Jonathan Teo

Jonathan Teo

Computational Systems Biologist & Data Scientist
Learning to model;
Modeling to learn

All Posts

Can a neural network perform PCA? Part 2 - The transformation

Can a neural network perform PCA? Part 2 - The transformation

guide

Published: October 23, 2020

PCA-like Autoencoder The goal of creating a PCA-like autoencoder is to provide context to the underlying weights of the neural network. In a neural network, for linear activation, the first layer is essentially a projection of the data onto a space determined by the weights of the layer. The cartoon below illustrates how a neural network projects data onto a vector space determined by the weights. To avoid confusion, all subsequent mentions of weights refer to the weight vectors as show...

Can a neural network perform PCA? Part 1 - Setup

Can a neural network perform PCA? Part 1 - Setup

guide

Published: October 22, 2020

PCA using autoencoder When analyzing large datasets, it is important to preprocess the data to prevent potential overfitting (curse of dimensionality). Dimension reduction is one such technique that identifies a small set of features to represent a large dataset. Features are chosen based on how well they capture underlying structure in the data based on certain criteria. For example, in principle component analysis (PCA), features are selected from principle components that best explai...

How does the value of a HDB flat depreciate over time? Part 1

How does the value of a HDB flat depreciate over time? Part 1

opinion

Published: July 30, 2020

Suppose you buy a resale apartment with 80 years of lease remaining today. 10 years later, the apartment appreciates by 25% or ~2.5% per year. Ignoring potential rental yield, 2.5% is a respectable rate of return especially since the apartment is also functioning as a living space for the duration.

Taxonomic Classification - Centrifuge

Taxonomic Classification - Centrifuge

guide

Published: July 21, 2020

Centrifuge Create an FM-Index of the reference sequence. Uses a BWA to align reads to an FM-index of reference genomes. Outputs the matches. If the number of matches is large, the algorithm will climb the phylogenetic tree to assign a higher phylo instead.

Taxonomic Classification - Kraken

Taxonomic Classification - Kraken

metagenomics

Published: July 20, 2020

Kraken Kraken 1 Kraken is a k-mer based taxonomic classifying tool for metagenomic sequencing data. It generates a k-mer hash lookup table from known reference genomes and assigns the lowest common ancestor (LCA) to each k-mer. Given a seqquencing read (S), break down S into a series of k-mers. Instead of aligning sequencing reads to the reference genomes, look up the hash table for an exact match of the k-mers and assign it the respective LCA.