Issue #307 - The ML Engineer ??
Alejandro Saucedo
AI & Data Executive @ Zalando | Advisor @ UN, EU, ACM, etc | Join 70k+ ML Newsletter
Thank you for being part of over 60,000+ ML professionals and enthusiasts who receive weekly articles & tutorials on Machine Learning & MLOps ?? You can join the newsletter for free at https://ethical.institute/mle.html ?
If you like the content please support the newsletter by sharing with your friends via ?? Email, ?? Twitter, ?? Linkedin and ?? Facebook!
This week in Machine Learning:?
Almost 30% use a custom built in-house tool for ML model training, with Databricks being the 2nd most popular choice with 20%, closely followed by AWS SageMaker. Insightful results from our survey on the State of Prod ML in 2024; it seems that although the area of ML model training is one of the most consolidated areas in regards to tooling, there is still quite a significant percentage of organisations that do not use an off-the-shelf framework and build custom in-house tools for their ML Training. Not surprisingly it does seem like the main choices are the cloud providers, however with quite a large gap between Azure/Google vs AWS/Databricks. If you have a chance we would be grateful if you could spend a few minutes on the survey, as you'll contribute valuable information about the machine learning tools and platforms you use in your production ML development. Your input will help create a comprehensive overview of common practices, tooling preferences, and challenges faced when deploying models to production, ultimately benefiting the entire ML community ???
This is an exciting time for time series machine learning, with a new Billion-parameter Mixture-of-Experts Forecasting model entering the arena released by Princeton, Squirrel AI and Griffith University, together with the largest OSS TS dataset: Following the steps of Google, Amazon, Nixtla and others, Time-MoE is the latest time series foundation model which brings new innovative approaches with a Mixture-of-Experts Transformer architecture aiming to handle universal zero-shot forecasting. They have been able to scale this model to 2.4 billion parameters by pre-training on "Time-300B", the largest open-access time series dataset comprising over 300 billion data points from more than nine domains. It is an exciting time for research in this space, not only new models and architectures being released, but also huge datasets to support further benchmarking and research.
TimescaleDB comes with a controversial take on Vector Databases - "they are the wrong abstraction": In production AI applications, managing vector embeddings tends to bring complexities due to vector databases treating embeddings as standalone data disconnected from their source - this is raised as an abstraction that should be addressed as it would otherwise lead to synchronization issues and stale data. TimescaleDB proposes treating embeddings as derived data similar to database indexes, which is interesting given recent extensions from DBs like planetscale to integrate embeddings natively into indexes, similarly through a "native vectorizer" abstraction. In this case however they still leverage the OSS pgai Vectorizer for PostgreSQL which helps automating the synchronization of embeddings with their source data within the database, however this does provide an insight on some of the open challenges that are yet to be addressed, and perhaps also one of the reasons why we still see lack of standardisation on VectorDBs in the State of Prod ML 2024 Survey.
Although many may claim that LLMs and GenAI are driving the current AI revolution, there is a strong reason to believe that one of the biggest breakthroughs and drivers are the innovations within embeddings, implemented beyond just text: Vector embeddings have grown to become a hugely powerful tool in ML, not only powering the GenAI / LLM products that have been blowing minds recently, but also powering things like product recommendations, similarity search and beyond. This is a great intuitive overview of embeddings with a few practical examples that can provide a high level conceptual understanding for individuals that may not have come across some of these internals - there are other interesting applications, such as how AirBNB used embeddings for similar listing recommendations all the way back in 2018.
One of the most pressing challenges when deploying GenAI products is ensuring robust guardrails are in place - companies are investing heavily on increasing the safety of their products, but data poisoning and jailbreaking attacks keep advancing at a faster pace, and here is a great example: Even advanced and robust GenAI powered products like GPT-4o are suceptible to security vulnerabilities - this is a great example showcasing data poisoning and a potent new attack called "jailbreak-tuning," where attackers inject harmful behaviors by fine-tuning models on poisoned datasets, even when advanced moderation systems are in place. This is a great deep dive into three threat models, including 1) malicious fine-tuning, 2) imperfect data curation, and 3) intentional data contamination, which really make it clear that that larger models become more susceptible to these attacks as they scale. This also makes it clear that it only continues to become more critical to ensure robust evaluation metrics, fine-tuning safeguards, and ongoing stress-testing to prevent exploitation of these vulnerabilities in production environments - this means that organisations also have to invest into their existing security / red-teaming capabilities to support this.
领英推荐
Upcoming MLOps Events
The MLOps ecosystem continues to grow at break-neck speeds, making it ever harder for us as practitioners to stay up to date with relevant developments. A fantsatic way to keep on-top of relevant resources is through the great community and events that the MLOps and Production ML ecosystem offers. This is the reason why we have started curating a list of upcoming events in the space, which are outlined below.
Upcoming conferences where we're speaking:
Other upcoming MLOps conferences in 2024:
In case you missed our talks:
Open Source MLOps Tools
Check out the fast-growing ecosystem of production ML tools & frameworks at the github repository which has reached over 10,000 ? github stars. We are currently looking for more libraries to add - if you know of any that are not listed, please let us know or feel free to add a PR. Four featured libraries in the GPU acceleration space are outlined below.
? If you know of any open source and open community events that are not listed do give us a heads up so we can add them!
OSS: Policy & Guidelines
As AI systems become more prevalent in society, we face bigger and tougher societal challenges. We have seen a large number of resources that aim to takle these challenges in the form of AI Guidelines, Principles, Ethics Frameworks, etc, however there are so many resources it is hard to navigate. Because of this we started an Open Source initiative that aims to map the ecosystem to make it simpler to navigate. You can find multiple principles in the repo - some examples include the following:
If you know of any guidelines that are not in the "Awesome AI Guidelines" list, please do give us a heads up or feel free to add a pull request!
About us
The Institute for Ethical AI & Machine Learning is a European research centre that carries out world-class research into responsible machine learning.
Data Enthusiast | Mentor | AWS | Spark | Kafka | Java
4 个月Insightful
Lead Design Engineer | Machine Learning | Deep Learning | Computer Vision | Data Analysis | Statistics | Decision Making | Science engineer | Researcher
4 个月?? ??