<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Stacey Svetlichnaya on Medium]]></title>
        <description><![CDATA[Stories by Stacey Svetlichnaya on Medium]]></description>
        <link>https://medium.com/@staceysvetlichnaya?source=rss-bf69d40ad788------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ixbrEs-bNFOMc84L5p5VBQ.jpeg</url>
            <title>Stories by Stacey Svetlichnaya on Medium</title>
            <link>https://medium.com/@staceysvetlichnaya?source=rss-bf69d40ad788------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 04:37:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@staceysvetlichnaya/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Hyperparameter Search with Iterative Sweeps]]></title>
            <link>https://medium.com/data-science/hyperparameter-search-with-iterative-sweeps-3799df1a4d45?source=rss-bf69d40ad788------2</link>
            <guid isPermaLink="false">https://medium.com/p/3799df1a4d45</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning-tools]]></category>
            <category><![CDATA[hyperparameter-tuning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[visualization]]></category>
            <dc:creator><![CDATA[Stacey Svetlichnaya]]></dc:creator>
            <pubDate>Mon, 22 Jun 2020 17:47:51 GMT</pubDate>
            <atom:updated>2020-07-09T18:34:47.104Z</atom:updated>
            <content:encoded><![CDATA[<p>How to run an effective hyperparameter search on deep learning models, with specific visual examples from Weights &amp; Biases</p><p>I’ve spent several years reproducing and optimizing <a href="https://app.wandb.ai/stacey">various deep learning models</a>, primarily for computer vision and NLP, often with extremely short deadlines. My distilled high-level strategy for hyperparameter search is bounded exploration (try a wider range of values for fewer variables) and faster iteration (more short phases of exploration building on each other). I hope this overview of hyperparameter search helps you tune deep learning models a bit faster regardless of the framework or tools you use.</p><h3>What is hyperparameter search?</h3><p>Hyperparameter search — or tuning, or optimization — is the task of finding the best hyperparameters for a learning algorithm. Such tuning could be done entirely by hand: run a controlled experiment (keep all hyperparameters constant except one), analyze the effect of the single value change, decide based on that which hyperparameter to change next, run the next experiment, and repeat. Of course, this would be painfully slow for deep learning algorithms like convolutional or recurrent neural networks: each experiment (one combination of hyperparameter values) requires training the model end-to-end. We can make this process more manageable by going meta on the manual search.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0uheZBSEA9YVBu-a" /><figcaption>An example Sweeps workspace in the Weights &amp; Biases UI. Image by author.</figcaption></figure><h3>Methods: Manual, grid, random</h3><p>The core promise of algorithmic or automated hyperparameter tuning is to be smarter about choosing the next experiment to run. If we need to try a bunch of different values for a bunch of different hyperparameters anyway, how can we optimally pick the next combination — and run fewer experiments, waste fewer computing resources, and get results even faster? The standard methods for this are <strong>grid</strong>, <strong>random</strong>, and Bayesian or <strong>bayes</strong> (though you’ll hear a long tail of other terms and strategies in this space, especially for model-based optimization). Grid search tries all possible combinations — think of this as a nested for-loop, where the total number of experiments will be on the order of number of hyperparameters times the number of possible values for each hyperparameter. This is useful if your search space is small and you want to guarantee finding the best option. Random search is generally faster and especially recommended if you have a larger search space, as it will sample a wider range of combinations and give you a sense of the possibilities faster than grid search. However, it is not guaranteed to find the best combination.</p><h3>Bayes</h3><p>Bayesian optimization builds a probabilistic model, or educated guess at, the function mapping from your hyperparameter choices to a target measure of your performance, usually validation accuracy. This Bayesian model of the actual model you’re training predicts what values to try next, based on all the values and results seen so far. The process aims to balance the explore-exploit tradeoff, trying both values that it’s not sure about to get more information about the space and ones that it expects to be near optimal, according to the preset objective. In practice, Bayesian optimization works faster than grid and random because it runs fewer “bad” experiments. Bayes does require a target metric to optimize, which can default to validation loss for your model. You can read <a href="https://towardsdatascience.com/a-conceptual-explanation-of-bayesian-model-based-hyperparameter-optimization-for-machine-learning-b8172278050f">a more detailed walkthrough of Bayesian optimization here</a>.