Vaclav Kosar's face photo

Vaclav Kosar

Welcome back to my blog.

Vaclav Kosar's Software & Machine Learning Blog

Learn about machine learning, self-attention in transformer architecture, software engineering.

I reduce costs and increase revenues with software and machine learning systems. For ScaleVoice I am developing on applications of LLMs in customer acquisition. For GLAMI, I optimized a product feature detection system performing 40 million predictions (e.g. product category) each month, so that the customers can find what they are looking for quickly. I also publish resources which you for machine learning research and education. For example, with GLAMI we recently published a GLAMI-1M dataset together with a practise notebook combining of texts which are in multiple languages and images.

Find more about me, Vaclav Kosar, over here.
Read the latest posts below.
Sort by category instead.
I also write what I learned about business, investing, and health.

Created on 30 Oct 2017. Updated on: 27 Dec 2022.

Vaclav Kosar's face photo

Vaclav Kosar

Let's connect! I may unlock opportunities or help you climb over obstacles.

Latest posts

Vaclav Kosar's face photo

Vaclav Kosar

Explore this area for additional insights to learn and apply tomorrow.

Precision and Recall Intuitive Mnemonic

Remember the precision and recall definitions with this visual story.

OpenAI Tool Call Messages Full Example

Complete example of conversation with user, function call with arguments, function response messages.

Multilingual Open LLMs

A list of good open multilingual large language models.

Bash Tricks I Use

Error handling, star expansion, exporting variables from an env file.

Llama 3 Performance and Deployment

Evaluations, Quantization, Fine-tunings

How I Cook One Pot High Protein High Fiber Low Histamine Food

Simple, cheap, fast, and relatively healthy food I cook in 40 minutes.

Synthetic Data for LLM Training

How I think about using generated training data for large language model training.

Bellman Update and Synthetic Data in Q-Transformer

Notes on Q-learning, temporal difference, Monte Carlo, and others methods related to Q-Transformer.

How to Install Your Python Version on Ubuntu

No need for Deadsnakes nor Apt, use pyenv or conda instead!

Dangers of Python Lambda: Repeated Values due to Late Binding

Avoid hidden bugs due to unexpected values in loops with AsyncIO or Multi-threading Python.

J.D. Rockefeller's Quotes, Gratitude, and Fictional Letters

Push away fictional letters, and discovered family letters full of gratitude.

Encoder-Only vs Decoder-Only vs Encoder-Decoder Transformer

Wrap your head around the main Transformer variants in 5 minutes.

Create Your Google Calendar Event Link in Seconds

Fill the form or upload your ICS file to share events with Your participants without having Google Calendar yourself.

How Deep Neural Networks Learn

Notes on Superposition, Memorization, Regularization, Double Descent, Model Ensembling in neural network training.

Multiple Interval Averaging Stopwatch

Measure multiple time intervals and get average calculated and copy data from a table.

Validate Reliability of a Research Paper

Save time on diligence using 5 questions on authority, bias, evidence, comparison, impact based on Doomberg's framework, common Ray Dalio's believability, TRAAP Test, CRAAP Test, RADAR.

Feature-wise Linear Modulation Layer

Guide models with FiLM layers via a simple linear transformation conditioning.

How to Create a Machine Learning Dataset

Trade-offs and time investment in publishing a dataset for reproducible research and advertising.

OpenAI's Image-Text Model CLIP

Encode image, and text into similar embedding vectors for multimodality.

Forward Deployed Engineer

Technical Account Manager, Solutions Architect, Professional Services Consultant.

Prompting Techniques That Squeeze the Best Out of Your LLM

From the simplest to the most advanced, instruct your GPT for the best generation.

Applying David Deutsch's Principles in Life and Business

Exploring principles of Fallibilism, Optimism, Explanations, Conjecture and Refutation, and Evolution in daily work.

Symbolic vs Connectionist Machine Learning

Reason without hallucinations in large language models by hybridizing neural networks with code in neural-symbolic systems.

Fill Versions From Python Environment

Script that fills in versions into your requirements file

Cooking with Psyllium Fiber to Reduce Sugar Spikes

My month-long experience using this plant-based fiber ingredient.

Neural Network Pruning Explained

Reduce on-CPU prediction and model storage costs by zeroing-out weights while minimally increasing the loss.

