Practical Homework

A curated collection of solved practical assignments from Machine Learning, Deep Learning, Efficient Deep Learning Systems, Reinforcement Learning, Generative AI, and Matrix Analysis courses at HSE and the Yandex School of Data Analysis (YSDA).

Each course lives in its own top-level directory named Topic-Org, and every assignment is a self-contained folder following a uniform hwN-topic / labN-topic convention.

Contents

• Machine Learning (HSE)
• Deep Learning (HSE)
• Deep Learning in Natural Sciences (HSE)
• Generative AI (YSDA)
• Efficient Deep Learning Systems (YSDA)
• Reinforcement Learning (YSDA)
• Matrix Analysis (HSE & YSDA)

---

Machine Learning (HSE)

#	Assignment	Description
1	Linear Regression	Data analysis, feature engineering, visualization, and model fitting with scikit-learn.
2	Gradient Descent	From-scratch gradient descent variants (descents.py) and linear regression training (linear_regression.py).
3	Linear Classification	SVM, logistic regression, probability calibration, feature transformation, and multi-class classification on a near real-world business case.
4	Decision Trees	Classification trees, hyperparameter analysis, a custom tree implementation, and regression trees with linear models in the leaves.
5	Boosting	Custom gradient boosting, optimization techniques, and experiments with CatBoost.

Deep Learning (HSE)

#	Assignment	Description
1	Feed-Forward Neural Networks	Fully connected network with forward/backward propagation, hyperparameter tuning, and performance visualization.
2	Convolutional Neural Networks	CNN for image classification with convolutional and pooling layers.
3	Recurrent Neural Networks	LSTM-based sequence model with a custom dataset loader, model definitions, and training pipeline.
4	CLIP	Contrastive Language–Image Pretraining: text and image encoders with a projection head, trained on paired image–text data.

Deep Learning in Natural Sciences (HSE)

#	Assignment	Description
1	Bioinformatics	Deep neural networks for genomic sequence analysis and functional annotation prediction.
2	Materials Science I	Convolutional and graph-based networks for forecasting material properties from atomic structures.
3	Materials Science II	ML workflows for materials discovery with uncertainty quantification and robustness analysis.
4	Physics-Informed Neural Networks	Networks with physical laws embedded in the loss to solve partial differential equations.

Generative AI (YSDA)

#	Assignment	Description
1	Flow Matching	Flow Matching model training with JIT compilation and a REPA-based architecture.
2	Flow Map Models	Exploration of flow map models for generative tasks.
3	MMD Distillation	Few-step generator distillation with an added Maximum Mean Discrepancy (MMD) objective.
4	MAR with Flow Matching Head	Masked Autoregressive image generation with a per-token flow matching head, built on a VAE latent space.

Efficient Deep Learning Systems (YSDA)

#	Assignment	Description
1	Testing and Experiment Management	Debugging and testing a DDPM training pipeline, W&B logging, Hydra configuration, and reproducible experiments with DVC.
2	Fast Training Pipelines	Custom static and dynamic loss scaling, efficient sequence batching, and profiling data-loading and GPU workloads.
3	Distributed Data Parallel Training	Custom SyncBatchNorm, gradient synchronization, distributed metric aggregation, and Ring All-Reduce.
4	Fully Sharded Data Parallel	A custom FSDP implementation with parameter sharding and overlapping communication with forward and backward computation.
5	Model Service Deployment	Containerized instance-detection service with HTTP and gRPC APIs, Prometheus metrics, and deployment tests.
6	Inference Algorithms	W8A8 weight and activation quantization, optimized integer matrix multiplication, and speculative decoding.

Reinforcement Learning (YSDA)

#	Assignment	Description
1	Cross-Entropy Method	Deep cross-entropy method with neural network function approximation for control tasks.
2	Dynamic Programming	Value iteration and policy iteration for solving Markov Decision Processes.
3	Model-Free RL	Monte Carlo, Temporal Difference, and on-/off-policy methods with sample-efficiency experiments.
4	Deep Q-Networks (DQN)	DQN with experience replay and target networks for discrete action spaces.
5	Continuous Control (TD3 & SAC)	Twin Delayed DDPG and Soft Actor-Critic for continuous action spaces.

Matrix Analysis (HSE & YSDA)

#	Lab	Description
1	Image Search via SVD	Singular value decomposition for eigenfaces and similar-image retrieval.
2	Recommendations via ALS	A recommender system based on low-rank approximation of sparse matrices.
3	Tomography	Tomogram reconstruction from ray-intensity data.
4	Resistor Network Voltages	Iterative methods for solving large sparse linear systems in a resistor network.

---