Press "Enter" to skip to content

Category: Machine Learning

Integrating Azure ML and Power BI

I have a new video:

In this video, I show off how easy it is to integrate Azure ML and Power BI, at least once you get past all of the trouble trying to integrate them.

I expected this to be easy. It turns out that the “make it look easy” depends on having several things in place already and using the correct (by which I mean “old”) deployment type.

Leave a Comment

Getting Started with Semantic Kernel in C#

Matt Eland tries out Semantic Kernel:

Generative AI systems use large language models (LLMs) like OpenAI’s GPT 3.5 Turbo (ChatGPT) or GPT-4 to respond to text prompts from the user. But these systems have serious limitations in that they only include information baked into the model at the time of training. Technologies like retrieval augmentation generation (RAG) help overcome this by pulling in additional information.

AI orchestration frameworks make this possible by tying together LLMs and additional sources of information via RAG. Additionally, AI orchestration systems can provide capabilities to generative AI systems, such as inserting records in a database, sending emails, or calling out to external systems.

In this article we’ll look at the high-level capabilities building AI orchestration systems in C# with Semantic Kernel, a rapidly maturing open-source AI orchestration framework.

Click through to see how things work.

Leave a Comment

Oracle OCI Labeling with Bounding Boxes

Brendan Tierney continues a series on image classification:

In a previous post, I gave examples of how to label data using OCI Data Labeling. It was a simple approach to data labeling images for input to AI Vision. In that post, we just gave a label for the image to indicate if the image contained a Cat or a Dog. Yes, that’s a very simple approach, and we can build image classification models, and use the resulting model to predict a label for new images. These would be labeled as a Cat or a Dog with a degree of certainty. Although this simple approach can give OK-ish results, we typically want a more detailed model and predictions. For a more detailed approach, we can use Object Detection. For this, we need to prepare our data set in a slightly different way and Yes it does take a bit more time to prepare. Or perhaps it takes a lot more time to prepare the data. But this extra time in preparing the data should (in theory) give us a more accurate model.

This post will focus on creating a new labeled dataset using bounding boxes, and in a later post, we’ll examine the resulting model to see if it gives better or more accurate results.

Read on for the process.

Comments closed

The Rise of Single-Purpose ML Frameworks

Pete Warden describes a phenomenon:

The GGML framework is just over a year old, but it has already changed the whole landscape of machine learning. Before GGML, an engineer wanting to run an existing ML model would start with a general purpose framework like PyTorch, find a data file containing the model architecture and weights, and then figure out the right sequence of calls to load and execute it. Today it’s much more likely that they will pick a model-specific code library like whisper.cpp or llama.cpp, based on GGML.

This isn’t the whole story though, because there are also popular model-specific libraries like llama2.cpp or llama.c that don’t use GGML, so this movement clearly isn’t based on the qualities of just one framework. The best term I’ve been able to come up with to describe these libraries is “disposable”. I know that might sound derogatory, but I don’t mean it like that, I actually think it’s the key to all their virtues! They’ve limited their scope to just a few models, focus on inference or fine-tuning rather than training from scratch, and overall try to do a few things very well. They’re not designed to last forever, as models change they’re likely to be replaced by newer versions, but they’re very good at what they do.

Pete calls them disposable ML frameworks, though I’d call them single-purpose frameworks to contrast with general-purpose ML frameworks like PyTorch and TensorFlow.

Comments closed

Creating an Image Classification Model in Oracle OCI Vision

Brendan Tierney separates the cats and the dogs:

In this post, I’ll build on the previous work on preparing data, to using this dataset as input to building a Custom AI Vision model. In the previous post, the dataset was labelled into images containing Cats and Dogs. The following steps takes you through creating the Customer AI Vision model and to test this model using some different images of Cats.

This post is part four of a series (first part, second part, third part) on custom image classification in Oracle.

Comments closed

ML with Keras and TensorFlow over Streaming Kafka Data

Paul Brebner gives us a streaming scenario for model training:

One of the goals of incremental learning is to train a model continuously from streaming data. Incremental learning from streaming data means you don’t need all the data in memory at once, and the model is as up-to-date as possible, which can matter for real-time use cases. The third driver for incremental learning that I mentioned in the previous blog is when there is concept drift in the data itself—but we’ll ignore this aspect for the time being. 

In the last blog we demonstrated batch training with TensorFlow, and mentioned that TensorFlow, being a neural network framework, has the potential for incremental learning—just like animals and people do. In this blog, we will set ourselves the task of using TensorFlow to demonstrate incremental learning from the same static drone delivery data set of busy/not busy shops that we used in the last blog. 

Read on to see the code, results, and warnings.

Comments closed

A Primer on Vector Search

Phil Booth takes a look at vector search systems:

Recently I built a system that uses vector search to logically truncate long documents and retain the most significant parts according to some search term. I’m a dummy, with no background in machine learning or mathematics, so there were new concepts for me to understand and implementation details to figure out. This post summarises what I learned.

Vector search and vector databases are becoming a fairly hot topic, so this at least grounds you on what they are.

Comments closed

BotChat BiWeekly

Mala Mahadevan starts a newsletter:

I do my best to find trustworthy sources to learn from, but you know how it is – sometimes it’s tough to tell what’s legit. So, if you ever see me post something that seems a bit off, please cut me some slack. These aren’t necessarily my opinions, just things that caught my eye.

What I learn is just my take on what I heard or read. It might not always jive with what the original speaker or writer means, or understand. I don’t use any fancy AI bots like ChatGPT to help me out. I just quote stuff and break it down in my own words.

Mala focuses on a pair of videos. I snuck into the newsletter with a few bomb-throwing statements, particularly around anthropomorphism (the assignment of human or human-like qualities to non-humans). Anthropomorphism is extremely common in language. It’s all well and good as metaphor, but once you start to believe it for real, that’s when you end up in trouble.

Comments closed

Training a Code-First Model in Azure ML

I have a new video:

In this video, we walk through the code in an Azure Machine Learning project and see how the pieces fit together.

There are a few more videos to go in this Azure ML series and I would recommend going through them in order to understand how we got to this video, but this one is what I’ve been building toward.

Comments closed

SeamlessM4T: Multimodal Speech and Text Translation

Facebook has announced a new library:

Today, we’re introducing SeamlessM4T, the first all-in-one multimodal and multilingual AI translation model that allows people to communicate effortlessly through speech and text across different languages. SeamlessM4T supports:

  • Speech recognition for nearly 100 languages
  • Speech-to-text translation for nearly 100 input and output languages
  • Speech-to-speech translation, supporting nearly 100 input languages and 36 (including English) output languages
  • Text-to-text translation for nearly 100 languages
  • Text-to-speech translation, supporting nearly 100 input languages and 35 (including English) output languages

The open source library is available on GitHub and you can also get the model itself on HuggingFace. The nicest thing about all of this is that, unlike existing translation services, you can run it entirely offline and perform the inference on local compute.

Comments closed