Jeremy Howard on Platform.ai and Fast.ai (Full Stack Deep Learning - March 2019)

Jeremy Howard argues that “augmented machine learning”—tight human–computer collaboration—beats fully automated ML pipelines for most practical problems, and that the fastest path to strong results often comes from combining what humans do well (rapid perception, similarity/difference judgment, and targeted labeling) with what computers do well (speed, memory, and large-scale optimization). He frames this as a direct challenge to AutoML’s goal of minimizing human involvement, saying that until computers become better than humans at everything (a far-off scenario), the best systems will keep humans in the loop.

Howard illustrates the point with Platform.ai, a workflow designed to make labeling and dataset building dramatically more efficient. The system starts with unlabeled or minimally labeled images (cars, faces, and other categories) and uses interactive projections to let humans quickly find clusters and outliers. Humans rapidly identify where examples look similar, then zoom in to spot differences—turning perception research into a practical labeling strategy. When the model struggles (for example, distinguishing car fronts from backs), the workflow adapts: humans provide a few examples, and the system generates “find similar” results using a pretrained ImageNet model. The interaction becomes a kind of visual dialogue—humans indicate what they’re trying to separate, while the model proposes projections that maximize the difference between chosen groups. With surprisingly few labels (on the order of hundreds), the system trains a classifier that reaches high accuracy (he cites 92% in the car example) and produces better embeddings after each round.

He extends the same idea beyond labeling. In a face dataset, the first projection already separates men and women well enough that bulk selection becomes feasible; later projections reveal additional structure (like sunglasses as an emergent category). The payoff is iterative: improved models yield better projections, which speed up the next labeling cycle. Howard also connects this approach to broader research on “human-augmented” training, emphasizing that studying human strengths—especially perception—can lead to better ML systems than trying to remove humans entirely.

From there, the conversation shifts to how data scientists can get strong results quickly without massive compute or exhaustive hyperparameter searches. Howard highlights research from Fast.ai’s ecosystem and its fellowship program, where the central theme is “spend a little human time instead of a lot of GPU time.” Examples include using a learning-rate finder (a quick procedure that identifies a good learning rate by running short experiments) and findings that, for transfer learning, default hyperparameters often work nearly as well as elaborate tuning. He describes classroom outcomes where beginners frequently reach near-perfect validation accuracy with only 100–200 images after using transfer learning and sensible defaults.

He then stacks additional practical accelerators: test-time augmentation (TTA) to average predictions over multiple inference-time transforms; progressive resizing to train first on smaller images and later on larger ones (speeding training while improving generalization); and “one-cycle” learning rate schedules paired with momentum changes to train faster and more reliably. Howard also discusses training reliability tricks for transfer learning (training newly initialized layers more aggressively), and optimizer details such as correct decoupled weight decay (AdamW) and ways to prevent Adam instability via gradient clipping or adjusting epsilon. The overall message is consistent: strong ML results come from combining smart defaults, efficient training heuristics, and deliberate human–computer interaction rather than brute-force automation or compute-heavy search.