Hint: forget about the p-values — Data drift is a concern to anyone with a machine learning model serving live predictions. The world changes, and as the consumers’ tastes or demographics shift, the model starts receiving feature values different from what it has seen in training, which may result in unexpected outputs. Detecting feature drift appears…

Are vision foundational models catching up with LLMs? — Self-supervised training methods continue to deliver breakthrough after breakthrough. Last week, Meta AI released the second version of their self-DIstillation with NO labels or DINO model. The model can supposedly be used as a backbone to solve virtually any computer vision task without fine-tuning! Have the foundational models in computer…

A new foundational model for computer vision, making sense of the spree of open-source LLMs, and should AI be open-source? — AI Pulse is also available at pulseofai.substack.com. In this edition: DINOv2, a universal computer vision backbone; A spree of open-source LLMs emerges following LlaMa’s leak; Should AI models be open-sourced?

Reduce your models' latency, storage, and inference costs with quantization and pruning — Over the last few years, machine learning models have seen two seemingly opposing trends. On the one hand, the models tend to get bigger and bigger, culminating in what’s all the rage these days: the large language models. Nvidia’s Megatron-Turing Natural Language Generation model has 530 billion parameters! On the…

Bard, Sparrow, and multimodal chatbots will render it obsolete soon, and here is why. — ChatGPT, the OpenAI chatbot released last autumn, has taken the Internet by storm. Arguably, no other machine learning model has ever made so many headlines outside of the AI community. It provides a near-human experience when talked to and can help many of us do our work faster: from SEO…

You don’t need a perfect oracle to get your probabilities right. — Machine learning models are often evaluated based on their performance, measured by how close some metric is to zero or one (depending on the metric) but this is not the only factor that determines their usefulness. In some cases, a model that is not very accurate overall can still be…

How to train models with only a few labeled examples — A large portion of the value delivered by AI so far comes from supervised models trained on increasingly large datasets. Many of these datasets have been labeled by humans, a job that is mundane, time-consuming, error-prone, and sometimes expensive. Self-supervised learning (SSL) is a different learning paradigm allowing machines to…

Good old PCA can alert you when the distribution of your production data changes — The ultimate purpose of any machine learning model is to bring value to its owner. Typically, this value comes in the form of the algorithm doing things better or faster (or both) than a human can. The cost of investment in developing and deploying the model is often high. In…

Compress the model for your edge device without losing the accuracy — As the popular cliché has it, data scientists spend 80% of their time preparing the data, and only 20% developing the models. While there might be some truth to it, we should never underestimate the effort needed for the remaining 20%. Choosing the architecture, training, fine-tuning, and evaluating the model…