n
- This is an old idea.<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- They do a concrete demonstration by using a poisson model with log rates<\/li>\n\n\n\n
- Show that it outperforms standard methods on sparsity, neural plausibility, etc.<\/li>\n\n\n\n
- Show that it outperforms amortized inference methods.<\/li>\n<\/ul>\n\n\n\n
Jacobian-Based Interpretation of Nonlinear Neural Encoding Model<\/a><\/p>\n\n\n\n
- \n
- Metric for measuring nonlinearity by quantifying changes in input-output Jacobian of NN model with input.\n
- \n
- Per voxel, compute input output jacobian per sample.<\/li>\n\n\n\n
- Compute the mean across samples<\/li>\n\n\n\n
- Compute deviation relative to the mean<\/li>\n\n\n\n
- Summarize deviations across coordinates using L1 norm<\/li>\n\n\n\n
- Measure the variance of the mean absolute deviations across samples.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
Brain-Like Processing Pathways Form in Models With Heterogeneous Experts<\/a><\/p>\n\n\n\n
Poster Session 2<\/h2>\n\n\n\n
From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit<\/a><\/h3>\n\n\n\n
- \n
- Sparse auto-encoders are useful for capturing concepts.<\/li>\n\n\n\n
- Assume that images are linear combinations of concepts.<\/li>\n\n\n\n
- Natural scences have hierarchical structure.<\/li>\n\n\n\n
- Deep nets trained on natural scenes uncover hierarchical structure.\n
- \n
- Concepts at different layers tend to be orthogonal.<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- Vanilla sparse auto-encoders have trouble capturing this hierarchical structure.<\/li>\n\n\n\n
- Solution: replace inference stage of auto-encoder with matching pursuit:\n
- \n
- Iteratively find the concept most correlated with the current residual and subtract it out.<\/li>\n\n\n\n
- Has the property that the result residual is orthogonal to the selected concept.<\/li>\n\n\n\n
- Naturally captures the cross-hierarchy-layer orthogonality.\n
- \n
- At least for adjacent layers.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n
- Tested on surrogate data.\n
- \n
- Data has hierarchical structure if $p(parent|child) = 1$ but $p(child|parent) < 1$.<\/li>\n\n\n\n
- MP-SAE recovers the underlying dictionary element relationships<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
![\"\"](\"https:\/\/sinatootoonian.com\/wp-content\/uploads\/2025\/12\/image-2-1024x366.png\")
<\/figure>\n\n\n\n
<\/p>\n\n\n\nRepresentational Difference Explanations<\/a><\/h3>\n\n\n\n
- \n
- Want to find differences in representation<\/li>\n\n\n\n
- Outcome: clustering of samples that are considered similar in one rep but not the other.<\/li>\n\n\n\n
- Approach:\n
- \n
- Compute $n \\times n$ distance matrices $D_A$ and $D_B$ for the two reps.<\/li>\n\n\n\n
- Normalize distance matrices $d_A$ and $d_B$, using nearest neighbours.<\/li>\n\n\n\n
- Compute a relative difference of normalized distances: $$G_{A,B}^{ij} = {d_A^{ij} – d_B^{ij} \\over \\min(d_A^{ij}, d_B^{ij})}.$$<\/li>\n\n\n\n
- Compress these to $\\{-1,1\\}$ using $\\tanh$ and form an affinity matrix by passing through negative exponential: $$F_{A,B}^{ij} = \\exp(-\\beta\\tanh(\\gamma G_{A,B}^{ij}).$$<\/li>\n\n\n\n
- This will produce strong links between pairs that are more similar in $A$ than in $B$.<\/li>\n\n\n\n
- Cluster the resulting graph, forming explanations:\n
- \n
- Different configurations of samples that are more similar in $A$ than in $B$.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
![\"\"](\"https:\/\/sinatootoonian.com\/wp-content\/uploads\/2025\/12\/image-3-1024x373.png\")
<\/figure>\n\n\n\nConnecting Jensen\u2013Shannon and Kullback\u2013Leibler Divergences<\/a><\/h3>\n\n\n\n
- Different configurations of samples that are more similar in $A$ than in $B$.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
- Metric for measuring nonlinearity by quantifying changes in input-output Jacobian of NN model with input.\n
- Result: Prove that JL holds for the power distance.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
- Here Q is the data distribution, modelled as a mixture of t-distributions\n
- Finding a state-space representation that can be linearly decoded to extract behaviour.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
- Mapping spikes across units and sessions into a common vector state.\n
- Model has three parts:\n
- Continual learning\n
- Options and Knowledge\n
- Learning higher-order transitions by constructing reward-respecting subgoals<\/strong>.<\/li>\n<\/ul>\n\n\n\n