Writing

15 min

OlmoEarth v1.1 from Ai2: a small, runnable embedding + kNN demo

earth-observation

foundation-models

remote-sensing

olmoearth

sentinel-2

ai2

Ai2 released OlmoEarth v1.1 on May 19. This post unpacks what changed and runs a tiny embedding + k-NN demo on synthetic water / forest / urban Sentinel-2 chips using the 3.5 M-param Nano model. Runs in under 5 seconds on a laptop.

May 22, 2026

Can Gemini Generate Believable Thermal Faces? Four Prompting Strategies, Two Model Sizes, vs. a Real Thermal Photograph

Gemini

image-generation

thermal-imaging

evaluation

paper-prototype

Nano-Banana

I ran Gemini 3.1 Flash Image Preview (Nano Banana 2) AND Gemini 3 Pro Image Preview (Nano Banana Pro) through four prompting strategies for generating LWIR thermal face images, then asked Gemini-3.5 to judge each against a real thermal photograph. All 14 outputs (7 per Gemini model) make the same physical errors. The bigger Pro averages +0.9 judge points but ceiling stays at 7/10 — pure scaling doesn’t fix the modality gap. The literature priors (T-FAKE CVPR 2025, ThermalGen NeurIPS 2025, ThermVision-DB 2026) all use finetuned generators with physics-aware losses. So I ran ThermalGen’s pretrained RGB→thermal model on the same image — and it produces visibly more correct thermal physics (canthi-correct eyes, cool hair, no glowing pupils) than any Gemini output. Surprising kicker: the Gemini-3.5 judge scores ThermalGen LOWER (3-4/10) than Gemini’s own outputs (7/10), because the judge is biased toward Gemini-style ‘fancy-thermal-looking’ images. Don’t use LLMs to judge thermal physics.

May 21, 2026

42 min

Thermal Super-Resolution: Can Off-the-Shelf RGB SR Models Upscale Thermal Faces? (Five Methods Head-to-Head)

thermal-imaging

super-resolution

ESRGAN

EDSR

DRLN

MSRN

A2N

Stable-Diffusion

image-restoration

paper-prototype

Six super-resolution methods tested on a 4× thermal upscaling task (116×87 → 464×348): bicubic, EDSR, MSRN, A2N, DRLN (all RGB-trained CNNs), and Stable Diffusion x4 upscaler (diffusion). Classical CNN-based SR cleanly beats both bicubic and the diffusion upscaler on every metric and on downstream nostril localisation. DRLN tops PSNR/LPIPS; A2N achieves 0.7 px nostril error on a 256-px-wide face. The diffusion upscaler is the most visually sharp but the LEAST accurate downstream — it hallucinates plausible-looking detail that breaks pixel alignment, dropping nose localisation accuracy 7× vs bicubic.

May 21, 2026

9 min

Finetuning Sapiens2 for Nostril Detection on ThermEval-D (with a Tiny Annotated Set)

thermal-imaging

keypoint-detection

Sapiens2

fine-tuning

few-shot

transfer-learning

frozen-backbone

ThermEval

Sapiens2-0.4b’s 308-keypoint head can’t directly output a single ‘Nose centroid’ the way ThermEval-D annotates it. So I freeze the backbone, replace the head with a 410k-parameter 1-keypoint heatmap regressor, and train on 40 ThermEval crops. Test PCK@10 = 93%, mean error = 5.5 px on the 80-crop test split. Does not beat zero-shot DWPose (99% PCK@10, 2.7 px) on this dataset — but produces a predictor pinned to YOUR anatomical convention, distillable to a tiny inference model, and reproducible on data DWPose wasn’t trained for. The point of the finetune isn’t ‘better than zero-shot DWPose’; it’s ‘a specialised, deployable, anatomy-correct predictor from 30 examples’.

Train a YOLO for Nostril Detection on Thermal — Hierarchical Pipelines That Actually Work

computer-vision

keypoint-detection

YOLO

MediaPipe

BlazeFace

hierarchical-detection

two-stage

thermal-imaging

ThermEval

Instead of trying to make MediaPipe FaceMesh (RGB-trained) work on thermal crops, train a tiny YOLOv8n directly to detect ‘nostril’ as an object class on ThermEval-D thermal frames. With 60 training crops, the YOLO hits 88% detection rate and 1.8 px median accuracy on 60 held-out frames — beating the best MediaPipe pipeline (23%/7.5 px) by 4x on detection and 4x on accuracy. The right hierarchical pipeline for thermal isn’t ‘face detector + RGB-trained FaceMesh’; it’s ‘face crop + a dedicated thermal-trained nostril detector’.

