From Perceptron to MNIST: A Complete Deep Learning Experiment Course Analysis

This article analyzes the deep learning experiment projects from the Computer Science course at AGH Medical University, covering basic perceptrons, the XOR problem, MNIST handwritten digit recognition, and applications of the PyTorch framework. The course design follows a progressive path, helping learners build a solid understanding of core neural network concepts, and is suitable for those who want to delve into principles rather than just call APIs.

This experimental course comes from the deep learning course of the Computer Science major at AGH Medical University and was organized and open-sourced by students. The course design follows a progressive path from theory to practice, with its core value lying in systematicity—gradually moving from simple linear classifiers to multi-layer perceptrons and concepts preceding convolutional networks, and finally completing projects on real datasets. It is suitable for learners who want to deeply understand the principles of deep learning.

Lab1 focuses on perceptrons, revealing linear separability through OR and XOR problems: a single-layer perceptron can solve OR but cannot handle XOR (non-linear boundary), while a two-layer MLP (implemented in PyTorch) successfully solves XOR, proving the necessity of hidden layers and non-linear activation. Lab2 covers PyTorch basics: tensor operations (creation, computation, GPU transfer), comparison of gradient descent between manual and automatic differentiation, custom modules and training loops, to master the core tools of the framework.

Lab3 applies knowledge to the MNIST dataset: building a simple MLP baseline, using 5-fold cross-validation and majority voting ensemble to improve stability; using Shannon entropy to quantify prediction uncertainty; comparing SGD and Adam optimizers, using Optuna for hyperparameter search; experimenting with Dropout layers and mean response ensemble to enhance robustness.

Lab4 applies MLP to complex scenarios: using the California Housing dataset for regression, comparing MLPs of different depths and tuning them; performing multi-label classification (each sample has multiple categories) on Segment and Scene datasets, using StratifiedKFold and a custom Trainer, evaluated with Hamming loss and F1 score; exploring the impact of network width and depth on performance.

The course also includes TensorBoard integration for training visualization, a milestone project on multi-class air quality prediction; and LangChain-related content: LangChain framework (chains, prompts, tools, retrieval), LangGraph workflow orchestration, and LangSmith tracking and debugging platform.

The course's advantages lie in its completeness (clear objectives, runnable code, standard tools). Self-learners are advised to complete the labs in order and not skip basic experiments; educators can use it as university experiment materials or corporate training content, with clear code and sufficient annotations, using open-source tools like PyTorch and Optuna.

The threshold of deep learning lies in the interconnectedness of concepts; this course helps build a complete knowledge graph from perceptrons to deep networks through a progressive path. If you are looking for systematic introductory materials, this open-source project is worth investing time in.