</p><h3>The long tail of other methods</h3><p>Other methods include gradient-based optimization to explicitly compute the gradient with respect to certain hyperparameters, early stopping to avoid wasting time on unpromising runs, and evolutionary optimization applied to consecutive sets of chosen hyperparameters. There’s also population-based training, where multiple independent processes learn both hyperparameter values and network weights in tandem. There are many options — both <a href="https://medium.com/criteo-labs/hyper-parameter-optimization-algorithms-2fe447525903">algorithmic details</a> and black-box optimization services — than I can cover here. Once you’re looking into these details and tradeoffs, I would ask myself, am I at the stage where I’ve thought through the specific model I’m tuning and the specific problem I’m solving, and I’ve concluded there’s nothing else clever I can do — I just want to throw GPUs at this task? If yes, I’d suggest using some optimization, like Bayesian, because that will most likely work faster, and go think about something else. But in my experience, the more interesting, challenging, and rewarding work of hyperparameter tuning happens <em>before</em> that stage, and we don’t really have good algorithms for it yet — and that’s why many of us still have jobs in machinelearning. So, let’s dive into that part!</p><h3>The hyperparameter-to-code spectrum</h3><p>Before you start trying different hyperparameters, you decide what constitutes a hyperparameter. There’s an important distinction between the regular parameters of the network, like the model weights and biases which are themselves learned, or iteratively improved during the training process, and the hyperparameters, which are configured o<em>utside</em> the training process and generally fixed for its duration (though some could be adaptive, like learning rate). For a given network, these can include the number of training epochs, learning optimizer type, batch size, weight decay, and much more. The network architecture itself could be a hyperparameter: am I starting from a well-known base network like Inception or ResNet? How many layers am I freezing versus fine-tuning? If I’m designing a network from scratch, then the number, size, type, and connection patterns of the layers, as well as the dimensions of the sliding window in convolutions, can all be hyperparameters. To decide whether to keep a given hyperparameter fixed or variable, you might think about how well you can predict its effect in advance, and how it will affect the overall run time and complexity of your search.</p><h3>Epochs</h3><p>Epochs — or number of passes through your entire training data — are a very well-understood hyperparameter. When I’m running repeated experiments, I want to set this high enough to see definite improvement but not so high that I’m wasting time on a very slowly converging curve. I usually try 10 or so, if I can get away with it — as long as I can see the loss decrease and start to asymptote. Seeing improvement in validation loss is especially important to make sure the model isn’t just memorizing the training data.</p><h3>Dataset</h3><p>The size of your training data, and how many examples you use to validate your training, are another well-explored hyperparameter. Assuming your data and split are representative and unbiased samples, the larger the dataset, the better the performance — and of course the longer each of your hypotheses takes to test. I follow the conventional wisdom of the 80/10/10 split, where 10% of the data is held out until all training and tuning is complete, to get the best sense for the model’s generalization capabilities. In manual mode, I train on the 80% and validate on 10% after every epoch. In practice, an experiment training on 80% and validating on 10% each time can be too slow for hyperparameter search. I lower these requirements, taking about 20–50% of each split as standard for each experiment. I try to get the runtime of one experiment down to a few minutes.</p><p>After each set of experiment runs — which together form a <em>sweep — </em>you can test with the full training and validation sets. Hopefully the observed patterns hold and maybe even yield a slight increase. If adding more data doesn’t help or hurts, I might be overfitting and may need to increase the percentage of data I’m using for each sweep. Validation data is especially tricky to fix here, since the much smaller sample may be too noisy and not representative. One solution is to pick a validation subset randomly each time, keeping in mind that, if anything, you’re solving a harder problem.</p><h3>Iterative sweeps</h3><p>This is a good framework for the hyperparameter search process: iterative sweeps, or repeated phases of tuning on a small subset and then testing on the full training and validation set. In each sweep, you decide, perhaps implicitly, which hyperparameters stay fixed in the code and which ones you will be testing by varying them. So one of your first choices is how much to change at a time: if you run a sweep over every possible setting, or if you test too many values for one hyperparameter, you’ll need to wait days for results.