Part of my series of notes from ICLR 2019 in New Orleans.
Learning Protein Structure with a Differentiable Simulator
- goal: learn protein sequence-to-structure end-to-end
- weak interactions (water-liking & -avoiding) => global protein fold
- energy-based models make sense to use here, but inference intractable
- jointly learn energy function and inference
- deep energy function to capture multi-scale dependence
- “in more ICLR language: a deep conditional random field with deep features”
- efficient sampling with coordinate transformation
- benefits of this approach:
- probabilistic – both discrete and continuous uncertainty
- some generalization over structure
- more so than just predicting geometry (angles)
- two months to train though
Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset
- music is complex, partly because of wide range of timescales
- (this is another theme here at ICLR for me)
- use structured prior – i.e., composers make music with notes
- importance of real performance data vs. just midi files / quantized scores
- not much of this data available
- but if we had a good transcription model, we could generate way more data from “the wild”
- => Wave2Midi2Wave
-
Onsets & Frames model for transcription
- result: MAESTRO dataset
- Music Transformer to compose music – transformer model with relative attention
- piano roll-conditioned WaveNet
- also reproduces other sounds in a piano recital besides just the notes
- much more here
A Unified Theory of Early Visual Representations from Retina to Cortex through Anatomically Constrained Deep CNNs
- receptive fields of neurons
- retina – center surround
- cortex – edge detector
- why? (and why different?)
- CNNs as a model of visual systems
- learn similar receptive fields, but seem to skip center-surround and go straight to edges
- try to build a CNN that explicitly models biological visual system
- optic nerve as a physical bottleneck between two distinct anatomical entities
- if we build this bottleneck into the network, we get center-surround!!!
- diversity across species
- fairly linear retinal cells that don’t seem to be semantic feature extractors (monkeys)
- vs. many types of nonlinear cells that e.g. detect predators (mice)
- the difference here is in complexity / depth of the system after retinal stage
- retina is more linear with the brain is deep, extracts more useful features when brain is shallow
- this only really happens when we have the retinal bottleneck
- this is actually really fucking cool
