auto-generating paper titles
Well, all the cool kids seem to be training their own text bots so here's one which finetunes gpt-2 to generate titles of scientific papers (or anything else). All code and instructions are in scrape_finetune_sample.ipynb. Works with python 3.6 and tf 1.15.0.
settings for data gathering
- all title-generating models are trained by finetuning the 117M model
- ml: Trained for 1000 steps on article titles from stat.ML betweem 2017-08-01 and 2019-07-01 with 'learning' in the abstract.
- neuro/genomics: Trained for 100 steps on 2,870 article titles from arXiv q-bio GN (genomics) and q-bio NC (neurons and cognition) between 2016-08-01 and 2019-07-01.
- quantum: Trained for 150 steps on a couple thousand article titles from arXiv quant-ph between 2019-05-01 to 2019-07-03.
- witty quotes (note: some of the produced samples may be offensive)
- tweets: Trained on the super funny tweets by @SICKOFWOLVES (uses 355M model)
- kdrama plots: Trained on kdrama synopsis first paragraphs
samples
Here are some samples for ml titles (more in the samples folder, e.g. quantum titles, kdrama synopses, witty quotes)
- A Machine Learning Framework For Computer Vision
- A Semi-Supervised Csi-On-Icb Test For Differential Privacy
- Neural Control Variates For Efficient Inference And Meaningful Decision Making
- Unlearn What You Have Learned: Recurrent Neural Networks For Pre-Processing Speech Using Sparsified Emotional Speech
- Learning Representations For Long Short-Term Memory Linked Index Circuits
- Stochastic Gradient Mcmc For Compressing Neural Network Training For Training Fine-Grained Relationships
- Bayesian Inference And Wavelet Decomposition Of Iterative Random Code
- Robust And Parallel K-Svm Through Sparsely Correlated Matrix Decomposition
- Towards A Practical
$K$ -Dimensional Deep Neural Network Model Parameterization Theory Emphasizing Sequence Interaction - Deep Signal Recovery With Dual Momentum: A Complexity And Stability Analysis Based Analysis
- Fault Diagnosis Using Deep Signal Recovery
- Sparse Least Squares Regression: Robust Regularization And Recommendations For Improved Statistical Validation
- Stacking With A Neural Network For Sequential Intent Classification
- Deep Residual Auto-Encoders For Multi-Label Image Classification
- Sparse Least Squares Regression For Robust And Adaptive Classification
- Anatomical Coronary Stearage Reconstruction Using Deep 3D Convolutional Neural Networks
- Deep Neural Processes
- Gated-Cgroup Communications Solution For Ciliary Dysfunction Detection Using Hierarchical Directed Acyclic Graph Convolutional Network
- Understanding Batch Normalization
- Improving Gan Performance With Stochastic Gradients And More Via Optimized Alternative Estimation
- Graphoisheter Encephalization Encephalization
- Improving Gans Using Covariate Shift And Multivariate Spatial Embeddings
- Dynamic Event Graph Convolutional Networks For Adverse Event Forecasting
- Fast Asynchronous Parallel Training Of Deep Networks
- On The Non-Parametric Power Of Logistic Regression For Smooth Events
- Unifying Pac And Learning Mdps Using Influence Functions
- Machine Learning To Plan And Downlink Using Intrinsic Motivation
- Classifier Readiness Testing For Imbalanced Data
- Fast And Scalable Bayesian Deep Learning With Limited Observations
- A Comparison Of Deep Neural Networks And Adaptive Graph Neural Networks For Anomaly Detection
- Distributed Deep Learning With Gossip Networks Using Bidirectional Lstm Sensors
- Revisiting Reuse Of Super Categories
- Anatomical Visual Exploration
- Multimodal Social Learning With Active Interest Discovery
- Stochastic Variance-Reduced Cubic Regularization For Approximate Inference
- Predicting County Level Corn Yields Based On Time Series Data
- A Deep Residual Network Approach For Predicting County Level Eegs Using Sparse And Incomplete Data
other things to train on
- limericks
- haikus
- song lyrics
reference
- troubleshooting: if you get an error, it might help to run
export PYTHONIOENCODING=ASCII
before finetuning - uses gpt-2 simple
- uses arxiv-scraper
- uses tweetscraper
- website based on this example
- adorable robot from here