Tokenization in Machine Learning Explained

Tokenization is splitting the input data into a sequence of meaningful parts e.g. pice data like a word, image patch, document sentence.

Embeddings in Machine Learning Explained

Embedding is a task specific lower dimensional vector representation of data like a word, image, document, or an user.

Multimodal Image-text Classification

Understand the top deep learning image and text classification models CMA-CLIP, CLIP, CoCa, and MMBT used in e-commerce.

Strong Static Typing vs Weak Dynamic Typing

Master programmers Carmack, Blow, Fowler, and Uncle Bob on strong static typing in contrast to weak dynamic typing.

Python functools.cmp_to_key Explained

Understand the functools' comparison function to key function conversion quickly.

Transformer Positional Embeddings and Encodings

How transformers encode information about token positions?

Transformer Embeddings and Tokenization

How transformers convert text and other data to vectors and back using tokenization, positional encoding, embedding layers.

Bits-Per-Byte and Bits-Per-Character

BPB and BPC are metrics used in compression and language modelling related to compression ratio.

Neural Data Compression

Lossless bit reduction with machine learning by minimizing cross-entropy. Examples: NNCP and TRACE models.

Understand Large Language Models like ChatGPT

In 9 slides from TF-IDF, Word2vec, knowledge graphs, and transformers to LLMs and ChatGPT basics explained.

OpenAI's DALL-E 2 and DALL-E 1 Explained

Compare of text-to-image generation models DALL-E 1, 2, and understand related models VQ-VAE, CLIP, and GLIDE

Google's Pathways Language Model and Chain-of-Thought Prompting

PaLM, the largest language model of early 2022, outperformed average human on grade school logic and math (BIG-bench) by simulating reasoning steps.

Word Alignment for Sentence Similarity

Semantic similarity increases with similar semantic units of similar semantic contexts in the monolingual word alignment.

Sparse Matrix Why and When?

Sparse representations and matrix formats like CSR, LOL, COO compress and speed up certain operations on mostly zero matrices.

MassiveText Dataset introduced for pre-training of DeepMind's Gopher

Private diverse 10-lingual textual dataset composed of web, Github, news, Wikipedia, Books, C4.

Transformer's Self-Attention Mechanism Simplified

How transformer models like BERT and GPT work?

SRU++ Model Speeds Up Transformer with Simple Recurrent Unit

Reducing compute by combining RNN with self-attention from Transformer architecture.

Twitter Bullet Points to Copy & Paste

Copy-paste your bullet point symbol and other tweeting tips and emojis. Usable also for LinkedIn, Meta, Inkscape ...

7 Powers' Moats Through the Lens of DiBello's Business Mental Model

Protect margins with Helmer's moats by understanding dimensions of demand, supply, and capital.

DeepMind's RETRO Retrieval-Enhanced Transformer

Retrieval-Enhanced Language Model cross-attends trillions of tokens for SoTA on Wikitext103 and The Pile with 25x fewer parameters.

Cross-Attention in Transformer Architecture

Merge two embedding sequences regardless of modality, e.g., image with text in Stable Diffusion U-Net with encoder-decoder attention.

Ten Commandments for Business Failure Book Summary

On inflexibility, infallibility, fouls, and more from D. R. Keough with foreword by Warren Buffett.

Manipulate Item Attributes via Disentangled Representation

Using attribute-specific embedding subspaces for image manipulation retrieval, outfit completion, conditional similarity retrieval.

ELECTRA - How to Train BERT 4x Cheaper

Reducing training flops 4x by GAN-like discriminative task compared to RoBERTa-500K transformer model.

Expire-Span: Scaling Transformer by Forgetting

Reducing computational costs by differentiably dropping memorized embeddings from self-attention context.

Scout Mindset Book Summary

Galef on clear thinking by avoiding tribal soldier mindset, self-awareness, no-illusions, changing mind, criticism.

Quilt Data Versioning Review & How-to

How to version data using Quilt data for Python on AWS S3 for machine learning.

Wav2vec: Semi-supervised and Unsupervised Speech Recognition

Word2vec for audio quantizes phonemes, transforms, GAN trains on text and audio from Facebook AI.

