Contents
This site has two reading paths: a brief account of my background and research direction, and a growing collection of explanations to learn from.
Research
- About
Background in software, signal processing, and machine learning, based in Lalitpur, Nepal.
- Research interests
Neural decoding, temporal computation, brain–computer interfaces, and biologically inspired models.
Projects
- Sleep-stage classification
Completed internal study of sleep-stage generalization across people; publication and deployment remain pending.
- GPT from scratch
Completed character-level decoder transformer with annotated code and derivations, following Karpathy’s Zero to Hero.
- Motor-imagery decoding
A planned EEG project to distinguish instructed imagined-movement tasks.
- Projects
Completed sleep-stage research and a GPT learning implementation, with motor-imagery decoding planned next.
Explorables
- Backpropagation, one step at a time
Follow values and gradients through a tiny neural network.
- A leaky integrate-and-fire neuron
Change a constant input current and observe membrane voltage and spike times.
- EEG bands, separated
Separate a synthetic signal into conventional frequency ranges and inspect the result.
- Learn with me
Three interactive explanations with worked examples, mathematical derivations and model limitations.
Notes
- Why a shared value needs gradient accumulation
Follow the two paths through a tiny computation graph to see why backward rules must add their contributions.
- A bigram model is a table of conditional probabilities
Count adjacent characters, normalize each row, and understand what the loss actually measures.
- From character IDs to a learned context
Track the shapes from an embedding lookup through a multilayer perceptron language model.
- Batch normalization does not make a distribution Gaussian
Separate centering and scaling from distribution shape, and distinguish training from inference.
- Why the softmax gradient is prediction minus target
Derive the logit gradient of mean cross-entropy and keep the batch scaling explicit.
- Growing context by combining neighboring representations
Follow a hierarchy of character groups and distinguish the learning exercise from the original WaveNet.
- Notes
Reviewed learning notes from Neural Networks: Zero to Hero and signal-processing study.
- What an EEG band can—and cannot—tell us
Frequency bands describe parts of a signal; they do not provide a direct label for a person's thoughts.
Concepts
- Backpropagation
How the chain rule computes the sensitivity of a loss to every parameter.
- Gradient descent
An update rule that moves parameters against the local loss gradient.
- Multilayer perceptron
A composition of weighted sums and nonlinear functions.
- Spike train
Representing neural firing as events in time.
- Leaky integrate-and-fire neuron
A passive membrane equation coupled to a threshold and reset rule.
- Concepts
Mathematical mini-lessons supporting the projects and educational explorables.
- EEG and frequency bands
What scalp voltage recordings and frequency components can tell us.
- Confusion matrix
Counts that reveal which classes a classifier confuses.
- Cohen's kappa
Agreement relative to an independent-marginal comparison.
Now
- Now
Radiology engineering, completed sleep-stage research, and a completed GPT-from-scratch learning project.
CV
- Experience & teaching
Medical-imaging ML, web systems, real-time audio, simulation, and practical teaching.
- Curriculum vitae
Academic CV: education and merit scholarship, completed EEG research, GPT implementation, radiology engineering, and teaching.
- Scholarship and academic record
Full tuition funding across eight semesters through Kathford’s Merit (Jehendar) Scholarship.
Contact
- Contact
Email and GitHub for PhD and research conversations.