-
Notifications
You must be signed in to change notification settings - Fork 9
Schedule
Volunteer by editing below. Small contributions are highly encouraged, and don't take the topic of the week too seriously.
Remi M: Can CCLF predict in vitro tumor growth with ML?
Sam F: High-Fidelity Image Generation With Fewer Labels https://arxiv.org/pdf/1903.02271v1.pdf
Sam F: Venture beyond Empirical Risk Minimization with Mixup a simple and effective data augmentation strategy : https://arxiv.org/abs/1710.09412
Sam F: pix2pix Image-to-Image Translation with Conditional Adversarial Networks. https://arxiv.org/pdf/1611.07004.pdf
Sam F: snorkeling for training data: Snorkel: Rapid Training Data Creation https://arxiv.org/abs/1711.10160
Sam F: Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV) https://arxiv.org/pdf/1711.11279.pdf
Sam F: "Deep-learning cardiac motion analysis for human survival prediction" a de-noising autoencoder which predicts survival from 4D cardiac MRIs https://www.nature.com/articles/s42256-019-0019-2
Sam F: Attention is all you need Tracing the development in recurrent language models like "Neural Machine Translation by Jointly Learning to Align and Translate" https://arxiv.org/pdf/1409.0473.pdf It's standalone usage in "Attention Is All You Need" https://arxiv.org/pdf/1706.03762.pdf And recent applications augmenting convolutional nets: "Attention Augmented Convolutional Networks" https://arxiv.org/pdf/1904.09925.pdf
Joshua Batson: Noise2Self: Blind Denoising by Self-Supervision https://arxiv.org/abs/1901.11365
Alessandro Achille: Critical Learning Periods in Deep Neural Networks. https://arxiv.org/pdf/1711.08856.pdf
Sam F: Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks https://arxiv.org/pdf/1707.01836.pdf The paper frames processing ECGs as a sequence to sequence learning problem. We will also discuss and demo several other ways to structure this problem including multi-task classification, waveform regression, and captioning.
Jon Bloom: Morse ensemble learning (Morsembling?): how the (trivial) topology of Euclidean space forces geometric relationships between the critical points of any "smooth" loss function on a deep neural network. Together with known probabilistic results, this gives a theoretical foundation for existing ensemble learning methods that in turn suggests new recursive algorithms that trade off between the quality and quantity of minima. I'm excited to have a conversation about why ensemble (and consensus) methods may be particularly useful when applying ML to biology. The Loss Surfaces of Multilayer Networks (2014) Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs (NeurIPS 2018) Essentially No Barriers in Neural Network Energy Landscape (ICML 2018) Stochastic Gradient Descent Escapes Saddle Points Efficiently (2019)
Luca: Classifying and segmenting microscopy images with deep multiple instance learning (https://www.ncbi.nlm.nih.gov/pubmed/27307644)
Sam F: Bias in word embeddings (https://www.pnas.org/content/115/16/E3635) and one approach to combat it: Multiaccuracy: Black-Box Post-Processing for Fairness in Classification (https://arxiv.org/abs/1805.12317)
Sam F: Deep Sets (https://arxiv.org/pdf/1703.06114.pdf)
Mehrtash B: Neural Ordinary Differential Equations (https://arxiv.org/pdf/1806.07366.pdf)
Sam F: Re-usable Holdout (http://science.sciencemag.org/content/sci/349/6248/636.full.pdf)
Sam F: Deformable Convolutions (https://arxiv.org/pdf/1703.06211v2.pdf)
Sam F: Multi-view pooling (https://arxiv.org/pdf/1505.00880.pdf)
Sam F: BQSR Models
Daniel Kunin: Autoencoders and SVD
Marton Kanasz-Nagy: Information bottleneck Part 2: https://openreview.net/forum?id=ry_WPG-A-
Ben Kaufman: Information bottleneck Part 1: https://arxiv.org/abs/1503.02406
Sam F: GPU on google cloud
Sam F: Multi-Modal Learning
Takuto: Autoencoding Variational Bayes.
David B: Neural Autoregressive Distribution Estimator
Sam F: DNA Motifs, Deep Bind, etc
David B: Undirected Models, Contrastive Divergence, restricted Boltzmann machines
Sam F: Autoencoders in Keras. Preludes to Autoencoding Variational Bayes.
David B: Self-supervised learning, a clever extension of the autoencoder concept:
- Agrawal et al, "Learning to See by Moving"
- Wang and Gupta, "Unsupervised Learning of Visual Representations using Videos"
- Noroozi and Favaro, "Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles"
Sam F: Generalizations of translation, bi-directional RNNs, Heuristics for better GANs and CNN + RNN caption generation.
David B: overview of a reasonable chunk of Chapter 13
Sam F: Gene2Vec
David B: Joint learning of word embeddings in two languages
David B: Word embeddings: Bengio 2003, word2vec, and GloVe, plus the Devlin et al paper on translation.
Sam F: visualizing layers.
We will brainstorm some problems relating to Mutect 2.
We will brainstorm Mimoun's project for identifying cancer cells via imaging at the Cancer Cell Line Factory.
Rudimentary plans for applying deep learning to BQSR.
Rudimentary plans for Steve's application of deep learning to SV breakpoints.
Ray Jones on this paper, with roots in this one.
David: Either a paper on using RNNs to predict protein or miRNA binding affinity, maybe both papers.
Umut: Fiddle and the cyclic loss estimator
Sam F: GANs in Keras, software architecture for DL
David: Brainstorming deep learning in Mutect.
Sam F: RNNs and their vanishing/ exploding gradients. Video from the very funny Jurgen Schmidhuber. LSTMs and GRUs in keras
David: Basic RNN architectures and contrasting RNNs with HMMs
Joe: Schreiber et al "Nucleotide sequence and DNaseI sensitivity are predictive of 3D chromatin architecture" (bioRxiv)
Sam F: More CNNs in Keras, Deeper look at inception, vgg16, and resnet
David: Saxe et al "Random Weights and Unsupervised Feature Learning" -- the surprising observation that CNNs with random kernels don't perform much worse than CNNs with trained kernels.
Sam F: Old school convolution, MNIST CNNs
David: Romero et al, FitNets
Sam F: GPU Tasting Menu Sam F: Intro to Keras Functional API
Larson: Discussed two recent studies related to stochastic gradient descent 1) Li, 2014 on efficient minibatch stochastic optimization and 2) Kimgma and Ba, 2015 on the 'Adam' algorithm. Jon Bloom discussed geometric approaches to identifying and navigating critical points in high dimensional space.
David: examples of data augmentation for speech recognition from: 1) Schluter and Grill, "Exploring Data Augmentation for Improved Singing Voice Detection"; 2) Ko et al, "Audio Augmentation for Speech Recognition"; Salomon and Bello, "Deep Convolutional Neural Networks and DataAugmentation for Environmental Sound Classification."
Umut: Generative adversarial networks
David: quick overview of regularization and an extremely half-baked thought on why early stopping works.
Sam F: Keras notebook on chromatin state prediction with a 1D convolutional net
Mehrtash: universal approximation theorems
Sam F: getting started with Keras