12 min

Why Nostril Detection on Thermal is Fundamentally Harder than on RGB — and Four Ways Around It

thermal-imaging

keypoint-detection

computer-vision

anatomical-priors

synthetic-data

multimodal

breath-monitoring

ThermEval

RGB has skin texture; thermal does not. A small dark blob inside a face on thermal looks like every other small dark blob — anywhere on the face, anywhere in the room. This post unpacks why MediaPipe and hierarchical pipelines fail on the ThermEval-D bake-off (per parts 1 and 3) and proposes four research-grade fixes: anatomical priors as a regulariser, paired-RGB label projection, T-FAKE synthetic pretraining, and using the breath-rate signal itself as self-supervision.

9 min

A Bake-off for Nostril Detection on Real-World Thermal Scenes (ThermEval-D)

thermal-imaging

keypoint-detection

Sapiens2

DWPose

MediaPipe

ViTPose

pose-estimation

benchmark

breath-monitoring

ThermEval

Off-the-shelf wholebody pose / face mesh models on the ThermEval-D real-world thermal dataset (192x256, indoor multi-person scenes). On the smallest-face thermal regime that actually matches real deployment, DWPose wins decisively: 100% detection rate, 99% PCK@10px. Sapiens2-0.4b is accurate (2 px median on the noses it finds) but its single-person inference misses 32% of multi-person scenes. MediaPipe FaceMesh detects only 14% of GT noses because the 20-25 px faces in ThermEval are below BlazeFace short-range’s working scale. The pipeline lesson is that for thermal deployment you need a model that includes its own multi-person face detector — not just a strong landmark head.

10 min

Gemini 3.5 Flash and Gemini Omni: What’s New, What’s Real, What I Could Measure

LLM

Gemini

benchmarks

latency

thinking

multimodal

Google announced Gemini 3.5 Flash and Gemini Omni at I/O 2026. I run 3.5 Flash against my April 3.1 baseline on the same Gemini API: thinking levels, tokens-per-second, a 500K-token needle test, agentic tool use, and a coding micro-benchmark. Omni is not yet on the developer API, so the second half is notes from the announcement rather than measurements.

12 min

From Vega-Lite to Matplotlib: A Tiny Spec-to-Code Translator

visualization

vega-lite

matplotlib

python

code-generation

dataviz

compilers

A small translator that lowers Vega-Lite JSON specs to readable Matplotlib code. v2 reworks it as a visitor class and adds aggregations, log scales, size encoding, area marks, layered charts, faceting, and filter transforms.

May 7, 2026

RS-CLIP on BharatEO: Tiny Contrastive Image-Text Model on WorldCover

earth-observation

foundation-models

clip

contrastive

remote-sensing

bharateo

worldcover

pytorch

Companion to the BharatEO MAE post: a pedagogical RS-CLIP-style model trained on the same ESA WorldCover ROIs. Tiny image and text encoders, InfoNCE loss, zero-shot land cover classification, image-text retrieval.

May 6, 2026

Building a Foundational EO Model for BharatEO: A Tiny WorldCover V0 Notebook

earth-observation

foundation-models

remote-sensing

bharateo

worldcover

pytorch

A pedagogical BharatEO-v0 experiment on real ESA WorldCover 2021 v200 pixels: 3-channel RGB MAE with L1 reconstruction, ESA patch classification, multi-ROI sampling, more chips, more epochs.

May 5, 2026

7 min

Fixing Slow GitHub Pages on a Campus Network with Split DNS

dns

github-pages

networking

macos

performance

GitHub Pages was painfully slow on a campus network even without Tailscale. The root cause was partial reachability to GitHub Pages IPs, and the clean fix was split DNS with a local forwarder.

