Techniques

We use several different classes of techniques within our lab, depending on the nature of research questions we’d like to answer. This ties into the types of projects that we run, which you can read more about on the projects page.

Lots of our lab members have familiarity with deep learning, and we now have quite a few users of PyTorch who are working on projects that involve improving our understanding or the performance of deep learning systems.

Thanks to Kai and other members of the lab, we have a page on Deep Learning lectures, talks and tutorials that is to be found here.

We are collaborating with several groups around Imperial to develop DL techniques for specific applications, including medical image analysis, visuo-motor control and other problems of inference. In some cases, this involves building training environments for AI systems, building robots or custom imaging systems.

Hypothesis Testing is another area we work on. We tend to favour building ad hoc statistical tests, based on the only way we approve of: drawing samples of data that correspond – as best as we can – to the so-called null hypothesis. This is not as crazy and left-field as it may sound. For more on this, contact the lab leader.

We have experience with the hand construction of image features. Again, this is not as dumb as it sounds. Sure, you can use the kth layer of your favourite deep network. But do you really need to…?