Eddie Cunningham

Excited to share this retrospective of papers shown at ML conferences in 2021 by folks from @umasscs including several #NeurIPS2021 papers: all.cs.umass.edu/ML2021.html #UMassML21

@Suuraj @sethaxen I don't know if Hutchinson's trick can be applied to the log determinant directly, but it can be used to estimate the gradient of the log determinant pretty nicely: d/dtheta logdet(J) = d/dtheta E_z[stop_gradient(z^T J^{-1})Jz]

@sethaxen The Hutchinson's trick computes the *trace* of a linear operator without computing the matrix itself, via random projections. Maybe an analogous trick can be used for the (log) determinant? Maybe logdet(M) = E_z log(z^T M z) ?

Is anyone aware of a method for computing eigenvalues of a linear operator without first computing the matrix representation of the operator?

Interested in using rectangular matrix-vector products in your normalizing flows? Come by our poster at 12:45 PST today, in session 4 of @aistats_conf "A Change of Variables Method For Rectangular Matrix-Vector Products" #realMVP by @_ecunningham_ proceedings.mlr.press/v130/cunningha…

Tomorrow, the brilliant @ank_parikh will be presenting his work on High Precision Text Generation at @umasscs #MLFL cics.umass.edu/event/machine-… the talk will also be made available on the youtube channel: youtube.com/channel/UC9Gp6…

Join our Q&A session at the ICML Workshop on Inductive Biases, Invariances and Generalization in Reinforcement Learning (@BIGICML) at 9.15 am EST today! We will be talking about using an explicit causal dynamics model for transfer learning in RL. Paper- …s-invariancesgeneralization.github.io/pdf/big_26.pdf

Check out our implementation, written using JAX github.com/Information-Fu…

Our method learns a low dimensional representation of data in an invertible fashion to significantly improve the generative quality of normalizing flows.

Check out our work on applying normalizing flows across dimensions at the INNF+ workshop at ICML today at 1:00 pm EST today! us02web.zoom.us/j/88113801013?… arxiv.org/pdf/2006.13070…