May 3, 2026

—title: “ReAct on a local model: Gemma 4 31B (4-bit MLX)”date: 2026-04-27code-fold: truecode-summary: “Show code”toc: truetoc-location: right—

ReAct on a local model: Gemma 4 31B (4-bit MLX)

Apr 27, 2026

Sapiens2 on a Mac: Posture Readouts, Gait Analysis on Video, and Tele-Derm ROIs

Sapiens2

computer-vision

pose-estimation

surface-normals

segmentation

gait-analysis

posture

tele-medicine

Apple-Silicon

MPS

human-centric

video

Two days after Meta’s Sapiens2 release, I run the 0.4B normal, segmentation, and 308-keypoint pose heads on an M2 Max via PyTorch MPS, then build six health-leaning downstream pipelines: joint-angle readout, body symmetry, tele-dermatology ROI extraction, foreground-only relighting, hair recolour, and a 12-frame gait analysis on video.

Apr 25, 2026

Kernel Fusion in MLX: What `mx.compile` and `mx.fast.*` Actually Do

MLX

Apple-Silicon

performance

kernel-fusion

training

inference

GPU

Metal

An educational walkthrough of MLX’s three kernel-fusion paths — graph compile, the mx.fast.* fused primitives, and custom Metal kernels — with measured speedups for training and inference on an M2 Max.

Apr 25, 2026

13 min

Gemini 3.1 Flash TTS: Audio Tags, English and Hindi in One Notebook

LLM

Gemini

TTS

audio

Hindi

multilingual

audio-tags

Google’s new Gemini 3.1 Flash TTS adds natural-language ‘audio tags’ like [excitement] or [like dracula] that steer voice style inline. I run it in English and Hindi and embed the results.

Apr 16, 2026

Brick Kilns from Above: Gemma 4 Calling Falcon Perception in a Loop

LLM

Falcon-Perception

Gemma-4

MLX

satellite

remote-sensing

brick-kilns

SentinelKilnDB

agent-loop

Hi-res ESRI imagery + Gemma 4 in the PR #926 agent loop on three brick-kiln classes. Twelve tiles, four Falcon queries each, three prompt variants — with per-tile mask grids and Gemma’s reasoning shown. 67% zero-shot, and the remaining 33% is instructive.

Apr 16, 2026

Grounded Visual Reasoning on a Mac: a Corridor and a Drone Pan

LLM

mlx-vlm

Gemma-4

Falcon-Perception

MLX

local-inference

grounded-reasoning

video

tracking

Apple-Silicon

Two worked examples — a still photo and a drone video — running Falcon Perception and Gemma 4 locally on an M2 Max. Counts you can audit, from an agent that never does arithmetic in its head.

Apr 16, 2026

Why VLMs Lie About Numbers (and How to Fix It With Grounding)

LLM

Gemma-4

Falcon-Perception

MLX

local-inference

grounded-reasoning

multimodal

Gemma 4 sounds confident. Falcon Perception actually measures. Together they beat either alone — here is the evidence.

Apr 5, 2026

Google released Gemma 4 yesterday — open-weight multimodal models under Apache 2.0. In a companion post, we compared Gemma 4 and Gemini via the API. Here, we run Gemma 4 enti…

Gemma 4 Locally on Mac: Performance & Vision Benchmarks

LLM

Gemma-4

MLX

local-inference

multimodal

vision

Apple-Silicon

benchmarks

Apr 3, 2026

Google just released Gemma 4 — open-weight models (Apache 2.0) available directly through the Gemini API. Same SDK, same API key — just change the model name.

Gemma 4 vs Gemini: Open-Weight Models Meet the Gemini API

LLM

Gemma-4

Gemini

multimodal

vision

segmentation

comparison

Apr 3, 2026

Paper: Koh et al., Concept Bottleneck Models (ICML 2020)

Concept Bottleneck Models — MNIST Demo

interpretability

deep-learning

pytorch

MNIST

educational

Mar 12, 2026

Agentic Browsing with agent-browser

automation

tools

A hands-on guide to automating web browsing for AI agents using agent-browser CLI

Jan 15, 2026

Building AI Tool Servers with FastMCP

python

mcp

llm

tools

Learn how to build MCP servers with FastMCP - enabling LLMs like Claude to interact with your custom tools, databases, and APIs.

