Deep Brain AI

"Neural networks are nature's way of implementing optimization algorithms. It's nothing but maximizing information gain in terms of compressing what's similar and contrasting what's different, while minimizing resources, say (or coding length). Hence a game." --- Yi Ma, UC Berkeley professor

In any data science project or research, don't get over exited about tech or jump into coding part too early, before understanding the problem and requirements well.

Do the following (your steps maybe different depending on requirements):

In research:

-- Understand the problem well.
-- Formulate research questions and hypotheses
-- Get the data ready, do necessary ETL, feature engineering, etc.
-- Implement the ideas to solve the problems
-- Experiment to validate hypotheses
-- Disseminate your findings (conference/journals/workshops).

In business:

-- Understand the business context very well and requirements (in a collaborative setting with client) thoroughly
-- Analyse customers requirements and communicate them to cross-check
-- Get the data ready, do necessary ETL, feature engineering, etc.
-- Dive into coding or implementation, predictive modelling, tuning, evaluation
-- Build the first prototype, send it to customers, and ask for their feedback
-- Address their comments and improve the prototype
-- Get feedback on the improved version
-- Iterate the process until they're happy
-- Communicate the models and business insights in the form of customer-centric story telling.
-- Take the prototype into production, scale up, deploy/deliver.

If you're a data scientist or a data engineer using Spark, you might be interested in SynapseML - a library for doing machine learning at scale with massive datasets.

SynapseML has several interesting features:

1. Responsible Al module, which gives detailed explanations of how features in opaque-box models affect the model prediction.
2. Support for Cognitive Services features like text-to-speech, text analytics, and multivariate anomaly detection.
3. Improved MLOps features with support of MLFlow
4. Support for geospatial intelligence to analyze distributed data on maps.
5. Support for tree ensemble model like LightGBM.

Release notes: https://Inkd.in/dukhsPg8
Paper: https://lInkd.in/dB3_r4rh

A distance measure is an objective score that outlines the relative difference between two objects (e.g., data points) in a problem domain.

In instance-based learning, distance measures play an important role. Usually training examples are stored verbatim and a distance function is used to determine which data instance of the training set is closest to an unknown test instance. Once the nearest training instance is located, its class is predicted for the test instance.

Following are a few algorithms in which distance measures at their core:

-- K-Nearest Neighbors
-- Learning Vector Quantization (LVQ)
-- Self-Organizing Map (SOM)
-- K-Means Clustering
-- K-medoids
-- Kernel-based support vector machines (SVM)
-- Knowledge graph embedding algorithm like TransE, etc.

To whom it may concern!

Don't expect only a fancy and flamboyant visualization (out of sorted or ordered data) to hit the mark, rather use them meaningfully to create a story that is understandable, interesting, memorable, and reproducible!

John Mccarthy, a professor emeritus of computer science at Stanford University, coined the term "artificial intelligence" and subsequently went on to define the field for more than five decades!

"Interviewers should not give a Data Scientist candidate a coding interview to figure out if they can really code. Instead a few questions could help them assess the candidate's coding skill even during a phone screen:

-- How do you structure your projects and code?

-- What are your coding best practice strengths and what are the worst practices you're working on?

-- If you run into a problem and there's no answer on Stack, how do you continue to research it?

-- What is your debugging process and what's the ugliest issue you've had to run down?

While the first 2 are about preventing things from going wrong, the last 2 are about todos when something does go wrong.

Answers may vary from candidates to candidates: strong coders will give substantive answers to all 4 questions, weak coders will give canned, high level answers.

Ideally, the assessment of the hiring manager should be focused on process, not syntax. He/she should focus on candidate's depth of experience, not length of experience. They should try to understand the foundational concepts & their implementations, not how much they have memorized" - Vin Vashishta

You can use LabML to organize machine learning experiments and monitor training progress and hardware usage from your phone. The current version supports PyTorch and Keras.

GitHub: https://github.com/lab-ml/labml

lab-ml/labml 🔎 Monitor deep learning model training and hardware usage from your mobile phone 📱 - lab-ml/labml

The following timeline of XAI covers most important algorithms and techniques wrt their scopes (i.e., global, local), methodology (e.g., backprop, perturbation), and usage level (e.g., intrinsic, post-hoc).

In case you see something missing, please comment.

Very insightful information on dimensionality reduction from Kyle McKiou.