Nicole Junkermann, NJF Capital founder, presents an A-Z of Artificial Intelligence – a series of short videos focused on key areas of interest in this hugely topical and consequential field.
Vast sums of computing power (and money!) are spent training AI systems to perform tasks like recognising visual data or interpreting human speech.
Inference is what happens when these systems are put to work for real.
In training, a neural network is fed large sums of data until it is able to correctly recognise a red traffic light or a human voice command.
A simplified version of the system is then used to infer things about data it encounters in everyday use, according to its training.
So, an autonomous vehicle stops at a red light and a digital assistant turns the lights on when you ask it to.
AI training and inference require different chip capabilities.
Where graphic processing units (GPUs) have dominated the chip market in AI training, field-programmable gate arrays (FPGAs) and specialised chips like Groq’s Tensor Streaming Processor are particularly suited for inference.