The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / B
7 terms, explained without the techno-murk.
A method for tracing a neural network's error backwards so its weights can be improved.
Imagine a team produced the wrong answer. Backpropagation works backwards through the team to estimate how much each hand-off contributed to the mistake, then makes small corrections.
After misreading an image, a network sends the error backwards through its layers before updating its weights.
A batch is a group of training examples processed together; batch size is how many are in it.
Instead of studying an entire library at once or updating after every page, training often works through manageable bundles. Bigger batches can use hardware efficiently but need more memory and can change how learning behaves.
A batch size of 32 means the model studies 32 examples before making its next weight update.
A standard test used to compare AI systems.
Benchmarks measure things such as accuracy, reasoning, coding or image recognition. They are useful but incomplete: a model can score well on a test and still struggle in real situations.
Several models might answer the same set of maths questions so their scores can be compared.
A pattern of unfair or skewed results in data or an AI system.
Bias can come from historical inequalities, unbalanced data, design choices or how a tool is used. It can affect some groups more than others, even when nobody deliberately programmed discrimination.
A hiring model trained on a narrow history of past employees may unfairly favour people who resemble them.
The balance between a model that is too simple and one that learns its examples too closely.
A very simple rule may miss real patterns; a very fussy rule may memorise every bump in the training data. The aim is a model flexible enough to learn, but steady enough to work on new cases.
A perfectly straight trend line may be too simple, while a wildly wiggly line may copy every accidental fluctuation.
Collections of data so large, fast-moving or varied that ordinary tools struggle with them.
Big data is not defined by one exact size. The phrase usually means that storing, combining or analysing the information needs specialised systems. A large quantity of data is not automatically accurate, useful or representative.
A transport authority may analyse millions of journeys, ticket taps and live vehicle locations.
A system whose inner reasoning is difficult to inspect or explain.
We can see what goes into a black box and what comes out, but not easily trace exactly why it reached that result. Many modern neural networks are partly black boxes because their decisions are spread across huge numbers of learned values.
A loan system rejects an application but cannot give a clear, reliable reason for the decision.