ICLR & Epi Submissions

# ICLR & Epi Submissions ## Top level todo: **Martin** - [ ] [epi] do a sensitivity analysis of the simulator wrt mobility / testing capacity - [ ] book a meeting with nasim about math iclr paper - [x] beautify the plots **Tegan** - [ ] Send an email out with a todo list for each person by Monday **Prateek** - [ ] get results... **Nasim** - [ ] get results... Akshay Eilif irina Yoshua ## ICLR Feedback - [ ] Specifically address ML Novelty (enumerate core elements here) - [ ] Add an overview figure (figure 1 should capture the main ideas) - [ ] Defend scalability of proposed method and that it can be deployed to real world (inc. runtimes) - [ ] Fit to real data or explain why simulation. Make _extremely_ clear what is claimed and not claimed about covid. - [ ] Make _extremely_ clear what is claimed about our people model (and what issues remain there). Contextualize these results wrt society. - [ ] Underline the domain randomization and how we ensure that our system relates to the real-world. - [ ] Address how rudimentary covid research is (this is not a reason not to publish, but to have ranges on key parameters and to update them), addressing how robust our method is to these limitations - [ ] Address intrinsic and extrinsic propoerties and how we think about adding data from multiple sources. - [ ] Exact loss terms, why MSE - [ ] Be clearer about baselines - [ ] Discuss "spillover and confounding". Directly address nonstationarity. - [ ] Cite additional related works: - [ ] Fan et al "A unifying variational inference framework for Hierarchical Graph-Coupled HMM with an application to influenza infection" - [ ] Makar et al "Learning the probabiltiy of activation in the presence of latent spreaders" - [ ] Address novelty of contributions - [ ] Cite simulator and sensitivity analysis (need to quantify validation to real-world data) - [ ] Defend the scale of the simulation and why this would generalize to larger simulations - [ ] Defend against "castle built on sand" problem; name key assumptions - [ ] make current state of the art clear - [ ] underscore that "differential benefit of one method over the other" is the key aspect, not necessarily transfer to realworld, or that the simulator is a perfect replication of real-world conditions. Specific Promises: - [ ] increased domain randomization - [ ] vary testing rates - [ ] vary test delay - [ ] vary population size - [ ] we will create an overview figure - [ ] mention runtimes - [ ] add description of demographic statistics (maybe not this...) ## Epi paper - [x] Do a first pass and just take notes - [x] add in prelim results for validation to real data - [ ] Get chi squared test to work - [ ] Start sim on march 20 (translate to the right), plot with only hospitalizations and mortalities), and try to get another 20 days of delay on the peak of the disease. Smooth them and make sure y-axes are the same. maybe put them on the same plot. - [ ] Run another validation experiment - [ ] ask epis for CNN-test of the epi paper - [ ] need to differntiate from pre-existing simulators: privacy model, ml? ### Major Notes - Introduction - Do we have a citation for the Rt of 2 during the last week of march? QC imposed lockdown on March 20. - Maybe change non-pharmaceutical interventions have been introduced or implemented, not adopted. - this sentence is borked: However, sustainedlimitation of human activity is likely not viable with regards to individuals’mental health (ref) and financial security (ref), children’s development (ref), aswell as national and global economies (ref). - promptly self-isolate? or immediately self-isolate. "in turn?" - find the range of days used in manual contact tracing notification - Maybe note here how low the rate of call answering has been - Has there been research on the impact of manual contact tracing wrt sars 2002? - cite asymptomatic case percentage (30 percent) - remove in-text bolding *never*, *all* remove the whole *never* sentence IMO - "App-based CT approaches" maybe, "the first step towards creating a more efficient intervention is to demonstrate it in-silico, and only then to deploy it and observe how traced individuals react to different warning signals" - Related work / state of the art -- this deserves its own paragraph - Related Work - First paragraph: We need a ton more citations - Second Paragraph: Is this true? there are not many agent-based models? Then we should talk about other ABMs, it's not an acceptable cope out to me. Need to flesh this out - Third paragraph on mathematical spread modeling: need to write it - Methods - Already onto page 5... - We actually don't model the app-based recommendations this way, we just reduce mobility directly (last sentence prateek) ### Minor notes - need to fix all the broken latex stuff - Need to fix the citations ### Tegan ## Epi paper - Point of our paper is to provide a testbed for different digital contact tracing methods, validate that it doesn't prefer one method over another - How our thing is different from other things - other diseases or privacy methods, review, but say why we felt the need to do this - Related work needs to be bulletproof ## ML paper - Related work - choose subsections - make clear how we're different from relworks and why we did what we did - Overview figure - Section on PRA that makes clear what PRA is - Organization of this section - distributed inference & privacy - Domaian rand & sim2real - retraining & nonstationarity - Different name??? -