</p><h3>Manageable exploration</h3><p>A general strategy, especially when pressed for time, is to start with a simple architecture or the best existing solution available for the task. Run a few manual tests to pick a good epoch count and train/validation size. Try an exploratory sweep over a small set of hyperparameters (like, 3–10). The more hyperparameters you start with, the longer this stage will be. For each hyperparameter, sample <em>fewer values in a wider range</em> — even if some of the runs fail, you will have a better sense of the search space. I wouldn’t get into very precise or unusual values at this point — common defaults are good, e.g. batch sizes in powers of two, and again not too many per hyperparameter — I can always narrow my focus in later sweeps.</p><p><strong>First candidates for tuning</strong></p><p>I find that exploring the training dynamics — how slow or fast the network learns — is a reliable first pass: one can get easy wins here by falling in the right range. With learning rate and batch size, for example, I can sample just a few values to get most of the information about how sensitive the network is to that hyperparameter. Getting the right order of magnitude for learning rate: 0.01 vs 0.001 vs 0.0001, say, is usually much more impactful than getting the perfect value of 0.0001 versus 0.00015. Optimizers will be highly correlated with learning rate and batch size, but they can have a huge effect on convergence as well and are good to explore (e.g. Adam is great for many applications, but I have seen projects where SGD massively outperforms Adam).</p><p><strong>Next candidates for tuning</strong></p><p>Once I set some ranges for the easier candidates in training dynamics, I pick a few of the more advanced details, like the layer size or number of filters. The overall network architecture is good to explore here, but I would avoid making too many changes at once or trying too many detailed hypotheses in one sweep — again because that produces a lot more cases to test. As an example, I would construct two or more architectures that capture the general difference between my ideas — let’s say, a regular RNN and a birdirectional RNN, or a CNN with max versus average pooling, and use those as categorical variables. Another way is to parametrize the number of convolutional blocks. But if one hyperparameter is the number of convolutional blocks, another is the number of filters in each convolutional layer, another is max or average pooling, I encounter the very combinatorial explosion I’m trying to avoid by constraining these sweeps to tight, manageable iterations. On this note, dropout can be good to test, but usually after I’m already learning well — when I increase it too early in my exploration process, a good fraction of my runs don’t learn much at all. One could throw in other details here, like weight decay, a learning rate schedule, training stages, number of layers to freeze, and so on. The possibilities are endless, and my main advice is, don’t try to capture too much too soon — stick with a simpler sweep than you might want first, and see what you learn.</p><h3>Babysitting and being done</h3><p>How can you tell how well your sweep is performing and when to stop it? I am often obsessively refreshing my terminal output to see the slowly-ticking stats. With a tool like <a href="https://www.wandb.com/sweeps">Weights &amp; Biases</a>, this part is much more pleasant because I can watch a real-time plot of my loss and accuracies instead. Still, this is not a productive use of human hyperparameter tuning powers. For any model on which I’m running a hyperparameter sweep, I’ve already verified that the training and validation losses generally decrease, and that the training and validation accuracies (or any other metrics of performance) generally increase. Once you’ve debugged the code and are sure the sweep is running correctly and logging results somewhere, I would strongly suggest not refreshing it, if you can resist the allure of improving metrics. Unlike in the manual exploration mode, experiment B is no longer conditioned on seeing how a hypothesis worked out in experiment A. Instead, we’re letting the automatic hyperparameter tuning <em>work for us</em>. An hour or a few hours is generally plenty of time for a sweep, assuming a single run can be constrained to a few minutes. With Weights &amp; Biases it’s also easy to <a href="https://docs.wandb.com/sweeps">scale a sweep</a>. Run the same launch command on another machine, and the logic managing the next choices of hyperparameters will parallelize across both machines, or agents. This lets you run a sweep across as many GPUs as you can get, without the cognitive overhead of tracking which experiment is on which machine.</p><h3>Setting constraints</h3><p>It’s useful to set a limit in advance, such as the total number of experiments I’d like to try or the amount of compute or time I want to spend, as much for efficiency as for my own sanity. I’ve wasted more time than I’d like to admit watching sweeps, hoping there’s a magical combination just around the corner. Really there’s no optimal answer — my heuristic is something like, stop if the pattern of results looks like it’s not changing or not going to yield anything surprising soon. At worst, I stop if I’ve already run for twice as long as it took to get the last best value.