Invariant object recognition : biologically plausible and machine learning approaches
Understanding the processes that facilitate object recognition is a task that draws on a wide range of fields, integrating knowledge from neuroscience, psychology, computer science and mathematics. The substantial work done in these fields has lead to two major outcomes: Firstly, a rich interplay between computational models and biological experiments that seek to explain the biological processes that underpin object recognition. Secondly, engineered vision systems that on many tasks are approaching the performance of humans.\ud \ud This work first highlights the importance of ensuring models which are aiming for biological relevance actually produce biologically plausible representations that are consistent with what has been measured within the primate visual cortex. To accomplish this two leading biologically plausible models, HMAX and VisNet are compared on a set of visual processing tasks.\ud \ud The work then changes approach, focusing on models that do not explicitly seek to model any biological process, but rather solve a particular vision task with the goal being increased performance. This section explores the recently discovered problem convolution networks being susceptible to adversarial exemplars. An extension of previous work is shown that allows state-of-the-art networks to be fooled to classify any image as any label while leaving that original image visually unchanged. Secondly an efficient implementation of applying dropout in a batchwise fashion is introduced that approximately halves the computational cost, allowing models twice as large to be trained. Finally an extension to Deep Belief Networks is proposed that constrains the connectivity of the a given layer to that of a topologically local region of the previous one.