Portfolio · Written for this search · ml
ML
Computation desk
Machine learning remains the widest industrial use of gradient-based optimisation, tensor algebra, and high-dimensional statistical inference at scale.
- Data input
- Forward pass
- Loss calculation
- Gradient computation
- Parameter update
Training architecture
A training loop feeds batches of input data through a parameterised function, computes a scalar loss, and uses backpropagation to derive gradients with respect to every weight. The optimiser then adjusts parameters in the direction that reduces the loss, repeating until convergence or a resource budget expires. Frameworks such as PyTorch and JAX handle the graph construction and gradient tape automatically, so practitioners write only the forward pass and loss definition.
Parallelism appears at every level: data-parallel replication across devices, model-parallel sharding for large architectures, and pipeline-parallel staging when depth exceeds accelerator memory. Techniques like mixed-precision arithmetic and gradient checkpointing trade numerical stability and recomputation cost for memory footprint. The result is that billion-parameter models train in hours rather than months, provided the cluster and communication fabric are correctly configured.
Model families
Convolutional networks exploit spatial locality in images and sequences, pooling and striding to build hierarchical features. Transformers replace convolution with self-attention, allowing each token to attend to every other token within a context window, and have become the default architecture for language and vision-language tasks. Diffusion models generate samples by learning to reverse a noise process, training a denoiser at multiple time steps and then iterating that denoiser at inference to recover signal from pure noise.
Reinforcement learning trains agents to maximise cumulative reward by sampling trajectories and adjusting a policy network with algorithms such as PPO or SAC. Offline RL learns from logged data rather than environment interaction, important when simulation is expensive or unsafe. The boundary between supervised and reinforcement methods continues to blur: language models trained with RLHF use policy gradients to align outputs with human preference, and world models replace hand-coded simulators in robotics and planning.
Deployment and inference
Inference optimisation begins with quantisation, reducing weights and activations from 32-bit floating point to 8-bit or even 4-bit integers, sometimes with negligible accuracy loss. Pruning removes low-magnitude weights or entire channels, and knowledge distillation transfers the behaviour of a large model into a smaller student network. Compiled inference engines fuse operations, eliminate redundant memory copies, and schedule kernels to match hardware capabilities, achieving latencies measured in milliseconds rather than seconds.
Production systems wrap models in serving infrastructure that handles batching, caching, and version rollout. Monitoring tracks drift in input distributions and output confidence, triggering retraining when performance degrades. Edge deployment moves inference to devices with limited power and memory, requiring co-design of model architecture and chip accelerators. The engineering effort to maintain a live ML system often exceeds the research cost of the original model, and reliability depends on the same disciplines of testing, versioning, and observability that govern any distributed service.
Finding from this desk
ML methods now span supervised classification, generative modelling, and reinforcement learning, each relying on automatic differentiation, stochastic gradient descent, and parallelised matrix operations that match the architecture of modern accelerators.