approximation-theory

cnn

gnn

Generalisation

Empirical Risk vs Population Risk

  • Approximation
  • Optimisation
  • Generalisation

rnn

transformers

Probabilistic Graphical Models

  • joint distribution of random variables modelled as DAGs.
  • TODO: write more about them

Reference:

representation-learning

gaussian-processes

generative-modelling

Beyond IID assumption

References

  • Murphy, Kevin P. Probabilistic machine learning: Advanced topics. MIT press, 2023.

Transfer Learning

Few-shot learning: Learning with very little data

Ways to perform transfer learning:

  • Knowledge of mapping
    • Finetuning
  • Knowledge of outputs
    • Distillation
  • Knowledge about inputs
    • Prompting

Contrastive Learning + Generative Modelling

Meta Learning

Scaling Laws

Practical

Question

  • what effects does L2 norm, layer norm, RMS norm, batchnorm have on the data geometry? How does the statistic change? How to choose? and which to prefer?
  • Why layernorm can be a footgun in low dimensional spaces?

Data

  • Pre-processing
    • Summary statistic
    • Visualisation
    • Shape
      • For testing, check dimensions are not the same. Say, for CNN, the dimensions are different.
      • Einops library
    • Type-checking
  • Code assertions for size and type
  • Data augumentation
  • How to maximise learning from data? How to augument data so that it’s hard to learn from it?