Deep Learning with PyTorch
A hands-on course in modern PyTorch: tensors and autograd, nn.Module design, data loading, a training loop you understand line by line, GPU and mixed-precision training, torch.compile, and exporting a model for deployment.
mod Track
Designing, fine-tuning, aligning and evaluating machine learning models.
Choose technical study according to the model problem you face. Public topics span classical machine learning, transformers, vision, speech, fine tuning, alignment and model evaluation. More specialized listings address multimodal systems, interpretability and robotics. Compare prerequisites carefully rather than treating every topic as a beginner starting point. Use the published outcomes to separate learning about model behavior from implementing a training method, and check course availability before committing to a sequence.
Build your grounding with AI foundations, then check prerequisites for this track’s available courses.
Available courses are open for enrollment. Waitlisted courses are still being prepared or reviewed.
Compare this track in the course catalogRelated learning: Data for AI · Training at Scale · Inference, Serving and MLOps
Track bundle
Save
on sale today, with every lesson, exam and certificate. If you are signed in, courses you already own are left out and the price at checkout drops to match.
Only the courses on sale in this track today are included. Courses added to the track later are not part of the bundle; buy them separately, or get the All-Access Pass, which includes new courses while it lasts.
A hands-on course in modern PyTorch: tensors and autograd, nn.Module design, data loading, a training loop you understand line by line, GPU and mixed-precision training, torch.compile, and exporting a model for deployment.
Build a decoder-only transformer piece by piece: tokenization, embeddings, scaled dot-product attention, multi-head and grouped-query attention, positional schemes such as RoPE, and the training and sampling loop, then map it onto real open-weight model architectures.
Gradient-boosted trees and linear models still win most tabular problems. This course covers leakage-proof pipelines, honest validation, tuning, calibration and explainability, then packaging, serving and monitoring for drift.
From CNNs and vision transformers to self-supervised backbones, real-time detectors and promptable segmentation: how modern vision models work, how to fine-tune them on your own labeled data, how to measure them with mAP and IoU, and how to deploy them fast.
How speech and audio models turn waveforms into text, labels and voices: features and codecs, CTC, transducer and encoder-decoder ASR, diarization, text-to-speech, audio language models, and how to measure word error rate and latency before you ship.
The foundations of reinforcement learning taught by building them: Markov decision processes, dynamic programming, Monte Carlo and temporal-difference learning, deep Q-networks, policy gradients and PPO. Every algorithm is implemented or trained on Gymnasium environments, with the experiment discipline that RL demands.
A practitioner's course on adapting large language models: when fine-tuning beats prompting or RAG, how to build the dataset, how much GPU memory full, LoRA and QLoRA training really need, and how to evaluate, merge and serve the result.
The full preference-optimization stack after SFT: designing and collecting human and AI preference data, Bradley-Terry reward models, PPO-based RLHF with a KL penalty, DPO and its variants (IPO, KTO, ORPO, SimPO), and the evaluation and failure modes that decide whether alignment worked.
Model-level evaluation done rigorously: how benchmarks are scored, why the same model gets different numbers in different harnesses, how contamination and saturation distort leaderboards, how to build your own eval suite with statistics you can defend, and how to evaluate reasoning, agentic and safety capabilities.
How small language models get their capability: logit and sequence-level knowledge distillation, on-policy distillation, synthetic-data distillation from a frontier teacher, structured and unstructured pruning with distillation-based recovery, and how to evaluate and choose small models honestly.
From the forward noising process and score matching to latent diffusion, diffusion transformers and rectified flow: how modern image and video generators work, how to train and fine-tune them with Diffusers, how to steer them with guidance and control adapters, and how to make sampling fast.
The model-side craft of pretraining: building a web-scale data pipeline with filtering and deduplication, training a tokenizer, choosing architecture and hyperparameters from small-scale experiments, keeping a large run stable, scheduling data and learning rate, and evaluating base-model checkpoints.
Reinforcement learning from verifiable rewards, end to end: building verifiers for math, code and tool use, the GRPO objective and its successors (Dr. GRPO, DAPO, GSPO), distributed rollout systems, reward design for partially verifiable tasks, and the evaluation discipline that separates real reasoning gains from hacked rewards and contaminated benchmarks.
A practical course in mechanistic interpretability: hooking model activations with TransformerLens and nnsight, probing and the logit lens, activation and attribution patching, sparse autoencoders and transcoders with SAELens and Gemma Scope 2, attribution graphs for circuit tracing, and using these tools for steering, safety audits and model debugging.
How modern multimodal models are trained: contrastive encoders (CLIP, SigLIP 2), connector-based vision-language models (LLaVA, Qwen-VL, Molmo), native and early-fusion designs, the web-scale and synthetic data pipelines behind them, staged pretraining and visual instruction tuning, extension to video and audio, and evaluation that catches hallucination and benchmark leakage.
The learning side of modern robotics: robot data and teleoperation, imitation learning with ACT and diffusion policies, vision-language-action models such as OpenVLA, pi0 and Isaac GR00T, simulation and sim-to-real with Isaac Lab and MuJoCo, world models, and the evaluation and safety practice needed before a learned policy controls real hardware.
Trains, fine-tunes and evaluates models.