Jan 13, 2026

Brave Browser Setup for Teaching & Presentations

productivity

teaching

privacy

browser

Configure Brave browser to hide personal browsing history during classroom demos and presentations

Jan 4, 2026

In this notebook, we explore how Vision-Language Models excel at Optical Character Recognition (OCR) and structured document understanding. We’ll focus on:

VLM for OCR and LaTeX Generation: From Images to Structured Text

VLM

OCR

LaTeX

document-understanding

math-OCR

table-extraction

Jan 4, 2026

In our VLM series, we’ve built models that can caption images, detect objects, answer questions, and handle multiple tasks. But can they reason about what they see?

VLM

chain-of-thought

reasoning

visual-reasoning

multimodal

deep-learning

Dec 30, 2025

Brick kilns are a significant source of air pollution in India, particularly in the Indo-Gangetic Plain. Monitoring these kilns for environmental compliance is challenging…

Brick Kiln Detection using SAM3 and Satellite Imagery

SAM3

geospatial

object-detection

satellite-imagery

segmentation

environmental-monitoring

Dec 29, 2025

In the multi-task VLM notebook, we built a model that handles caption, OD, and VQA - but it requires explicit task prompts:

Intelligent Task Routing for Multi-Task VLMs

VLM

task-routing

auto-detection

classifier

multimodal

deep-learning

educational

Dec 28, 2025

When to clear memory: - After training/evaluation before starting a new notebook - When switching between large models - If you get CUDA out of memory errors

Multi-Image Vision-Language Model: Compare, Reason, and Understand

VLM

multi-image

interleaved

multimodal

deep-learning

educational

Dec 27, 2025

When to clear memory: - After training/evaluation before starting a new notebook - When switching between large models - If you get CUDA out of memory errors

Multi-Task Vision-Language Model: Caption, Detect, and Answer

VLM

multi-task

instruction-tuning

multimodal

deep-learning

educational

Dec 26, 2025

In previous notebooks, we: 1. Built a minimal VLM from scratch and trained it for image captioning 2. Instruction fine-tuned it for object detection with JSON output

Instruction Fine-Tuning a VLM for Visual Question Answering

VLM

instruction-tuning

VQA

multimodal

deep-learning

educational

Dec 25, 2025

In the previous notebook, we built a minimal Vision-Language Model (VLM) from scratch and trained it on image captioning. The model learned to describe images in natural…

Instruction Fine-Tuning a VLM for Object Detection

VLM

instruction-tuning

object-detection

multimodal

deep-learning

educational

Dec 25, 2025

Vision-Language Models (VLMs) have revolutionized how AI systems understand and reason about images. Models like GPT-4V, LLaVA, and Gemini can describe images, answer…

Building a Minimal Vision-Language Model from Scratch

VLM

vision-language

multimodal

deep-learning

transformers

educational

Dec 24, 2025

Consider this realistic scenario:

Fine-Tuning FunctionGemma for Multi-Table Data Analysis

LLM

fine-tuning

function-calling

gemma

data-analysis

SQL

Dec 24, 2025

Meta’s SAM3 introduces text-based prompting for segmentation - just describe what you want to segment. This notebook shows:

SAM3 + Gemini: Intelligent Image Segmentation

SAM3

Gemini

segmentation

vision

multimodal

Dec 23, 2025

In the previous post, we explored SAM3’s image segmentation. Now let’s use SAM3 for video tracking - following a tennis ball across frames using just a text prompt.

SAM3 Video Tracking: Following a Tennis Ball

SAM3

video

tracking

segmentation

Dec 23, 2025

Gemini’s code execution with images reveals how AI thinks. Instead of just getting an answer, you see:

Gemini Code Execution: See How AI Thinks, Codes, and Analyzes Images

LLM

Gemini

code-execution

computer-vision

reasoning

Dec 19, 2025

Brick kilns are a significant source of air pollution in India, and monitoring their temporal patterns across multiple locations is crucial for environmental management.…

Multi-Location Temporal Brick Kiln Analysis Using Async Gemini 3 Pro API

LLM

Gemini-3-Pro

satellite-imagery

temporal-analysis

async

environmental-monitoring

Dec 17, 2025

We often need to process multiple images. The question is: What is the fastest way to do it?

