Course Objective & Scope
Choosing a neural architecture governs not only accuracy but the trade-off between inductive bias, context length, memory footprint and inference latency. This programme compares modern convolutional backbones (ResNet, EfficientNet, ConvNeXt), vision and general-purpose Transformers (ViT, Swin), linear-time State Space Models for long sequences (S4, Mamba), sparsely activated Mixture-of-Experts (MoE) layers and multimodal fusion architectures.
Participants adapt pretrained backbones from timm and Hugging Face with custom task heads, fuse heterogeneous modalities (tabular + text + image) in a unified network and build an architecture selection matrix for their enterprise use case.
Who Should Attend?
- Data scientists and ML engineers designing or adapting deep learning architectures
- R&D teams building vision, sequence, audio and multimodal systems
- Technical leads balancing model accuracy against hardware latency
- Prerequisite: PyTorch and foundational neural network (CNN/RNN/MLP) knowledge
Key Takeaways
- Choose between CNN, Transformer, SSM and hybrid backbones based on data scale and inductive bias
- Customise ConvNeXt, ViT and Swin backbones with timm for transfer learning
- Benchmark attention against Mamba/SSM blocks on memory and speed for long sequences
- Fuse tabular, text and visual inputs using cross-attention and gated multimodal layers
- Make architecture decisions grounded in FLOPs, parameter counts and real hardware latency
Curriculum
Architectural Building Blocks and Inductive Bias
- Residual connections, pre-norm vs post-norm, RMSNorm and SwiGLU
- Depth, width, receptive field and neural scaling laws
Modern Convolutional and Vision Transformer Backbones
- Evolution of convnets from ResNet to EfficientNet and ConvNeXt
- Vision Transformers (ViT), windowed Swin attention and hybrid stems
Long Sequences: State Space Models (Mamba) and Hybrids
- The quadratic attention bottleneck and linear-time alternatives
- Selective state space models (S4, Mamba) and Jamba-style hybrid blocks
Sparse Architectures and Mixture-of-Experts (MoE)
- Top-k routing, expert capacity and load-balancing auxiliary losses
- Total vs active parameters: implications for training and inference
Multimodal Fusion and Architecture Selection Lab
- Early, late and cross-attention multimodal fusion patterns
- Benchmarking FLOPs, memory and accuracy on the enterprise case
Other Courses in This Category
-
Self-attention mechanics, Multi-Head and Grouped-Query Attention, RoPE positional encodings, encoder/decoder/encoder-decoder families, KV caching and FlashAttention, implemented from scratch in PyTorch and with Hugging Face.
- Live Course
- 5 modules
-
The Keras 3 and TensorFlow 2 ecosystem: Functional/Subclassing APIs, high-throughput tf.data pipelines, custom GradientTape loops, tf.distribute and enterprise deployment via SavedModel, TF Serving and TFLite.
- Live Course
- 5 modules
-
Quantisation (PTQ, QAT, INT8/INT4), structured pruning, knowledge distillation and deployment with ONNX Runtime, TensorRT, OpenVINO and LiteRT (TFLite) for lower cloud GPU costs and fast edge inference.
- Live Course
- 5 modules
-
Foundations and architectures of deep generative modelling: VAEs, GANs, Normalizing Flows, Denoising Diffusion (DDPM/Latent Diffusion), Flow Matching and autoregressive models for synthetic data and multimodal generation.
- Live Course
- 5 modules
-
MDPs, Q-Learning, DQN, Policy Gradients (PPO, SAC), contextual bandits and Offline RL using Gymnasium and Stable-Baselines3 for dynamic pricing, resource allocation and sequential control problems.
- Live Course
- 5 modules
-
Deep fraud models over transaction sequences, session clickstreams and payment networks: Transformer/RNN sequence encoders, autoencoder anomaly scores, GNNs for fraud rings and hybrid GBDT+DL architectures.
- Live Course
- 5 modules
Get in touch
