Lmst

@hotdogsladies Born Floridians who refuse to wear yellow unite 🤝

@siracusa With a fixed RNG seed it's 100% deterministic :)

@RecDiffs minor correction, at 1:22:02 John says GPT is deterministic, but this is not true, all LLMs sample outputs from a distribution of possible next tokens, and the degree of how variable the ouput is is controlled by a parameter called 'temperature', which is often user controllable. here's the random draw code for LLaMa 2, a similar open source model https://github.com/karpathy/llama2.c/blob/6c5d78fa41d35e57c510181309f89b189f6a536a/run.c#L686

@tpoisot maybe this whole time i should have applied BFS to tomatillo finding

@tpoisot where'd you get the tomatillos?

@tpoisot 🙋‍♂️

@atpfm The explanation of standard error with respect to John's poll is good but slightly incorrect and incomplete. Classic application of CLT https://en.wikipedia.org/wiki/Central_limit_theorem

@siracusa They name dropped "transformer" twice, but it's a six year old model. They've probably been using it for several years already

@gvdr Ah no worries, can't make it to Vienna at that time as its too close to my PhD submission deadline, but looking forward to next year!

@gvdr Hi, is this going to be hybrid or just in person in Vienna?

michael catchen boosted:

🐞 Hey all! We are running a satellite event at this year NetSci2023. 🐞

https://www.complexity.eco/

Come chat about ecological networks complexity, food webs, fancy network mathematics, and all thing interactions.

#🐞 #ecology #networks #mathematics #complexity #netsci2023

michael catchen boosted:

Reminder: I'll be mentoring a joint @ArviZ and @TuringLang #GSoC #JSoC project working on model-refitting for leave-one-out cross-validation. This project will make model comparison and assessment in Turing.jl more ergonomic and lay key groundwork for similar integration with other PPLs. If that sounds like something you'd like to work on, let us know! #JuliaLang

For details, see https://github.com/arviz-devs/arviz/wiki/GSoC-2023-projects#reloo-for-turing-models-julia. Please boost to help reach potential applicants!

michael catchen boosted:

I'm very happy to see this paper out in @MethodsEcolEvol - as ecologists are increasingly trying to predict networks, we need methodological guidelines.

The main take-home message is: ecological network data has class imbalance, and we need to account for this in training.

Long story short: train on balanced datasets, evaluate on datasets with the correct class imbalance!

https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.14071

Response of MCC, Informedness, ROC-AUC, and PR-AUC to changes in the training set balance (on the axis) for a series of increasing connectances (colour). All of these values approach 1 for a good model, but should be lower when the prediction is more difficult. Informedness is consistently high, and by contrast, MCC increases with additional training set balance. Across all models, training on a more connected network is easier. ROC-AUC is consistently high, and therefore not properly able to separate good from poor classifiers. On the other hand, PR-AUC responds to changes in the training set.

@tpoisot @jorgeapenas there's probably a way to do iterative WF in a single line

Now live in NeutralLandscapes.jl v0.1.3, you can make patchy landscapes, e.g. here via `rand(Patches(), (100, 100))` and `rand(Patches(numpatches=25), (100, 100))`. You can adjust percent habitat, the distribution of patch sizes, and more!

michael catchen boosted:

It lives - SpeciesDistributionToolkit.jl is now released, and we're working to integrate it with our boundaries finding package.

An overview of zones of ecological transitions for the Hawaii archipelago

michael catchen boosted:

Fingers crossed, but tomorrow should be the day we release the one package to rule species distribution work in #julialang - it's got GBIF integration, can download predictor data for you, uses Makie for plotting, and the very next step is direct integration with MLJ for ML applications!

michael catchen boosted:

This paper, led by @michael, is removing the most important roadblock in our ability to estimate the reliability of our prediction of ecological networks: we can estimate the false negative rate, and turn this into information about the required sampling effort. This is a qualitative step forward.

https://ecoevorxiv.org/repository/view/4929/

We conclude by estimating how false-negatives impact or interaction prediction and estimates of network structure, and by emphasizing how simulating the sampling process can aid in the design of samples to actionable monitoring of biodiversity change.

We also show there are positive associations between co-occurrence and interactions in empirical spatial/temporally sampled networks, and this actually _increases_ the rate of false-negatives. These means interactions between 'rare' species may be much more common than we know, and the nestedness we see in empirical bipartite networks may just be an artifact of observation error.

Client Info