</p><h3>Understanding your results</h3><p>Now for the analysis. This phase is crucial — dig into the results and ideally summarize, in writing, any patterns and especially the relationships you observe. For example, does decreasing learning rate tend to increase validation accuracy? Does increasing dropout help, but only when there are at least three layers? Consider what experiments would confirm or disprove your observations, and what you want to try next. The more concrete I can make these notes, the easier it is for me to be intentional and for my future self — let alone teammates or a wider audience — to understand what the heck my past self was trying to do and why.</p><p><strong>Parallel coordinates plot</strong></p><p>It’s really helpful to visualize how all of the hyperparameters interact with the goal metric. The Weights &amp; Biases parallel coordinates plot is a particularly convenient way to do this. The rightmost column plots the result metric (e.g. validation accuracy, validation loss) of each run in a sweep. A wavy line connects this outcome to all the corresponding hyperparameter values for that particular experiment: these are the points at which the wavy line crosses each of the other vertical number lines (e.g. learning rate, batch size, dropout). Each wavy line thus represents a single experiment and is colored according to the outcome (e.g. more purple if the accuracy is lower, more yellow if it is higher). This lets you see correlations at a glance. In the example below, intermediate values for momentum are better, and setting both learning rate and momentum high leads to lower validation accuracy. You can read more about and <a href="https://app.wandb.ai/stacey/fmnist_sweep/reports/Hyperparameter-Relationships-in-Fashion-MNIST--VmlldzozNjY3NA">explore this plot interactively here</a>, as well as in this <a href="https://app.wandb.ai/stacey/pytorch_intro/reports/Meaning-and-Noise-in-Hyperparameter-Search--Vmlldzo0Mzk5MQ">longer report on quantifying significance</a> in hyperparameter searches.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/560/0*sedWfB5YtuwFJUkF.png" /><figcaption>Visualizing 138 CNNs training on Fashion-MNIST. Image by author.</figcaption></figure><p><strong>Parameter importance panel</strong></p><p>The Weights &amp; Biases parameter importance panel makes these visual correlations more concrete. It runs a random forest over your sweep progress (i.e. which hyperparameter values led to which results) to calculate the importance and correlations of each hyperparameter to a result metric. As a bonus, this metric doesn’t have to be the one you were trying to optimize in the sweep itself. Below, I show the parallel coordinates plot and two parameter importance panels for a <a href="https://app.wandb.ai/stacey/deep-drive/reports/The-View-from-the-Driver&#39;s-Seat--Vmlldzo1MTg5NQ">UNet trained in Fastai for semantic segmentation on driving scenes</a> (identify each pixel as belonging to a car, human, bicycle, building, sidewalk, etc, for 20 classes total). My performance metric is IOU: Intersection over Union averaged over all my classes, which will be highest when my model’s predicted regions overlap identically with the ground truth annotations and lower when my predicted regions do not match the ground truth. The first plot clearly shows that ResNets are better encoders than Alexnet to increase IOU, while the effect of learning rate is inconclusive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/779/0*5N5USunsVC3h_jrx.png" /><figcaption>The effect of learning rate is hard to disentangle. Image by author.</figcaption></figure><p>When I consult the parameter importance panel on overall IOU across classes (left chart below), I see that learning rate has a strong negative correlation: I should set the learning rate lower to increase average IOU. However, if I query for parameter importance with respect to <em>human</em> IOU — IOU for the class of humans alone — I see that using Alexnet as an encoder is much more important, and this is the best choice if my goal is to detect humans. You can <a href="https://app.wandb.ai/stacey/deep-drive/reports/The-View-from-the-Driver&#39;s-Seat--Vmlldzo1MTg5NQ">read this report for more details</a>: basically, Alexnet tends to predict larger false positive regions, improving IOU for a rare class like humans but weakening performance overall.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/704/0*5aTbLJSKiAx6AjUU.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/656/0*jrE44lfEkIS18LKd" /><figcaption>Hyperparameter importance with respect to overall average IOU (left) and human IOU only (right). Images by author.</figcaption></figure><h3>Sweep, visualize, repeat</h3><p>Once you’ve analyzed your sweep, documented your findings, and made a plan for the next steps, you can run a few more manual tests or try another sweep. I would recommend expanding the window in the recommended direction (e.g. if lower learning rate increases validation accuracy, try even lower values) or increasing the precision (sampling more closely around the best value) on more fruitful hyperparameters. If any hyperparameter settings are clearly great, note that and consider promoting these to constants: keeping these already-well-tuned values fixed, perhaps with a comment in the code to explain why. On the other hand, if any settings are breaking your script with outliers (e.g. a too-large batch size causes an OOM error), narrow the range for future runs. You can also add a few new hyperparameters to explore, and you’re all set to repeat this process! In Weights &amp; Biases, you can even <a href="https://docs.wandb.com/sweeps/add-to-existing#seed-a-new-sweep-with-existing-runs">start a sweep from existing runs</a>: mark some of your best runs from the sweep you just finished to use as inspiration or a rough draft of your next sweep configuration. They will be taken into account when creating the sweep, and their config not repeated, in running the next sweep.