← Quant playground

Free practice library

Interview practice

Solve complete questions with progressive hints, worked solutions and Python implementations. Each question explains its assumptions and source match. Original interpretations are labelled explicitly.

752

Ready to practise

174

Not yet playable

926

Catalogue titles

Public standard problems and original exercises; hidden Trading Interview questions have not been verified.

174 titles have been investigated but remain unresolved. Their missing details are recorded without adding substitute exercises. Every playable question has a guided animated answer; selected questions also have custom diagrams.

Practise Learnr’s own questions →Follow a role-specific curriculum →

752 matching entries · page 1 of 16

Machine Learning · Hard

Backpropagation by Hand

Original interpretation · Match confidence: medium

Company use unverified
About this question’s source match

The title with subcategory Neural Networks specifies the method assessed. This is an original worked scenario using the standard method documented in the public reference. The inaccessible Trading Interview statement has not been verified; numerical data, conventions and framing are newly authored rather than claimed source reproductions.

A scalar neural network is a=w*x+b, h=ReLU(a), yhat=v*h+c and loss=(yhat-y)^2/2. For x=2,y=1,w=1,b=-1,v=3,c=0 compute the forward values and gradients with respect to w,b,v,c. Then explain the gradients when a<0 and the convention needed at a=0.

Assumptions and constraints

  • All data and timing conventions are those explicitly stated in the prompt.
  • Show intermediate calculations; use the specified normalization and boundary conventions.

Examples

x=2,w=1,b=-1,v=3,c=0,y=1
→ dw=12

Residual 2 propagates through output weight 3 and input 2.

Company evidence · unverified

No company-specific source has been matched to this exercise. Its references support the topic or solution method; they do not establish employer use.

References

Read the supporting material without leaving your question.

  • d2l.ai

    Dive into Deep Learning: backprop

    Standard method and mathematical definitions; not the hidden catalogue question or the newly authored numerical data.

Catalogue metadata captured 2026-10-12. Metadata source.