XAI = Explainable Artificial Intelligence

Dec 14 2023

Alessio Malizia and Fabio Paternò write:

Numerous papers argue for using XAI methods in the literature, as well as multiple suggestions for brand-new XAI family approaches. Nevertheless, finding instances of practical XAI technique implementations that have enhanced the business in industry/societal/real-world applications is more challenging, even if some interesting work in this area has been put forward, for example in the health domain

Read Why Is the Current XAI Not Meeting the Expectations?

Apr 17, 2017 Matt Brems writes:

Principal component analysis (PCA) is an important technique to understand in the fields of statistics and data science… but when putting a lesson together for my General Assembly students, I found that the resources online were too technical, didn’t fully address our needs, and/or provided conflicting information. It’s safe to say that I’m not “entirely satisfied with the available texts” here.

As a result, I wanted to put together the “What,” “When,” “How,” and “Why” of PCA as well as links to some of the resources that can help to further explain this topic. Specifically, I want to present the rationale for this method, the math under the hood, some best practices, and potential drawbacks to the method.

While I want to make PCA as accessible as possible, the algorithm we’ll cover is pretty technical. Being familiar with some or all of the following will make this article and PCA as a method easier to understand: matrix operations/linear algebra (matrix multiplication, matrix transposition, matrix inverses, matrix decomposition, eigenvectors/eigenvalues) and statistics/machine learning (standardization, variance, covariance, independence, linear regression, feature selection). I’ve embedded links to illustrations of these topics throughout the article, but hopefully these will serve as a reminder rather than required reading to get through the article.

Read A One-Stop Shop for Principal Component Analysis

cross-posted from: https://programming.dev/post/9436800

December 28 2023 Pankaj Singh writes:

In big data processing and analytics, choosing the right tool is paramount for efficiently extracting meaningful insights from vast datasets. Two popular frameworks that have gained significant traction in the industry are Apache Spark and Presto. Both are designed to handle large-scale data processing efficiently, yet they have distinct features and use cases. As organizations grapple with the complexities of handling massive volumes of data, a comprehensive understanding of Spark and Presto’s nuances and distinctive features becomes essential. In this article, we will compare Spark vs Presto, exploring their performance and scalability, data processing capabilities, ecosystem, integration, and use cases and applications.

Read Spark vs Presto: A Comprehensive Comparison

cross-posted from: https://lemmy.world/post/11196216

Hi all,

For those wanting a quick repo to use as a basis to get started, I’ve created jen-ai.

There are full instructions in the readme. Once running you can talk to it, and it will respond.

It’s basic, but a place to start.

cross-posted from: https://programming.dev/post/9428234

Apr 18, 2023 Matthijs Cox writes:

The “two language problem” was globally accepted for several decades. You just learn to live with it. Until one day it was challenged by the Julia language, a programming language that promises both speed and ease of use. As I feel the pain of the two language problem deeply, I wanted to try out this new solution. So together with several allies I went on a mission to adopt this new technology at work, and remove the bottleneck.

While we had initial success and attention, we quickly stumbled into resistance from the existing groups of researchers/scientists and developers. Over time I have named this the “two culture problem”. In the beginning I didn’t see the cultures clearly, which limited our success. I was too focused on the technological problem itself.

I will refer to the two cultures as “scientists” versus “developers”. However, the “scientists” group generalizes to anyone who codes quick and dirty to explore, such as domain experts, data analysts and others like that. I do hope everyone is doing their exploration somewhat scientifically, so the generalization should makes sense. Scientists typically want to get their stuff done, perhaps with code, but they don't care about the code. Software developers care deeply about the code craftsmanship, sometimes obsessively so, but often developers barely understand the business domain or science. There are people near the middle, trying to balance both, but they are a rare breed.

Read My Target Audience

The article discusses the Adam optimizer, a popular algorithm in deep learning known for its efficiency in adjusting learning rates for different parameters.

Unlike other optimizers like SGD or Adagrad, Adam dynamically changes its step size based on the complexity of the problem, analogous to adjusting stride in varying terrains. This ability to adapt makes it effective in quickly finding the minimum loss in machine learning tasks, a key reason for its popularity in winning Kaggle competitions and among those seeking a deeper understanding of optimizer mechanics.

AI at Meta writes:

Today we’re releasing Code Llama 70B: a new, more performant version of our LLM for code generation — available under the same license as previous Code Llama models.

Download the models

• CodeLlama-70B
• CodeLlama-70B-Python
• CodeLlama-70B-Instruct

CodeLlama-70B-Instruct achieves 67.8 on HumanEval, making it one of the highest performing open models available today.

CodeLlama-70B is the most performant base for fine-tuning code generation models and we’re excited for the community to build on this work.

Nitter link

Try the model on Hugging Face

Installable using Ollama

Real World Data Science Writes:

The 2024 International Cherry Blossom Prediction Competition will open for entries on February 1, and Real World Data Science is once again proud to be a sponsor.

Contestants are invited to submit predictions for the date cherry trees will bloom in 2024 at five different locations – Kyoto, Japan; Liestal-Weideli, Switzerland; Vancouver, Canada; and Washington, DC and New York City, USA.