</p><p>I hope this overview helps you tune deep learning models a bit faster regardless of the framework or tools you use. At the high-level, I recommend constrained exploration (a bigger range of values for fewer hyperparameters) and tightening your feedback loop (shorter experiments, smaller sweeps, done in more iterations) as much as possible. Here are <a href="https://www.wandb.com/sweeps">more resources to get started with Weights &amp; Biases Sweeps</a>. I’d love to hear from you how we can make W&amp;B Sweeps work even better for your use cases.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3799df1a4d45" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/hyperparameter-search-with-iterative-sweeps-3799df1a4d45">Hyperparameter Search with Iterative Sweeps</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deep Learning for Climate Adaptation: Detecting Drought from Space]]></title>
            <link>https://medium.com/@staceysvetlichnaya/deep-learning-for-climate-adaptation-detecting-drought-from-space-b633b429d791?source=rss-bf69d40ad788------2</link>
            <guid isPermaLink="false">https://medium.com/p/b633b429d791</guid>
            <category><![CDATA[social-good]]></category>
            <category><![CDATA[climate-change]]></category>
            <category><![CDATA[hacking]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Stacey Svetlichnaya]]></dc:creator>
            <pubDate>Tue, 13 Aug 2019 21:21:57 GMT</pubDate>
            <atom:updated>2019-08-13T21:24:23.369Z</atom:updated>
            <content:encoded><![CDATA[<p>by Stacey Svetlichnaya and Andrew Hobbs</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XPtlpvtHFRrjtG1VHP4tZQ.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@mickaeltournier?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Mickael Tournier</a> on <a href="https://unsplash.com/search/photos/satellite-river?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>TL;DR — our new project helps drought resilience through deep learning on satellite images — join our efforts <a href="https://app.wandb.ai/wandb/droughtwatch/benchmark">here</a>.</p><h3>The Drought Watch benchmark</h3><p>The <a href="https://www.wandb.com/benchmarks">Weights &amp; Biases Benchmarks</a> feature offers a centralized platform for open and efficient collaboration on deep learning projects. The latest is Drought Watch, developed with Andrew Hobbs, Nathan Jensen, and Jin Baek through our recent<a href="https://www.wandb.com/applied-deep-learning"> Applied Deep Learning class</a> and the <a href="https://www.cv4gc.org/">Computer Vision for Global Challenges workshop</a> at CVPR 2019. The goal of this project is to improve drought detection from satellite images so<a href="https://www.resilience.org/stories/2019-07-26/tackling-drought-in-kenya-livestock-insurance-policy-to-help-pastoralists-beat-climate-change/"> index insurance can more effectively support Northern Kenyan families</a>, especially as climate conditions worsen.</p><h3>The challenge</h3><p>The dataset contains about 100,000 satellite images of Northern Kenya in 10 frequency bands, collected by the <a href="https://www.ilri.org/">International Livestock Research Institute</a>. Local experts (pastoralists, or nomadic herders) manually labeled the forage quality at the corresponding geolocations — specifically, the number of cows from {0, 1, 2, 3+} that the location <em>at the center of the satellite image</em> can feed. Each satellite image is 1.95km across, and each pixel in it represents a 30 meter square. Pastoralists estimate the forage quality within 20 meters when they stand on location, an area slightly larger than a pixel in the full 65x65-pixel satellite image. The satellite images thus provide a lot of extra context, which may prove useful since forage quality is correlated across space. The challenge is to learn a mapping from a satellite image to forage quality so we can more accurately predict drought conditions. Furthermore, the current labeling is very sparse, and we want dense predictions of forage quality at any pixel in a satellite image, not just at the center.