Monitoring Models at Scale

Catch this Domino webinar on monitoring models at scale, Dec 11 @ 10am PT, covering detecting changes in pattern of real-world data your models are seeing in production, tracking how model accuracy and other quality metrics are changing over time, and getting alerted when health checks fail so that resolution workflows can be triggered.

New R Support in Azure Machine Learning

A new R package azuremlsdk (available to install from Github now, and from CRAN soon), provides the interface to the Azure Machine Learning service. With R functions, you can provision new computing clusters in Azure, and use those to train models with R and deploy them as prediction endpoints for use from any app. You can also launch R-based notebooks in the new Azure Machine Learning studio web interface, or even launch a complete RStudio server instance on your cloud computing resources. Azure Machine Learning service supports the latest version of R (3.6.1) and all R packages (from CRAN, Github, or elsewhere).

Design of Experiments for Your Change Management

A step-by-step Guide to Design of Experiments. Data science professionals, have you ever faced any of the following challenges?
Story 1: Machine learning does not mean experimental design
You are asked to design an experiment due to your statistical expertise, but realized your machine learning tools do not help you design an experiment.
Story 2: Observational data cannot be repeated
Your team observed the sales trend relating to a few factors. Your team wanted to propose new business strategies. But the team cannot decide the right factors to replicate the process.
Story 3: Correlation is not causation
Your team identified a strong correlation between two factors and proposed a business change plan. Your team got the buy-in of the clients to implement the business proposal in a promising pilot test. After the testing period, you got no claimed outcome.
If so, you may have committed the fallacy of using correlation to prove causation. In my post ‘Machine Learning or Econometrics?’ I shared that many analytic questions are about causality. These causal questions require the designs of data generation and cannot be derived from big data. Today the divergence of machine learning and statistical experiments seems converging into the big data science umbrella. However, the two distinct disciplines can and should cross-pollinate to solve numerous business initiatives.

Why your AI might be racist and what to do about it

Even well-designed AI systems can still end up with a bias. This bias can cause the AI to exhibit racism, sexism, or other types of discrimination. Entirely by accident. This is usually considered a political problem, and ignored by scientists. The result is that only non-technical people write about the topic. These people often propose policy recommendation to increase diversity among AI researchers. The irony is staggering: A black AI researcher is not going to build an AI any different from a white AI researcher. That makes these policy recommendations racist themselves. It still makes sense to increase diversity among AI researcher for other reasons, but it certainly won’t help to make AI system less racist.

The Akaike Information Criterion

In this article, we’ll cover the following topics:
• We’ll learn about the AIC, its definition, its properties, and its uses.
• We’ll learn about the concepts that the AIC formula is based upon.
• We’ll run an experiment using Python and statsmodels in which we’ll use the AIC score to select an optimal regression model from a collection of over 4000 candidate models.

An Alternative To Batch Normalization

The development of Batch Normalization(BN) as a normalization technique was a turning point in the development of deep learning models, it enabled various networks to train and converge. Despite its great success, BN exhibits drawbacks that are caused by its distinct behavior of normalizing along the batch dimension. One of the major disadvantages of BN is that it requires sufficiently large batch sizes to generate good results(for-eg 32,64). This prohibits people from exploring higher-capacity models that would be limited by memory. To solve this problem Facebook AI Research(FAIR) developed a new normalization technique, Group Normalization(GN). In this article, we will be mainly focussing on Group Normalization(GN)and how it can be used as an alternative to Batch Normalization(BN) and other normalization variants(Layer Normalization(LN), Instance Normalization(IN)).

How to Manage Your Machine Learning Workflow with DVC, Weights & Biases, and Docker

Managing a machine learning workflow is hard. Beyond the usual challenges in software engineering, machine learning engineers also need to think about experiment tracking, reproducibility, model deployment, and governance. In this article, I want to show 3 powerful tools to simplify and scale up machine learning development within an organization by making it easy to track, reproduce, manage, and deploy models.

Let’s build an Intelligent chatbot

Step by step approach to build an intelligent chatbot using python. In the article Build your first chatbot using Python NLTK we wrote a simple python code and built a chatbot. The questions and answers were loosely hardcoded which means the chatbot cannot give satisfactory answers for the questions which are not present in your code. So our chatbot is considered not an intelligent bot. Here in this article, we will build a document or information-based chatbot that will dive deep into your query and based on that it’s going to respond.

Transfer Learning in NLP

Welcome to the first chapter of Modern NLP. For obvious reasons, it makes sense to start with the story of transfer learning – the reason for rapid progress in NLP. This chapter was originally written for my book ‘NLP with BERT’ with Packt publication but I found it better to rather publish it openly on medium. Natural language processing is a powerful tool, but in real-world we often come across tasks which suffer from data deficit and poor model generalisation. Transfer learning solved this problem by allowing us to take a pre-trained model of a task and use it for others. Today, transfer learning is at the heart of language models like Embeddings from Language Models (ELMo) and Bidirectional Encoder Representations from Transformers (BERT) – which can be used for any downstream task. In this chapter, we will understand different types of transfer learning techniques and how they can be used to transfer knowledge to a different task, language or domain.