Gemini API Performance: Sequential vs Async vs Single-Prompt

LLM

Gemini

Performance

Batching

Python

Dec 12, 2025

In a previous post, we explored Google’s Gemini 3 Pro multimodal capabilities. The model showed impressive performance on object detection, segmentation, and complex visual…

Testing Gemini 3 Pro: Are VLMs Still Blind and Biased?

LLM

Gemini-3-Pro

vision

VLM

benchmarks

evaluation

Dec 10, 2025

Multimodal AI models can understand and generate content across different types of data—text, images, video, audio, and documents—through a unified interface. This notebook…

Gemini-3-Pro: Mastering Multimodal AI Capabilities

LLM

Gemini-3-Pro

multimodal

vision

audio

video

object-detection

Dec 6, 2025

In this post, we compare different approaches for extracting structured information from text using two tasks:

Structured Outputs: Extraction vs Classification with LLMs

LLM

structured-generation

evaluation

Nov 26, 2025

Google’s Alpha Earth project represents a breakthrough in satellite imagery analysis by providing pre-trained deep learning embeddings of Earth’s surface. These embeddings…

Exploring Satellite Embeddings with Google’s Alpha Earth

satellite-imagery

deep-learning

earth-observation

google-earth-engine

Nov 26, 2025

Handling Missing Input Channels with Mask-Aware Neural Networks

machine learning

deep learning

missing data

robust models

Learn how to build neural networks that gracefully handle missing input channels by explicitly encoding missingness patterns

Oct 29, 2025

Vector Search with Text and Images

Understanding vector databases through text and image search

Oct 7, 2025

Using REST APIs in Google Sheets with Apps Script

Learn how to call REST APIs from Google Sheets using a custom Apps Script function.

Aug 24, 2025

DINOv2 vs CNN: Transfer Learning vs Training from Scratch

Deep Learning

Computer Vision

Comparing DINOv2 fine-tuning vs training CNN from scratch for binary image classification

Aug 17, 2025

Visualizing ERA5 Data

Visualizing ERA5 data grid coverage over India

Aug 13, 2025

Temperature Scaling in Softmax: Controlling Randomness in Language Models

deep learning

natural language processing

softmax

temperature scaling

text generation

Exploring how temperature scaling affects the randomness and diversity of language model outputs through mathematical analysis and interactive visualizations

Jul 9, 2025

Multi-Image Classification with VOC Dataset Visualization

computer vision

deep learning

multi-label classification

data visualization

Visualizing and filtering multi-image classification samples from the VOC dataset for improved training data quality

Jul 7, 2025

Logit Masking for Sleep Stage Classification

machine learning

deep learning

sleep

sequence modeling

Demonstrating how logit masking enforces valid transitions in sleep stage classification

Jul 4, 2025

Here are some useful Slack tips to improve your team communication and channel management.

Slack Tips for Better Team Management

productivity

tools

Jun 30, 2025

Linux Remote Server Setup with uv and CUDA

setup

linux

remote

development-environment

gpu

cuda

Setting up Python Environment on Linux Remote Servers with GPU Support

Jan 15, 2025

Faster Quarto Builds with Incremental Rendering and Caching

quarto

github-actions

caching

performance

Tired of waiting 5+ minutes for Quarto to rebuild all 145 posts when you only changed one? Here’s how to make GitHub Actions only rebuild what actually changed.

Jan 8, 2025

8 min

Parallel File Transfer Over SSH When You Have Thousands of Files

ssh

scp

performance

terminal

Transferring large projects with thousands of small files over SSH can be painfully slow. Here’s how we solved it with parallel transfers.

