for the KTH TCS Seminar series
on March 26, 2019
by Hugo Mougard
Harder than Computer Vision, NLP
3 sub-problems
2 projects
Some gory details
gitbase
bblfsh
Types shared across languages
apollo
tmsc
&
snippet-ranger
id2vec
ml
3 targets initial targets
Goal: automate formatting
Must explain false positives
OR
Must not have false positives
Unsupervised learning with explainable rules
Reproduction task: ~94.3% precision
Ongoing, Early stage. Goal: use expressive models
Predict formatting characters before each leaf of the AST
Supervised learning?
2 steps
Each repository is considered as a task $\mathcal{T}_i$
$$\min_\theta\sum_{\mathcal{T}_i \sim p(\mathcal{T})} \mathcal{L}_{\mathcal{T}_i}(f_{\theta})$$Model-Agnostic Meta-Learning approach
Minimize loss AFTER having optimized for a task $\mathcal{T}_i$
$$\min_\theta\sum_{\mathcal{T}_i \sim p(\mathcal{T})} \mathcal{L}_{\mathcal{T}_i}(f_{\theta'_i})$$$\theta'_i$ is optimized from $\theta$ on $\mathcal{T}_i$ at each iteration
For the next seminar!
Thank you for your attention!
Questions & Discussion
Hugo Mougard <hugo@sourced.tech>