TensorShin · an interactive lab for machine learning

Learn the math.
Code the ideas.

Abstract. Build an intuition for machine learning. Connect the concepts, put them into code, and see how far you’ve come.

Start your journey Read the curriculum

topics: 644
domains: 10
problems: 160
papers: 832

Fig. 1Gradient descent,

\theta_{t+1} = \theta_t - \eta\,\nabla L(\theta_t)

, on a non-convex loss surface. Runs started on different slopes settle in different minima; 1b shows the same surface in plan view.

01Understand

Built around Stanford CS229, MIT OpenCourseWare, fast.ai and the original papers.

Knowledge graph

Follow your curiosity. Give it a direction.

10 domains, from the mathematics underneath every model to the systems that ship them. Choose a row to see where it begins.

Curriculum domains grouped by stage, with topic counts and estimated study hours. Select a domain to see its topics.
#	Topics	Est. hours
AFoundationsThe ideas underneath every model
01	40	129
02	63	112
03	58	128
BModelsGive those ideas something to learn
04	61	122
05	66	106
06	54	106
07	63	100
CSystemsFrom attention to working systems
08	92	139
09	72	97
10	75	97

Table 1The curriculum by domain. Bars are proportional to the largest domain; hours are the sum of each domain’s topic estimates.

02Implement

Every exercise runs in an isolated Python workspace. No GPU, no local setup.

Beyond “I get it.” To “I built it.”

Implement the algorithms you’re learning, then inspect what they do. Move the temperature and watch the distribution respond.

Listing 1 · Python

import numpy as np

def softmax(logits, temperature=1.0):

# Small change. Different confidence.

z = np.array(logits) / temperature

z = z - np.max(z)

exp_z = np.exp(z)

return exp_z / np.sum(exp_z)

logits = [2.0, 1.0, 0.1, -1.0]

probs = softmax(logits, temperature)

p_i = \dfrac{e^{z_i/T}}{\sum_j e^{z_j/T}}

Fig. 2

TemperatureT = 1.00.2 · sharperflatter · 3.0

H(p): 0.961 nats
max p: 0.638
Σ p: 1.000

Listing 1 · Fig. 2Temperature-scaled softmax. Low T concentrates probability on the largest logit; high T spreads it out, raising the entropy H(p).

Further experiments

All 160 problems

03Remember

The curve is a simple exponential model to show the idea; your own schedule adapts to how confident you feel.

Learning fades. Reviews bring it back.

Every review lets a memory last longer before the next one. TensorShin schedules them for you, from one day to one month out.

R(t) = e^{-\Delta t / S}

Retention on day 60no review 0%spaced review 70%

Fig. 3Retention after learning, without review and with reviews on days 1, 3, 7, 16 and 35. Each review resets R and lengthens the stability S.

Map: A living picture of what you know, topic by topic.
Review: Confidence-based intervals that grow as a topic sticks.
Record: Study sessions, streaks and milestones.

Start tracking your progress

AAppendix

Questions

A.1How is this different from following a course or a playlist?

A course gives you one fixed order. Here every topic links to its prerequisites, so you can see what an idea depends on and start from where you are. Each topic connects to curated lectures, the original papers, Python problems, and a review schedule, so understanding, implementation and recall happen in one place.

A.2How do I know my implementation is actually correct?

Each of the 160 problems runs in an isolated Python sandbox, with nothing to install. While you work you see visible tests and their output; submitting runs hidden tests as well, so a solution has to hold up on inputs you haven’t seen.

A.3What happens when I forget something I marked as learned?

When you complete a topic you rate your confidence, and that schedules a review between one day and one month later. Recall it and the interval grows; feel unsure and it shrinks; forget it and the topic reopens for learning. Daily recall quizzes feed the same schedule.

A.4What does a Paper Lab involve beyond reading the paper?

Each lab pairs a landmark paper, such as ResNet, BERT, LoRA or DDPM, with three challenges: build its core mechanism, repair a realistic bug, and run a controlled experiment. A lab counts as complete only when your code passes its checks, the measurements are captured, and you write a conclusion the evidence supports.

A.5Can I follow my own order instead of the curriculum’s?

Yes. Build an ordered path from any topics, make it your active path, and track progress through it. Paths can stay private, be shared, or be forked from someone else’s and adapted.

04Begin

One topic. See where it takes you.

Choose something you want to understand, and start there. Sign in with Google.

Start learning