Jan 7, 2025

Keyboard shortcuts on Mac

setup

macos

shortcuts

productivity

keyboard-shortcuts

workflow

Keyboard shortcuts on mac

Jun 12, 2024

Object detection

computer-vision

object-detection

deep-learning

neural-networks

yolo

pytorch

from ultralytics import YOLO, checks, hub
import pandas as pd

Attention in Sequence to Sequence

attention-mechanism

sequence-to-sequence

deep-learning

neural-networks

transformers

nlp

pytorch

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset
import numpy as np
import matplotlib.pyplot as plt

Time Series Forecasting - Comparing MLP and Transformer Models

time-series

forecasting

deep-learning

transformers

pytorch

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
import torch
import torch.nn as nn
import torch.nn…

Character-Level Attention Mechanism for Name Generation

1165 rows × 1 columns

Character-Level RNN for Name Generation

1165 rows × 1 columns

Sampling Distribution Analysis - Mean vs Standard Deviation

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
import torch
import torch.nn as nn
import torch.nn…

May 7, 2024

Mixture of Experts

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
import torch
import torch.nn as nn
import torch.nn…

Jan 18, 2024

Logo

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
%config InlineBackend.figure_format = 'retina'

Jan 2, 2024

Basic Neural Process

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import torch
import torch.nn as nn
import torch.nn.functional as F
%config…

Dec 28, 2023

Using a neural network as a covariance function

import torch
import torch.nn as nn
import matplotlib.pyplot as plt
import numpy as np
# Retina mode
%config InlineBackend.figure_format = 'retina'

# Create…

Dec 22, 2023

Towards Transformers

transformers

attention-mechanism

deep-learning

neural-networks

nlp

sequence-modeling

pytorch

import tiktoken

encoding = tiktoken.get_encoding("cl100k_base")

encoding.encode("Hello World! This is a simple notebook")

[9906, 4435, 0, 1115…

Dec 21, 2023

Naive implementation of Strassen’s algorithm

import numpy as np 
import time
import matplotlib.pyplot as plt
import pandas as pd
# Retina display
%config InlineBackend.figure_format = 'retina'

log_size …

Dec 20, 2023

YouTube video to transcript using openAI whisper and summary using OLLama

Reinforcement Learning

Super Resolution using U-Net like architecture

import matplotlib.pyplot as plt
import torch

%matplotlib inline
%config InlineBackend.figure_format='retina'

# Download some MNIST to demonstrate…

Jul 12, 2023

Trees using NetworkX

import networkx as nx
import numpy as np
import matplotlib.pyplot as plt
import pandas as pd

%matplotlib inline

# Retina display
%config InlineBackend.figure_format =…

Jun 12, 2023

Positional Encoding

import numpy as np
import matplotlib.pyplot as plt
import torch
import seaborn as sns
import pandas as pd

dist =torch.distributions

sns.reset_defaults()
sns.set_co…

Jun 9, 2023

Bayesian Active Learning with Disagreement (BALD)

import numpy as np
import matplotlib.pyplot as plt
import torch
import seaborn as sns
import pandas as pd

dist =torch.distributions

sns.reset_defaults()
sns.set_co…

May 26, 2023

PINNs are a class of neural networks for solving partial differential equations. They are trained to minimize the residual of the differential equation, and the boundary and…

PINN

May 5, 2023

Gradient wrt input for a simple model for adversarial attacks

SIREN paper implementation

Reconstruction from ReLU

Reconstruction from Sine

Apr 28, 2023

In this post, I’m noting some observations from the SIREN paper.

SIREN paper

Apr 27, 2023

Heteroskedastic and Homoskedastic MLPs in PyTorch for regression

Tensorboard

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

import numpy as np
import matplotlib.pyplot as plt

from torch.utils.da…

Mar 21, 2023

Residual Connections in PyTorch

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

import numpy as np
import matplotlib.pyplot as plt

from torch.utils.da…

Mar 18, 2023

In this post, we look at AutoPyTorch, a framework for automated machine learning.

AutoML PyTorch

automl

neural-architecture-search

hyperparameter-optimization

automated-machine-learning

pytorch

Feb 25, 2023

linear-algebra

matrix-decomposition

dimensionality-reduction

visualization

jax

from jax import vmap, jit, grad, vmap
import jax.numpy as jnp

# Enable 64-bit mode
from jax.config import config
config.update("jax_enable_x64", True)
import matplotl…

Feb 1, 2023

This is a work in progress. I will be adding more content to this post in the coming days.