Read details about Real World Data Science Competition | When will the cherry trees bloom?

cross-posted from: https://programming.dev/post/8724281

Itamar Turner-Trauring writes:

These sort of problems are one of the many reasons you want to “pin” your application’s dependencies: make sure you only install a specific, fixed set of dependencies. Without reproducible dependencies, as soon as NumPy 2 comes out your application might break when it gets installed with new dependencies.

The really short version is that you have two sets of dependency configurations:

• A direct dependency list: A list of libraries you directly import in your code, loosely restricted. This is the list of dependencies you put in pyproject.toml or setup.py.
• A lock file: A list of all dependencies you rely on, direct or indirect (dependencies of dependencies), pinned to specific versions. This might be a requirements.txt, or some other file dependencies on which tool you’re using.

At appropriate intervals you update the lock file based on the direct dependency list.

I’ve written multiple articles on the topic, in case you’re not familiar with the relevant tools:

• “Faster Docker builds with pipenv, poetry, or pip-tools” covers using those three tools to maintain lockfiles.
• For Conda, see “Reproducible and upgradable Conda environments with conda-lock”.

Read NumPy 2 is coming: preventing breakage, updating your code

January 1, 2024 - Omar Sanseviero writes:

In this blog post, we’ll do an end-to-end example of the math within a transformer model. The goal is to get a good understanding of how the model works. To make this manageable, we’ll do lots of simplification. As we’ll be doing quite a bit of the math by hand, we’ll reduce the dimensions of the model. For example, rather than using embeddings of 512 values, we’ll use embeddings of 4 values. This will make the math easier to follow! We’ll use random vectors and matrices, but you can use your own values if you want to follow along.

Read The Random Transformer | Understand how transformers work by demystifying all the math behind them

cross-posted from: https://programming.dev/post/8391233

Dr. Chris Rackauckas (@chrisrackauckas@fosstodon.org) writes:

#julialang GPU-based ODE solvers which are 20x-100x faster than those in #jax and #pytorch? Check out the paper on how #sciml DiffEqGPU.jl works. Instead of relying on high level array intrinsics that #machinelearning libraries use, it uses a direct kernel generation approach to greatly reduce the overhead.

Read Automated translation and accelerated solving of differential equations on multiple GPU platforms

Sheldon Axler Writes:

I am happy to announce publication of the fourth edition of Linear Algebra Done Right as an Open Access book. The electronic version of this new fourth edition with a Creative Commons BY-NC license is availble without cost at the link below.

• PDF file for Linear Algebra Done Right, fourth edition

The print version of the new fourth edition of Linear Algebra Done Right is now available from Amazon and from Springer at the links below.

• Amazon: print version of Linear Algebra Done Right, fourth edition
• Springer: print version of Linear Algebra Done Right, fourth edition

The fourth edition of Linear Algebra Done Right contains over 250 new exercises and over 70 new examples, along with several new topics and multiple improvements throughout the book. See page xvi in the file linked above for a list of major improvements and additions in the fourth edition.

This best-selling textbook for a second course in linear algebra is aimed at undergraduate math majors and graduate students. The novel approach taken here banishes determinants to the end of the book. The text focuses on the central goal of linear algebra: understanding the structure of linear operators on finite-dimensional vector spaces. The author has taken unusual care to motivate concepts and to simplify proofs. A variety of interesting exercises in each chapter helps students understand and manipulate the objects of linear algebra.

No prerequisites are assumed other than the usual demand for suitable mathematical maturity. Thus the text starts by discussing vector spaces, linear independence, span, basis, and dimension. The book then deals with linear maps, eigenvalues, and eigenvectors. Inner product spaces are then introduced, leading to the finite-dimensional spectral theorem and its consequences such as the singular value decomposition. Generalized eigenvectors are then used to provide insight into the structure of a linear operator. Determinants are cleanly introduced via alternating multilinear forms.

Read Linear Algebra Done Right

January 7, 2024 by Omar Sanseviero:

This series aims to demystify embeddings and show you how to use them in your projects. This first blog post will teach you how to use and scale up open-source embedding models. We’ll look into the criteria for picking an existing model, current evaluation methods, and the state of the ecosystem. We’ll look into three exciting applications:

• Finding the most similar Quora or StackOverflow questions
• Given a huge dataset, find the most similar items
• Running search embedding models directly in the users’ browser (no server required)

You can either read the content here or execute it in Google Colab by clicking the badge at the top of the page. Let’s dive into embeddings!

Read Sentence Embeddings | Everything you wanted to know about sentence embeddings (and maybe a bit more)

Emily Riederer Writes:

Switching languages is about switching mindsets - not just syntax. New developments in python data science toolings, like polars and seaborn’s object interface, can capture the ‘feel’ that converts from R/tidyverse love while opening the door to truly pythonic workflows

Just to be clear:

• This is not a post about why python is better than R so R users should switch all their work to python
• This is not a post about why R is better than python so R semantics and conventions should be forced into python
• This is not a post about why python users are better than R users so R users need coddling
• This is not a post about why R users are better than python users and have superior tastes for their toolkit
• This is not a post about why these python tools are the only good tools and others are bad tools

The Stack

WIth that preamble out of the way, below are a few recommendations for the most ergonomic tools for getting set up, conducting core data analysis, and communication results.

