Diagnose a training loop with a tiny controlled case
Trace a batch from input and target to gradient and parameter update. Choose a controlled diagnostic and explain what its result does not establish.
Text lesson
MLE PathChecking your access…
Also in Technical membership bundle, Career membership bundle, Interview Accelerator bundle, ML Residency bundle, AI Engineer Interviews learning bundle, Deep learning foundations learning bundle.
Shared topics and concepts are exploration links, not universal prerequisites.
27 of 27 modules
Trace a batch from input and target to gradient and parameter update. Choose a controlled diagnostic and explain what its result does not establish.
Text lesson
Compute an estimator mean and variance under explicit sampling. Interpret loss jumps and gradient cancellation without declaring convergence.
Text lesson
Select a checkpoint by a predeclared validation rule. Compare capacity and regularization without using training fit or repeated test selection as proof.
Text lesson
Derive a scalar gradient and execute two steps. Distinguish gradient, change and next parameter; diagnose overshoot.
Text lesson
Recover an example-weighted gradient from unequal microbatches. Distinguish example, valid-token and microbatch averages.
Text lesson
Collapse affine-only layers and identify what a nonlinearity changes. Compute ReLU and exact-GELU behavior at stated points.
Text lesson
Identify trainable tensors and retained base dependencies. Distinguish freezing a parameter from cutting derivatives needed upstream.
Text lesson
Compute accumulated versus exponentially decayed squared-gradient state. Compute Adam bias correction and separate its estimates from variance.
Text lesson
Accumulate branch contributions in reverse order. Check a vector-Jacobian product using explicit dimensions and upstream seed.
Text lesson
Compute a training normalization and a fixed-statistics evaluation. Predict which examples or spatial positions influence an output.
Text lesson
Compute inverted-dropout output and expectation. Distinguish ordinary evaluation from an explicitly stochastic evaluation protocol.
Text lesson
Calculate fan-based variance or standard deviation under stated assumptions. Separate second moments, variance and gradient-path products.
Text lesson
Compute LayerNorm and RMSNorm on one vector. Explain shift sensitivity, epsilon limits and learned affine effects.
Text lesson
Compute a scalar meta-gradient through one inner gradient step. Separate support/query roles within tasks and training/testing across tasks.
Text lesson
Compute a stateful update under a stated convention. Explain why restoring weights alone does not resume the same optimizer.
Text lesson
Compute forward and backward contributions through identity and projection shortcuts. Check shapes and identify cancellation counterexamples.
Text lesson
Compute norm clipping and distinguish it from component clipping. Execute a schedule with an explicit counter and accumulation order.
Text lesson
Derive binary and multiclass cross-entropy gradients and logit Hessians. Compute finite stable logit losses and identify numerical domain failures.
Text lesson
Identify inference information and output space. Separate retained-slice regression from population change.
Reading
Compare replay allocation, storage and influence. Derive EWC and identify approximation boundaries.
Reading
Compute trajectory summaries and conventions. Identify measured forward-transfer evidence.
Reading
Distinguish LLM adaptation interventions. Evaluate separated capability and release predicates.
Reading
Specify shared and task-specific update paths. Audit task sampling, reduction and weighting.
Text lesson
State shared weights, forward-state boundary, loss timing and backward horizon. Sum two shared-weight contributions and distinguish a detach from a reset.
Text lesson