95-876 Practical Deep Learning: Models & Applications
6 units, Carnegie Mellon University
Neural Nets for Realworld Data builds on introductory machine learning by shifting focus from tabular data and general methodology to the modeling challenges posed by complex, realworld data. In practice, data is often unstructured, highdimensional, and multimodal, including time series, discrete sequences, graphs, images, and text. These data types require specialized modeling paradigms and inductive biases that traditional approaches do not always capture well. This course equips students with a practical understanding of neural network architectures designed for different data modalities. A central theme is representation learning the use of neural networks to transform complex inputs into informative embeddings that can support downstream prediction and decisionmaking tasks. Students will learn how neural networks differ from traditional methods, when they are advantageous, and how their architectural choices reflect the structure of the data they model. The course is organized around a progression from classical alternatives to modern neural approaches, helping students understand both the motivation for and the tradeoffs of each architecture. Key model families include autoencoders for representation learning, RNNs and LSTMs for sequential data, CNNs for images, Transformers for continuous and discrete sequences as well as vision tasks, graph neural networks for relational data, multimodal architectures for integrated data sources, and generative model families. In addition, the course covers important training and evaluation topics such as optimization dynamics, regularization, model selection, scalability and efficiency, as well as interpretability. Through assignments and coding projects, students will gain both conceptual understanding and handson experience designing and implementing neural network solutions for realistic problems across diverse data modalities.
Prerequisites: none listed
Offered: Spring 2027
Courseweave needs JavaScript to show sections, workloads and planning tools.