To preview these recommendations:

Set Up

Installation: pyenv
IDE: VS Code

Analysis

Wrangling: polars
Visualization: seaborn

Communication

Tables: Great Tables
Notebooks: Quarto

Miscellaneous

Environment Management: pdm
Code Quality: ruff

Read Python Rgonomics

cross-posted from: https://programming.dev/post/8246313

Data science managers and leaders should make sure that cooperative work on models is facilitated and streamlined. In this post, our very own Shachaf Poran, PhD suggests one method of doing so.

For folks who aren't sure how to interpret this, what we're looking at here is early work establishing an upper bound on the complexity of a problem that a model can handle based on its size. Research like this is absolutely essential for determining whether these absurdly large models are actually going to achieve the results people have already ascribed to them on any sort of consistent basis. Previous work on monosemanticity and superposition are relevant here, particularly with regards to unpacking where and when these errors will occur.

I've been thinking about this a lot with regards to how poorly defined the output space they're trying to achieve is. Currently we're trying to encode one or more human languages, logical/spatial reasoning (particularly for multimodal models), a variety of writing styles, and some set of arbitrary facts (to say nothing of the nuance associated with these facts). Just by making an informal order of magnitude argument I think we can quickly determine that a lot of the supposed capabilities of these massive models have strict theoretical limitations on their correctness.

This should, however, give one hope for more specialized models. Nearly every one of the above mentioned "skills" is small enough to fit into our largest models with absolute correctness. Where things get tough is when you fail to clearly define your output space and focus training so as to maximize the encoding efficiency for a given number of parameters.

~n (@nblr@chaos.social) writes:

This is fine...

"We observed that participants who had access to the AI assistant were more likely to introduce security vulnerabilities for the majority of programming tasks, yet were also more likely to rate their insecure answers as secure compared to those in our control group."

[Do Users Write More Insecure Code with AI Assistants?](https://arxiv.org/abs/2211.03622?

Taggart (@mttaggart) writes:

Japan determines copyright doesn't apply to LLM/ML training data.

On a global scale, Japan’s move adds a twist to the regulation debate. Current discussions have focused on a “rogue nation” scenario where a less developed country might disregard a global framework to gain an advantage. But with Japan, we see a different dynamic. The world’s third-largest economy is saying it won’t hinder AI research and development. Plus, it’s prepared to leverage this new technology to compete directly with the West.

I am going to live in the sea.

www.biia.com/japan-goes-all-in-copyright-doesnt-apply-to-ai-training/

2023-12-29 by Steve C:

Rather than using R/python/SQL to analyse some data, the author of this post appears to have given a CSV and prompts (in plain English language) to OpenAI assistant, and it generated quite advanced plots. I find this fascinating as some of those plots would take some time in ggplot or matplitlib, especially finessing them to get the spacing/centering right.

https://incident.io/blog/festive-macbooks

Posted on rfordatascience.slack

cross-posted from: https://programming.dev/post/8025145

2023-12-29 by Novica Nakov:

I notice that Ctrl+Enter for running the code in Python and in R is not the same thing.

Read the whole article

May 2, 2023 by Vojtěch Juránek

excellent ML frameworks like TensorFlow, PyTorch or general data processing frameworks like Spark became available and made the writing of ML models much more straightforward. Since that time, these frameworks have matured, and writing models are even more accessible, as you will see later in this blog. However, data set preparation and gathering data from various sources can sometimes take time and effort. Creating a complete pipeline that would pull existing or newly created data, adjust it, and ingest it into selected ML libraries can be challenging. Let’s investigate if Debezium can help with this task and explore how we can leverage Debezium’s capabilities to make it easier.

Read Image classification with Debezium and TensorFlow

December 28, 2023 by Amy Ma

diving headfirst into the world of loss functions: the silent superheroes guiding our models to learn from mistakes. In this post, we'd cover the following topics:

• What is a loss function?
• Difference between loss functions and metrics
• Explaining MSE and MAE from two perspectives
• Three basic ideas when designing loss functions
• Using those three basic ideas to interpret MSE, log loss, and cross-entropy loss
• Connection between log loss and cross-entropy loss
• How to handle multiple loss functions (objectives) in practice
• Difference between MSE and RMSE

Read An In-Depth Guide to the Most Common Loss Functions

abstract :

How do sequence models represent their decision-making process? Prior work suggests that Othello-playing neural network learned nonlinear models of the board state (Li et al., 2023). In this work, we provide evidence of a closely related linear representation of the board. In particular, we show that probing for "my colour" vs. "opponent's colour" may be a simple yet powerful way to interpret the model's internal state. This precise understanding of the internal representations allows us to control the model's behaviour with simple vector arithmetic. Linear representations enable significant interpretability progress, which we demonstrate with further exploration of how the world model is computed.