Job details |
||||||||||||||||||||||||||||||
Job Opportunity: EPSRC DTP PhD Studentship: Using ant biology and natural environments to enhance mo:Job description:Qualification Type: PhD EPSRC DTP PhD Studentship: Using ant biology and natural environments to enhance models of vision and robot navigation Project Over the past decade, Deep neural networks (DNNs) have revolutionisedAI, achieving highlyimpressive perceptual feats especially in the domain of vision. In particular, convolutional neuralnetworks have surpassed humans in their capability to classify naturalistic images. These findings,together with their similarities in architecture and activity patterns, have sparked major renewed interestin DNNs as models of the brain. However, these models diverge from the perceptual abilities ofbiological vision in important and unexpected ways, e.g., through their weakness to noise,susceptibilityto adversarial attacks and incapability to recognise stylised representations. This has been shown to bedue to their over-reliance on diagnostic but brittle features such as texture [1] or even a single pixelcorrelated with the image class [2]. In part, these limitations arise from a lack of inductive biases in themodels to regularise their learning, coupled with unrealistic training environments. Some of theseshortcomings have recently been addressed through incorporating more of the properties of theprimate visual system into these models [3] or by using more broad and naturalistic trainingregimes [4]. The aim of this project is to leverage our expertise of the brain and foraging behaviour ofants to see if both incorporating biological constraintsandnaturalistic training data can generateDNN'swhose perceptual abilities have the robustness and generaliscapability of biological vision. References [1] Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A. & Brendel, W. (2019).ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy androbustness.arXiv:1811.12231. [2] Malhotra, G., Evans, B. D. and Bowers, J. S. (2020) Hiding a plane with a pixel: examining shape-bias in CNNs and the benefit of building in biological constraints.Vision Research 174, pp. 57-68. [3] Evans, B. D. Malhotra, G. & Bowers, J. S. (2021). Biological convolutions improve DNN robustnessto noise and generalisation.Neural Networks 148, pp. 96-110. [4] Mehrer, J., Spoerer, C. J., Jones, E. C., Kriegeskorte, N. & Kietzmann, T. C. (2021).An ecologicallymotivated image dataset for deep learning yields better models of human vision.Proceedings of theNational Academy of Sciences, 118(8). What you get Standard UKRI stipend of £16,062 stipend, research training grant of £1,650 per annum and full tuition fees up to the overseas rate for 3.5 years. Eligibility The stipend and fee waiver is available to: UK / EU / Overseas. Eligible candidates will have a 2:1 degree or equivalent in a related field. Deadline 10 June2022 How to apply Apply online for a full time PhD in Informatics using our step-by-step guide (http://www.sussex.ac.uk/study/phd/apply). Here you will also find details of our entry requirements. Please clearly state on your application form that you are applying for the EPSRC DTP 2022 under the supervision of Dr Benjamin Evans (B.D.Evans@sussex.ac.uk) Contact us For queries related to the admissions process Contact Us by PhD.Informatics@sussex.ac.uk. Timetable Deadline: 10 June2022 Interview Date:TBC Decision:TBC Start Date: 19th September Skills:
|