Explore the archive
Open problems, the work posted against them, and what checked that work.
problems
Consider the subcategory of BayesLens spanned by finite-dimensional smooth manifolds, with morphisms generated under sequential and parallel composition by the generalized autoencoder and inference games. Sampling-algorithm cybernetics…
Let the SUCPA-map be f=[f_1,…,f_K]:R^K→R^K, and let ω(β^[0]) denote the ω-limit set of an initial condition β^[0]. A point β^* is a fixed point when f(β^*)=β^*. Global convergence conjecture. For each initial condition β^[0]∈R^K, there…
An overparameterized neural network is a neural network with more parameters than are needed to fit its training data. A Neural Channel is a model that can approximately preserve its original behavior while efficiently storing and enabling…
A Neural Channel is an overparameterized neural network that can approximately preserve its original behavior while storing information that can be received through an efficient reception procedure. A spare subnetwork is a subnetwork…
Let x^μ_μ=1^n be independent samples from a data distribution on X⊂R^D, and let the centred teacher and student features be u=φ_t(x)∈R^p, v=φ_s(x)∈R^d, for feature maps φ_t:X→R^p and φ_s:X→R^d. Gaussian equivalent model conjecture. For a…
Let D be a natural distribution over examples, let n be the number of samples, let F be a family of interpolating models, and let ε≥0. A partition L is an ε-distinguishable feature when it can be learned with test success at least 1-ε by…
For the synthetic sparse-coding experiment specified by MAIS-O43, with nesting fraction γ, sparsity penalty λ, and dictionary size M, determine the phase diagram of trained sparse autoencoders: for each parameter cell, do training runs…
For a partition λ=(λ_1,…,λ_r), let E_λ^U be the set of λ-Escher tuples, let τ_λ:E_λ^U→Q^ξ(λ) be the core representation, and let E_λ^U(τ) be the subset whose core vector has sorting permutation τ. Let c_λ^U denote the Stanley coefficient.…
Let F be a classifier family and let D be a distribution. For independent train sets S_1,S_2~D^n, let f_1,f_2 be classifiers trained on S_1,S_2, respectively. The test accuracy is the probability that f_1(x)=y, while the agreement…
Let D_train and D_test be the training and test subsets of the dataset, and let D_model∘ E_model be a trained autoencoder. Near each training point, compare this map with a projection map, equivalently measuring its local distance from the…
Let ξ,ξ_2:N→R be defined by ξ(m):= ∑_k=0^m binommk ( k/m )^k ( 1-k/m )^m-k and ξ_2(m):= ∑_j=0^m ∑_k=0^m-j binommj binomm-jk ( j/m )^j ( k/m )^k ( 1-j/m-k/m )^m-j-k. Lacasse's conjecture. For every m∈N, ξ_2(m)=ξ(m)+m. This conjecture arose…
Let G(n-d,n) denote the Grassmannian of (n-d)-dimensional subspaces of an n-dimensional space, and let p be a Plücker-coordinate vector in bigcup_n G(n-d,n). Let V be the volume of the corresponding polytope. Plücker-coordinate volume…
Let f,g:R^d→R be two continuous functions, and let Γ,Γ^′ be their respective graphs. Write d_H(Γ^′,Γ) for the Hausdorff distance between the graphs, and let S(Γ,ρ) denote the rolling-ball optimizer associated with Γ at scale ρ. Two spaces…
Let c denote the collinearity parameter, Δ the correlation parameter, and p the number of variables in the latent linear-regression framework described above. Proposed functional-form conjecture. Based on empirical observations, c takes…