PID Controller: A Simple Control Loop Mechanism

Proportional–integral–derivative controller calculates feedback to reduce the error in the next step.

DreamCoder: Wake & Sleep Program Learning

Learning to code by growing function library, fantasising coding tasks, and training neural search.

Google Product Taxonomy Viewer

Interactively explore Google Shopping's and Shopify's categories to configure products in your feed.

Automatically Expanding Taxonomy

Pinterest's Arborist model finds parents for unseen textual nodes using triplet-loss, StarSpace embeddings, & shortest path.

Submodularity in Ranking, Summarization, and Self-attention

Diminishing returns with a budget constraint in problems of coverage and results diversification.

Feed-Forward, Self-Attention & Key-Value

Feed-forward (MLP) layer is similar to cross-attention as observed in SwiGLU and All-attention.

Lambda Networks Transform Self-Attention

Is Lambda Layer similar to self-attention in a Transformer? What gives LambdaNet its power? LambdaResNet beats EfficientNet but does it loose to Performer?

Performers FAVOR+ Faster Transformer Attention

The Performer model attention approximation has linear complexity in contrast to square and outperforms Linformer.

Double Descent Contrary to Bias-Variance Trade-Off

Increasing parameter count leads to multiple test loss peaks and a global minima in the overparameterized regime.

To What Python Number Types Does json.loads Parse?

JSON specifies only a number value, so how to infer the correct type between int and float? How are NaN and Infinity handled?

Brutalist and Modernist Architectures Collide at Sunshine Plaza

Take a tour of the surprising merge of a modern and soviet era design on a Prague's public square with this photo album.

VR Holiday In Nature: Relax In Nature While Trapped In Your City

Escape the cityscape into the wild through your VR headset. Subscribe and download VR180 3D holiday.

Word Mover's Embedding: Cheap WMD For Documents

What is Word Mover's Embedding for documents and how it approximates Word Mover's Distance between documents.

Transfigure Stress into Energy by Drawing on Research

Apply research and Bezos' advice and your pounding heart will flash opportunity instead of doom.

OpenAI's Glow - Flow-Based Model Teardown

Interpretable latent representations by composing non-linear invertible functions and maximizing the exact log-likelihood.

BentoML vs Cortex - ML Serving Showdown

To find the best model serving tool, compare open-source MLOps platforms BentoML and Cortex.

StarSpace - Embeddings For Documents, Users, and Words

Create vectors of various entities in a single space with this general-purpose embedding model from Facebook AI.

When is the end of this quarter?

How many days are left in the quarter? What is the current calendar quarter-end date? What is the current quarter?

Thinkpad P52 Disassembly For Repaste, RAM Upgrade, Or Anything Else

Repaste, max RAM, or install antenna into Thinkpad P52 with these links and tips for full disassembly.

Python Context Manager Exception Handling and Retrying

Wrap your resource into a context manager with-statement to catch, handle exceptions, and close the resource.

Result Diversification in Web Search and Recommenders

Increase coverage in web search and recommendation via re-ranking diversification factor

I read papers on a podcast

To improve my pronunciation and speech, I read mostly Machine Learning scientific papers on a podcast.

Learn faster with a generated quiz

Reduce your effort of creating and revising learning material using a free AI-powered tool.

Thinkpad P53 vs P52 Thermals: Any Improvement?

Is Thinkpad P53 the cooler brother of P52?

Constant 1D Kalman Filter Is Exponential Or Cumulative Average

Understand the Kalman filter with a simple one-dimensional example. It converges to cumulative or exponential average in case of constant measurement uncertainty and process noise.

FastText Word Embeddings

How FastText works, word embeddings, n-grams, OOV words, and visualize embedding norms.

Highly Compressed Richard Hamming's Lectures

Get inspired by Hamming's lectures compressed into tiny downloadable files.

Thinkpad P52 vs ZBook 15 G5 vs Precision 7530

My experience with the best mobile workstation of 2019 with specs matching Thinkpad P52 and ZBook 15 G5 and P1000 GPU.

Spline: Data Lineage For Spark Structure Streaming (2018)

Vaclav Kosar and Marek Novotny presentation at Spark N AI Summit 2018 of a POC of Structured Streaming data lineage tool.

