Abstract

Please note that the speaker is no longer available on this date so the talk has been cancelled. We will look at arranging an alternative.

Occam’s razor states “Plurality is never to be posited without necessity.” We begin by examining how small a neural network can distinguish the digits in the MNIST data set. For continuous problems, there are Universal Approximation Theorems. For any function and a criterion of closeness, if there are enough neurons in a neural network, then there exists a neural network with that many neurons that does approximate the function that close. However, is this desirable? Simpler systems facilitate human insight. We look at the following challenges in data approximation also known as inference in machine learning.

• Ill-conditioned matrices.
• Modeling data, keeping the model simple while explaining the data adequately.
• Choice of model space.
• New data arriving.
• Model updates.
• Confidence in model predictions.
• Informed data collection.
• Limited computing power, up and down-link capacity.

And introduce a system which builds up complexity when it is necessary.

The zoom link is https://cam-ac-uk.zoom.us/j/81161988457?pwd=TB5DgLyL0RLQROGBA4LC9jLnlKAh5p.1 (Password 355996)