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kYcA4CbsBa2moMMeeSVHQQ.png" /></figure><h3>The index insurance strategy</h3><p><a href="https://basis.ucdavis.edu/agricultural-index-insurance-economic-development">Index insurance programs</a> [1] help thousands of pastoralists — farmers in Northern Kenya who travel with their cattle searching for forage. The programs monitor environmental conditions and use their predictions of drought to send resources to families in affected areas (e.g. money so herders can keep their animals alive, replace those that die, and/or invest in alternatives). A signature example is the Kenya Livestock Insurance Programme (KLIP), launched in 2014 to protect pastoralists entirely reliant on livestock, which has covered over 150,000 herders [2]. Such programs currently use a normalized differenced vegetation index (NDVI) from near-infrared and red bands of satellite imagery. This does not distinguish between edible and inedible plants and only considers two of the ten available frequency bands. We hope to collaborate on a computer vision model that can better leverage satellite imagery and this large expertly-labeled dataset. With more accurate models, insurance companies can monitor drought conditions — and send resources to families in the right areas — more effectively, lowering costs and making the program accessible to more farmers.</p><h3>Long-term impact opportunity</h3><p><a href="https://www.un.org/press/en/2018/gaef3499.doc.htm">According to the UN</a>, droughts and floods caused by climate change are already devastating agriculture in developing countries. This exacerbates hunger and poverty, as 75% of the poor in these countries <a href="http://www.fao.org/fileadmin/templates/wsfs/docs/Issues_papers/HLEF2050_Global_Agriculture.pdf">rely on agriculture to survive</a>. Increasingly extreme weather patterns threaten over 50 million pastoralists in Africa, and countries including Ethiopia, Niger, and West Africa seek to replicate KLIP, with a target of reaching 10 million herders. Another initiative, the G7 InsuResilience, intends to reach 400 million households worldwide by 2020 [3]. Index insurance and resilience-based social protection programs are a promising sustainable strategy for managing and <a href="https://basis.ucdavis.edu/publication/policy-brief-resilience-based-protection-blunts-climate-changes-impact-poverty">reducing poverty in the long-term</a>. Improving the accuracy of models like this one is an opportunity to make a concrete, scalable impact on many people’s lives and our collective climate resilience — all while honing your deep learning skills.</p><h3>Code for the changing climate</h3><p>You can contribute to this project by <a href="https://app.wandb.ai/wandb/droughtwatch/benchmark">joining the benchmark here</a>.</p><p>You will see simple instructions to clone the repo, download the data, and start training models in a few lines of code. As you run training scripts, you’ll see a link to a wandb page for each run, showing real-time graphs of relevant metrics. In the browser, your results will show up in the “Project workspace” tab to the left.</p><p>With wandb, you can easily visualize the training and explore various hyperparameters and architectures. We encourage you to include your code and share notes, findings, or observations about your process. You can use this platform to see what others have tried, discuss ideas, and synthesize different approaches. We hope to grow a collaborative community to learn, advance the state of the art, and build a better future together.</p><p>[1] How does index insurance like this work? Agricultural index insurance uses a common metric like rainfall, average yield per area, or vegetation growth to statistically model crop or livestock production with varying accuracy and computational cost. Indices that can be measured remotely, like rainfall or drought conditions from satellite, have lower computational costs — we can increase prediction accuracy through various post-processing techniques without needing to make additional local measurements. By relying on a shared, external, and maximally objective metric, index insurance avoids two main issues with regular insurance (aside from high cost): adverse selection (where only the folks more likely to suffer losses pay for the insurance) and moral hazard (counterproductive action with the intent of collecting a payout). This leaves the main cost: the basis risk, or the error between the index’s estimate and an individual farmer’s real losses. As agricultural indices of conditions like drought become more accurate with technological advances (in remote sensing, satellite imaging, and now computer vision and deep learning), the basis risk declines, reducing the cost of the program and making it accessible to more farmers. To measure the effectiveness of index insurance programs, economists consider the Minimum Quality Standard, comparing the program’s value over time — in terms of potential for income stabilization — to no insurance program and equivalent direct cash transfer. They also study “ex-ante” effects, increased investments in future productivity that pastoralists are able to make with the knowledge of coverage. <a href="https://basis.ucdavis.edu/agricultural-index-insurance-economic-development">You can learn more here</a>.</p><p>[2] <a href="https://www.resilience.org/stories/2019-07-26/tackling-drought-in-kenya-livestock-insurance-policy-to-help-pastoralists-beat-climate-change/">https://www.resilience.org/stories/2019-07-26/tackling-drought-in-kenya-livestock-insurance-policy-to-help-pastoralists-beat-climate-change/</a></p><p>[3] <a href="https://basis.ucdavis.edu/agricultural-index-insurance-economic-development">https://basis.ucdavis.edu/agricultural-index-insurance-economic-development</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b633b429d791" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>