Conformal Prediction

Jan 19, 2023

VSCode Settings and Tips

vscode

Shortcut	Action
Cmd + Shift + P	Open command palette
Ctrl + `	Toggle terminal
Cmd + B	Toggle sidebar
Cmd + Ctrl + F	Toggle full screen

Jan 17, 2023

Binomial and Poisson distribution

This is the first post in a Quarto blog. Welcome!

import jax.numpy as jnp
import jax
from jax import random
import tensorflow_probability.substrates.jax as tfp
tfd = tfp.distributions
import pandas as pd

import matpl…

Welcome To My Blog

news

Sep 14, 2022

Tristan O’Malley

Audio Filtering

import numpy as np
import matplotlib.pyplot as plt
import torch
import seaborn as sns
from functools import partial

sns.reset_defaults()
sns.set_context(context="ta…

Feb 24, 2022

Coordinate descent failure example

Pyro Conditioning

import numpy as np
import matplotlib.pyplot as plt
import torch
import seaborn as sns
import pandas as pd
import pyro

dist =pyro.distributions

sns.reset_defaults()
…

Feb 20, 2022

Linear Regression using Pyro

import numpy as np
import matplotlib.pyplot as plt
import torch
import seaborn as sns
import pandas as pd

t_dist =torch.distributions

sns.reset_defaults()
sns.set_…

Feb 17, 2022

Probabilstic PCA using PyTorch distributions

import numpy as np
import matplotlib.pyplot as plt
import torch
import seaborn as sns
import pandas as pd

dist =torch.distributions

sns.reset_defaults()
sns.set_co…

Feb 17, 2022

Drawing graphical models

Logistic Regression using PyTorch distributions

import numpy as np
import matplotlib.pyplot as plt
import torch
import seaborn as sns
import pandas as pd

dist =torch.distributions

sns.reset_defaults()
sns.set_co…

Feb 14, 2022

GMM

mixture-models

clustering

expectation-maximization

probabilistic-models

unsupervised-learning

import torch
dist = torch.distributions
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
%matplotlib inline

Feb 14, 2022

Let us first look at G1. Look at the illustration below. We have a normal distribution \(p\) and two other normal distributions \(q_1\) and \(q_2\). Which of \(q_1\) and \(q_…

Variational inference

variational-inference

bayesian

approximate-inference

probabilistic-models

optimization

Feb 12, 2022

\(\hat{\theta}_{MAP}=\dfrac{n}{n+\sigma^{2}} \bar{x}+\dfrac{\sigma^{2}}{n+\sigma^{2}} \mu\)

Maximum A-Posteriori (MAP) for parameters of univariate and multivariate normal distribution in PyTorch

PyTorch

Feb 11, 2022

Multiplying a matrix A with a vector x transforms x

Matrix as transformation and interpreting low rank matrix

Feb 11, 2022

Autograd in JAX and PyTorch

JAX

PyTorch

import torch
from jax import grad
import jax.numpy as jnp

Feb 9, 2022

Let us generate some normally distributed data and see if we can learn the mean.

Maximum Likelihood Estimation (MLE) for parameters of univariate and multivariate normal distribution in PyTorch

PyTorch

Feb 9, 2022

We will be studying the problem of coin tosses. I will not go into derivations but mostly deal with automatic gradient computation in TF Probability.

Coin Toss (MLE, MAP, Fully Bayesian) in TF Probability

TFP

Feb 7, 2022

In this notebook I’ll talk about sampling from univariate and multivariate normal distributions. I’ll mostly directly write the code and show the output. The excellent…

Simple Directed Graphical Models in TF Probability

TFP

Linear Regression in TF Probability using JointDistributionCoroutineAutoBatched

TFP

Sampling from univariate and multivariate normal distributions using Box-Muller transform

Feb 4, 2022

Let us first look at G1. Look at the illustration below. We have a normal distribution \(p\) and two other normal distributions \(q_1\) and \(q_2\). Which of \(q_1\) and \(q_…

Understanding KL-Divergence

Jan 29, 2022

Good reference https://tensorchiefs.github.io/bbs/files/21052019-bbs-Beate-uncertainty.pdf

Linear Regression in Tensorflow Probability

Jan 28, 2022

Let us generate some normally distributed data and see if we can learn the mean.