Debounce In Bash To Fix Lenovo Touchpad And Trackpoint Lost Sync

Functional programming in Bash for Yoga X260 and P52.

My First Contribution To A Major OSS Project Apache Spark

Finally my rather small pull request was merged into master of Apache Spark!

How To Create Custom Ubuntu Web Link App

Turn any web page into an Ubuntu application and prevent the web owner from tracking you around the web.

Modern Config Injection In Maven Plugins

Maven Mojo constructor injection of config parameters via Guice JSR-330 support.

Walking Desk: Cheap And Tiny

A review of my motor-less walking desk setup.

Boundary Control Entity Architecture Pattern

BCE is a source code structure pattern sometimes called ECB, EBC, Hexagonal, Onion, Clean architecture, or MVC.

Spring Integration Highlights - message driven architecture

Get familiar with Spring Integration implementation of Enterprise Integration Patterns and compare it to Java 8 Streams and RxJS.

Fish Roe vs Fish Oil

Healthiness and price of a salty delicacy versus oily softgels. EPA, DHA, Neu5Gc.

Restore Missing Punctuation with Keras Convolutional Text Punctuator

Simple neural network android app for restoring punctuation in text e.g. YouTube subtitles.

Easy Online Independence: Mail Backup, File Synchronization

Cheap way to increase your independence from the online giants with Syncthing, mbsync, Thunderbird.

Generic Class Name Signals Low Cohesion

Why and how to avoid non-specific class names like util, utils, or helper?

How to Structure Code

Localize Related, Inline over Extract, Specific over Generic. My view partially based on Carmack, Jonathan Blow, and Adam Bien's posts.

Is $15 USB Microscope Enough For You?

See yeast cells and pond water critters paying just $15 for an USB microscope.

Linux Text To Speech Comparison: Flite Vs Pico2Wave Vs Festival

Comparison of open-source text to speech (TTS) software in terms of pleasantness, comprehensibility, and modularity.

Try This Sped-Up Classical Music Attuned For Today's Sped-Up Age

Are you attracted to complexity and nobility of classical music, but deterred by its slow pace?

Drone Detecting White Marker for Stabilization

On a hackathon I implemented trivial image processing algorithm to locate white piece of paper on grey carpet floor to be used for drone horizontal stabilization.

Functional ForEach In Bash

Don't you hate verbosity of Bash's while-do statements when writing in-line scripts? No worries, you can improve on that!

Obsolete Git Branch Remover Maven Plugin

Having many branches left behind, abandoned, never deleted? How do you deal with them? We had same problem and I developed a automated DevOps solution.

How To Boost Your Jog Morale Using Military Cadence And Run Farther

Run beyond your max with this professional mind hack, and burn calories.

GitFlow Incremental Builder - Speed up your multi-module Maven build

Build only modules that changed compared to the master branch, plus their dependents. Java!

Hamiltonians with constant spectral intervals and time-dependent perturbation

On quantum systems determined by time-dependent Hamilton operators. Family of quantum systems, whose Hamilton operators take form H(t) = H 0 + V (t), where V (t) is perturbation and H 0 is self-adjoint with pure-point spectrum and constant gaps between eigenvalues in spectrum σ(H 0 ).

Feynman summation in finite-dimensional quantum mechanics

A summary and enhancement of existing literature regarding finite-dimensional quantum mechanics. In the later parts Feynman’s path summation is discussed.

Simulation of Soft Photon Calorimeter

Understand how an electromagnetic calorimeter works presented at Dubna JINR 2011.

Transverse momentum spectra and correlations in the blast wave model with resonances

This work provides a review of theories of properties of high energy density matter originating in heavy-ion high energy collisions (GeV/nucleus).

Programing Ateus Omega Over UDP and TCP/IP

Replacing Ateus Omega remote programming Xapi PC software with a built-in DsPIC module with ENC28J60.

About Vaclav Kosar How many days left in this quarter? Twitter Bullet Points to Copy & Paste Averaging Stopwatch Privacy Policy

Copyright © Vaclav Kosar. All rights reserved. Not investment, financial, medical, or any other advice. No guarantee of information accuracy.

Vaclav Kosar's face photo

Vaclav Kosar

You made it all the way here. Many doors opened. Where do we go?