đź«’ Olive these puns are here to make you smile
LinkedIn spills GenAI secrets, Stanford proposes synthetic medical data, I cover Operation Olive Branch and ML models for changing conditions. Enjoy the read!
Late to the Party 🎉 is about insights into real-world AI without the hype.
Hello internet,
it finally feels like summer is coming! I’ll head out and enjoy the weather right after I hit send!
In this issue, we have LinkedIn spilling the beans on GenAI, making great conference talks, and Stanford proposing stable diffusion could make synthetic medical data. I talk about Operation Olive Branch and saving Maram’s family, and I write about how to make machine learning models adaptable to changing environmental conditions. And I saw the aurora!
Let’s dive right in!
The Latest Fashion
- LinkedIn Engineering published a blog on their implementation of generative AI, which is quite insightful
- This extensive blog on making a great conference talk covers it all
- Stanford HAI published on Stable Diffusion for synthetic medical images (keep in mind they have a vested interest in “foundational models” while reading)
Worried these links might be sponsored? Fret no more. They’re all organic, as per my ethics.
My Current Obsession
I think I’m not alone in saying that it’s terrifying what is going on in the world right now. The fact that this could mean multiple possible things, from multiple wars and the rise of the extreme right to climate change, doesn’t make this better. This week, however, it has been the ongoing genocide in Palestine for me and the suspicious activity around Rafah lately. I feel powerless as governments seem to support the ethnic cleansing of the area, which is especially terrifying as a German. But it seems like campus protests around the world may slowly be shifting the sentiment. In the meantime, while governments are making up their minds, I found a grassroots movement called Operation Olive Branch that is collecting and vetting pleas for help from Palestinian families for support to possibly evacuate, buy food or get urgent medical care. This is particularly important, as scammers are somehow amoral enough to try and benefit from this atrocity. Can you believe it?!
They made a FAQ and an explainer video, and here’s the massive spreadsheet!
If you get overwhelmed by big spreadsheets, like me, you could help Maram’s family reunite across the border.
Or scroll the “Finisher list”, which has many fundraisers close to the finish line (and some pleasant surprises with updates from families that have escaped!)
Any small bit helps.
Interestingly, all this is happening while celebrities are attending the Met Gala, a $ 75,000-per-seat event with lavish gowns and lots of press. Unfortunately (for them), one of them made a TikTok saying, “Let them eat cake” in a Marie Antoinette-inspired dress. Now, instead of outrage, we have #BlockOut2024, where people started blocking celebrities to deny them ad revenue. So, I’ve been having some fun going through celebrity accounts and blocking them. But let’s be honest; I was never their target audience, to begin with.
Lastly, a treat for everyone who reads these! I saw the polar lights in the middle of Germany yesterday. There’s a strong likelihood to see them again today, so maybe check if you might get lucky to see the spectacle tonight! I posted images on the Latent Space.
Thing I Like
Let’s be honest… I’m moving next week, I regret each and every purchase I have ever made in my life and have to pack into boxes right now.
Hot off the Press
Python Deadlines
I added the conference dates for PyCon Africa, but the CfP isn’t open yet.
This week, three CfPs are closing: PyData Eindhoven, PyCon Portugal, and PyCon Russia.
I’ve done more digital archaeology, finding CfPs and conference dates for PyCon APAC all the way back to 2012.
Machine Learning Insights
Last week, I asked, How can machine learning models be designed to be more adaptable to changing environmental conditions?, and here’s the gist of it:
The worry of machine learning models only working on "historical data" is strong everywhere. Making these models adaptable to changing environmental conditions is crucial, especially in fields like meteorology and environmental monitoring, where the input data can vary significantly over time due to climate change and other factors. Here are some strategies for making machine learning models more adaptable. Sambit Kumar Panda had a fun answer on Linkedin! Here are some more thoughts on it:
Transfer Learning and Domain Adaptation
Transfer learning involves adapting a model trained on one task to a new, related task. This approach is particularly useful in environmental sciences, where data from one geographical location or time period can be leveraged to improve predictions in another. For example, a model trained to predict temperature patterns in one region might be fine-tuned to predict similar patterns in a different region with only slight modifications. Similarly, domain adaptation techniques are used when training and test data come from different distributions, which is often the case in environmental datasets due to changes in sensor technologies, methodologies, or environmental shifts. Techniques like feature normalisation and domain-invariant feature extraction help the model perform well across different data distributions.
Online Learning
Online learning is a method in which the model is continuously updated as new data arrives rather than being trained on a fixed dataset. This type of learning is ideal for environmental applications where conditions change over time. For instance, a model predicting air quality can adapt to new pollution sources by updating its parameters regularly as new data is collected. However, online learning itself is really tricky to stabilise and evaluate, so keep that in mind.
Ensemble Methods
Ensemble methods, which combine multiple models to make a prediction, can enhance adaptability by integrating diverse perspectives. This is particularly effective in environmental models, where different models might be better suited to different conditions. For example, one model might excel in dry conditions while another performs better in humid conditions. In fact, physical models for weather and climate prediction are usually run as ensembles to get a distribution of scenarios!
Using Robust and Resilient Models
Models inherently robust to noise and missing data tend to adapt better to varying conditions. Techniques such as regularisation and Bayesian methods can impart a degree of resilience to the model, allowing it to perform consistently even under different environmental conditions.
Hybrid Models
Similarly to Bayesian methods, combining physical models with machine learning models (physics-informed machine learning) can capitalise on the strengths of both. For example, integrating physical laws with data-driven models can help predict complex environmental phenomena like storm patterns more accurately.
Monitoring Systems
Implementing feedback mechanisms where the model's predictions are periodically evaluated and the model is adjusted based on performance can help adapt to changing environments. This monitoring and re-training is akin to a self-correcting system that evolves based on its accuracy in real-world conditions.
Data Augmentation
Augmenting training data with synthetic data generated to simulate various possible future conditions can prepare the model for a range of scenarios. This is useful in climate modelling, where future conditions may be unprecedented. Interestingly, test-time augmentation can also be used as an ensemble method. The best part is that these methods are not mutually exclusive and are often more effective when combined. By implementing these strategies, machine learning models can be better equipped to handle the dynamic and often unpredictable nature of environmental data, providing more reliable and accurate predictions over time.
Got this from a friend? Subscribe here!
Question of the Week
- Can AI assist in more accurate and timely prediction of volcanic eruptions?
Post them on Mastodon and Tag me. I'd love to see what you come up with. Then I can include them in the next issue!
Tidbits from the Web
- This needs no explanation.
- Help Maram’s family reunite
- Hope you have a better weekend than this guy
Jesper Dramsch is the creator of PythonDeadlin.es, ML.recipes, data-science-gui.de and the Latent Space Community.
I laid out my ethics including my stance on sponsorships, in case you're interested!