Testing out some distributions in Tensorflow Probability

Jan 26, 2022

Learning Gaussian Process regression parameters using mini-batch stochastic gradient descent

How to learn the parameters of a GP

Sep 3, 2021

Linear Regression from scratch in Julia

using Plots
theme(:default)

using LinearAlgebra
using LaTeXStrings

Sep 1, 2021

In this post, I will look at a simple application:

Probabilistic Programming in Pyro

Aug 20, 2021

Blurring an image selectively using Affinity Photo

setup

Blurring an image selectively using Affinity Photo

Jun 19, 2021

Audio filters

signal-processing

audio

filtering

digital-signal-processing

python

scipy

Audio filtering techniques and applications

Jun 18, 2021

Running Python scripts on server over ssh and getting back content

setup

python

ssh

remote-development

networking

terminal

Running Python scripts on server over ssh and getting back content

Jun 17, 2021

Some of my shortcuts on the iPad

setup

ipad

shortcuts

automation

productivity

ios

Some of my shortcuts on the iPad

Jun 16, 2021

My iPad Setup

setup

ipad

mobile-development

productivity

ios

My iPad computing setup

Jun 14, 2021

10 min

My Mac Setup

setup

macos

development-environment

productivity

tools

My Mac Setup

Jun 12, 2021

Excellent, now let us start from random weight initialisations and use backprop to come to our result

GAN

deep-learning

generative-models

adversarial-networks

neural-networks

keras

tensorflow

Feb 28, 2020

Bayesian Linear Regression

bayesian

linear-regression

uncertainty-quantification

probabilistic-models

statistics

A programming introduction to Bayesian Linear Regression.

Feb 20, 2020

An Example Markdown Post

markdown

A minimal example of using markdown with fastpages.

Jan 14, 2020

Gaussian Processes

gaussian-processes

bayesian

uncertainty-quantification

probabilistic-models

kernel-methods

An interactive exploration of Gaussian processes.

Aug 20, 2019

16 min

Placement-Preparation-2018-1-HashMap

academia

HashMaps for programming interviews

Aug 18, 2018

Visualising Electricity Access Over Space and Time

sustainability

How is the world changing over the years!

Jun 26, 2018

Mapping location of air quality sensing in India

air quality

AQ sensing in India

Jun 21, 2018

Active Learning

A programming introduction to query by committee strategy for active learning

Jun 16, 2018

Signal denoising using RNNs in PyTorch

Denoising

Jan 13, 2018

CS Ph.D. lessons to my younger self

academia

Some personal reflections..

Jan 7, 2018

17 min

Neural Networks for Collaborative Filtering

Neural networks to learn the embeddings! and how to combine them

Dec 29, 2017

I have written a few posts earlier about matrix factorisation using various Python libraries. The main application I had in mind for matrix factorisation was recommender…

Recommend keras

recommendation-systems

collaborative-filtering

deep-learning

neural-networks

keras

tensorflow

Dec 18, 2017

Adagrad based matrix factorization

Adagrad optimizer for matrix factorisation

Aug 13, 2017

In this post, I’ll be using Adagrad for solving linear regression. As usual, the purpose of this post is educational. This link gives a good overview of Adagrad alongwith…

Linear regression adagrad vs gd

optimization

gradient-descent

adagrad

linear-regression

algorithms

Aug 12, 2017

A while back, I read this wonderful article called “Top 50 ggplot2 Visualizations - The Master List (With Full R Code)”. Many of the plots looked very useful. In this post…

Fifty ggplot python 1

visualization

data-visualization

ggplot

plotnine

python

data-analysis

Aug 2, 2017

Linear regression with prior (using gradient descent)

What if we start from some prior!

Jun 15, 2017

Data exploration using widgets in Matplotlib

visualisation

Exploring data in Matplotlib

Jun 14, 2017

Constrained Non-negative matrix factorisation using CVXPY

Constrained NMF using CVXPY!

Apr 21, 2017

Out of Tensor factorisation

Out of tensor factorisation

Apr 20, 2017

Out of matrix non-negative matrix factorisation

What if we to predict for entries not within the matrix?!

Apr 19, 2017

In the previous post, I had discussed regarding the EM algorithm from a programmer’s perspective. That blog post sparked an interesting discussion on Twitter which has led…

MCMC coins

bayesian

mcmc

monte-carlo

probabilistic-models

parameter-estimation

Jul 1, 2014

Latexify Matplotlib

visualisation

Towards amazing plots in research papers!

Jun 2, 2014