Applied Machine Learning

# Build and reason about real modeling pipelines beyond the foundations: regression and classification from loss functions to Bayesian uncertainty, representation and generative models via latent spaces, and ill-posed inverse problems solved with deep priors.

# a suggested order, not a lock: jump anywhere, checkpoints track what you finish

1. Regression6 posts · 1h 6m

   From least squares as Gaussian maximum likelihood to regularization, kernels, deep regression networks and the Bayesian view, where a point estimate gives way to a predictive distribution.

   1. Foundations of Regression: Loss Functions, Linear Models, and the Probabilistic ViewDeep Dive10 min scheduled
   2. Regularization, Feature Engineering, and Basis Expansions in RegressionDeep Dive11 min scheduled
   3. Ill-Posed Regression: Priors, Stability, and the Geometry of AmbiguityDeep Dive12 min scheduled
   4. Nonlinear and Kernel-Based Regression: Beyond the Linear FrontierDeep Dive10 min scheduled
   5. Deep Regression Networks and Structured OutputsDeep Dive11 min scheduled
   6. Probabilistic and Bayesian Regression: From Point Estimates to Predictive DistributionsDeep Dive12 min scheduled
2. Classification5 posts · 54 min

   The probabilistic view of classification, then calibrated, uncertain and interpretable classifiers, attention-based architectures, and ensembles you can trust.

   1. Foundations of Classification: Loss Functions, Linear Models, and the Probabilistic ViewDeep Dive10 min scheduled
   2. Probabilistic and Bayesian Classification: Modeling Uncertainty in PredictionsDeep Dive11 min scheduled
   3. Feature Representations and Embeddings for ClassificationDeep Dive12 min scheduled
   4. Advanced Neural Networks for Classification: Architecture, Regularization, and AttentionDeep Dive10 min scheduled
   5. Interpretability, Uncertainty, and Ensembles: Making Classifiers TrustworthyDeep Dive11 min scheduled
3. Latent Spaces6 posts · 1h 12m 1 self-checks

   Learn representations, not just labels: autoencoders, variational models and the ELBO, the geometry of latent space, and learning from distances rather than features.

   1. Autoencoders: Learning to Compress and ReconstructDeep Dive12 min scheduled
   2. Variational Autoencoders: Learning Latent DistributionsDeep Dive12 min scheduled
   3. The Geometry of Latent Space: Manifolds and DisentanglementDeep Dive12 min scheduled
   4. Latent Dynamics: Sequential and Temporal Generative ModelsDeep Dive12 min scheduled
   5. Dissimilarity Spaces: Learning from Distances, Not FeaturesDeep Dive12 min scheduled
   6. Latent Dissimilarity in Practice: Regression and ClassificationDeep Dive12 min scheduledquiz
4. Inverse Problems and Deep Priors6 posts · 1h 12m optional

   When the forward operator destroys information the inverse is ill-posed. Regularization from Tikhonov to networks used as learned priors, and reconstruction when the operator bites back.

   1. Ill-Posed Problems and Deep Learning: When Classical Inversion FailsDeep Dive12 min scheduled
   2. Deep Networks as Learned Priors: Representation as RegularizationDeep Dive12 min scheduled
   3. Inverse Problems and Regularization: From Tikhonov to Learned PenaltiesDeep Dive12 min scheduled
   4. Denoising and Signal Restoration: Classical Filters to Deep ReconstructionDeep Dive12 min scheduled
   5. Solving the Objective: From Proximal Gradient to Learned OptimizationDeep Dive12 min scheduled
   6. Super-Resolution, Deblurring, and Inpainting: Reconstruction When the Operator Bites BackDeep Dive12 min scheduled
5. Advanced ML Topics2 posts · 24 min optional

   Two advanced corners worth a detour: energy-based models and their intractable partition function, and sliding-window optimization on streams.

   1. Energy-Based Models: Boltzmann Machines and Stochastic Latent VariablesDeep Dive12 min scheduled
   2. Sliding Window Optimization: Efficient Learning on Streams and SequencesDeep Dive12 min scheduled
6. Applied Machine Learning CapstoneDeep Dive5 minquiz

   Capstone: the modeling toolkit end to end, from the loss-likelihood correspondence and calibration to the VAE's ELBO and ill-posed inversion.

7. You can design a modeling pipeline end to end and reason about its uncertainty, representations, and stability