Core Neural Architectures & Generative Models

Microsoft via Coursera

Go to class Write review

Overview

Google, IBM & Meta Certificates – 40% Off

One Coursera Plus subscription covers most Professional Certificates on Coursera.

Unlock All Certificates

Building production-ready deep learning systems requires fluency across multiple architecture families and modern MLOps tooling. This course covers core neural architectures used in modern AI engineering, from convolutional and sequence models to Transformers and generative systems, alongside the practical toolchain required to build, track, and evaluate them. To succeed in this course, learners should bring strong Python programming skills, a foundational understanding of machine learning, and hands-on familiarity with PyTorch. Hands-on labs and projects utilize TorchVision and Albumentations for image transformations, Hugging Face Diffusers for generative pipelines, and TorchMetrics for specialized evaluation (such as Fréchet Inception Distance). Model tracking, experimentation, and workflows are integrated using MLflow within a Databricks workspace environment. You'll implement CNN architectures, including ResNet and ConvNeXt with transfer learning, build LSTM and GRU models for time series and classification, engineer Transformer attention blocks, fine-tune BERT and ViT models using Hugging Face, and implement VAEs, GANs, and Diffusion pipelines using Hugging Face Diffusers. By the end of this course, you'll be able to select, implement, and evaluate neural architectures for computer vision, temporal, NLP, and generative tasks, and produce documentation supporting production architecture decisions. This course is designed for deep learning engineers specializing in computer vision, time series, and generative AI. Labs use a dual-path design: a gradable track with precomputed outputs and logs supports workflow mastery without live GPU provisioning, while an optional live path is available for enterprise learners with Azure subscriptions and GPU quotas.

Syllabus

CNN Architectures: ResNet & ConvNeXt
- Move beyond basic convolutional layers by designing advanced network topologies. Learners explore, analyze, and implement residual connections, compound scaling, and depthwise separable convolutions to build efficient, state-of-the-art vision models deployable via Dual-Path cloud workflows.
Transfer Learning & Data Augmentation
- Maximize model performance on custom datasets with minimal compute. Learners compare frozen backbone feature extraction versus full fine-tuning, and integrate robust augmentation pipelines using a hybrid suite of TorchVision transforms and Albumentations to prevent overfitting within an enterprise Dual-Path cloud architecture.
Sequence Models: LSTMs, GRUs, and Packing
- Overcome the vanishing gradient problems of standard Recurrent Neural Networks. You will implement LSTMs and GRUs equipped with memory gates, and engineer PyTorch data pipelines utilizing padding and packing to efficiently process variable-length sequences on GPU hardware within a Dual-Path environment architecture.
Temporal Deep Learning & Architecture Selection
- Move beyond standard recurrent models to engineer massive-scale time series forecasters. You will explore Temporal Convolutional Networks (TCNs) and hybrid architectures, and analyze precomputed telemetry to select the optimal model based on latency and accuracy trade-offs.
Transformer Architecture: The Attention Mechanism
- Deconstruct the core engine of modern large language models. This module works through the self-attention mechanism mathematically and programmatically, implements causal masking, and evaluates how shifting from recurrence to attention allows for massive parallelization.
Fine-Tuning Transformers with Hugging Face
- Move from theory to application. This module builds enterprise-grade fine-tuning pipelines using Hugging Face to adapt foundation models—including BERT and ViT—to domain-specific natural language processing and computer vision tasks on Azure Machine Learning (Azure ML).
Latent Spaces: VAEs and GAN Stability
- Dive into the mathematics of generative modeling. You will implement VAEs to learn structured latent spaces, and then tackle the notorious instability of Generative Adversarial Networks (GANs) using advanced regularization techniques.
Diffusion Models & Generative Evaluation
- Transition to the architecture driving modern image synthesis. You will fine-tune Stable Diffusion models using DreamBooth via the Hugging Face Diffusers library, and learn how to quantitatively evaluate generative quality using FID and LPIPS metrics.
GenAI Module: Augmenting Vision and Sequence Workflows with Generative AI
- Learn to accelerate computer vision and sequence modeling workflows using generative AI utilities. You will write prompts to programmatically generate and validate custom Albumentations transform chains, automatically design packed variable-length LSTM topologies, and evaluate synthetic dataset expansions for training robustness.
Project Module: Multi-Domain Deep Learning Architecture Synthesis
- Synthesize your advanced architectural skills to engineer a unified multi-domain deep learning system. You will implement robust object-oriented custom neural modules that combine modern vision blocks, dynamic packed-sequence encoders, and transformer attention matrices into scalable, cloud-ready training routines monitored via MLflow.