Advanced learning strategies for visual understanding

The current success of deep neural networks (DNNs) in image understanding largely depends on the availability of very large training datasets. However, on one hand, this approach is data and computationally expensive, and on the other hand, leaves the user with poor control over what the network has actually
learnt. DCNNs trained in this way may be affected by biases due to their intrinsic tendency towards exploiting “shortcuts”, e.g., features that perform well on the training set but fail to generalize or may have  discriminatory effects. This PhD project focuses on novel strategies to reduce the cost of the training
procedure, while increasing the control of the practitioner on the characteristics of the trained models. To this aim, two complementary research directions will be pursued: the introduction of prior knowledge and stronger inductive biases in the training procedure, and the design of strategies for selecting, combining and scheduling training examples (with and without human in the loop). Results will be assessed on image interpretation tasks, through publicly available datasets, and in realistic settings represented, e.g., by the analysis of images collected from social media. It is expected that results obtained may be of interest for the machine learning community at large.

The PhD student will work on effective training strategies for DNNs focusing on two complementary objectives.

Objective 1: Leveraging prior knowledge The first objective investigates how incorporating prior knowledge (usually coming from linguistic corpora like WordNet or ConceptNet) or logical constraints can improve the training process by: i) compensating the lack of labelled training data, and ii) overcoming the natural tendency of DNNs towards shortcut solutions, by providing high-level inductive biases capable to guide the network towards the desirable solution; as an example, a classifier could be constrained to find a solution that is independent of sex, age or ethnicity, with the aim to counterbalance systematic biases present in many computer vision datasets (e.g., from social media) that rarely disappear by adding more data. This objective will be tackled by exploiting emerging Neural-Symbolic techniques, that merge statistical knowledge representation and reasoning with DNNs letting high-level information be backpropagated to the feature learning phase.

Objective 2: Design of novel training curricula The second objective will target novel strategies for automatically constructing training curricula for DNNs. The mainstream approach to training DNNs involves selecting random mini-batches sampled uniformly from a large labelled training set. This approach disregards the fact that not all training samples contribute equally. Additionally, neural networks, like
humans, have been shown to benefit from learning curricula where difficulty naturally progresses from easier towards more difficult tasks. Nonetheless, the design of a learning curriculum in a principled and automatic fashion is still an active research area. Scenarios with and without human in the loop will be considered.
Interaction between the two research objectives will be closely investigated, by pursuing a mixed curriculum when training can be performed by examples, as well as by shaping the prior knowledge / the logical constraints.

Candidates should have a solid background in machine learning, data analytics or a related field, and strong motivation to learn through advanced research. Expertise in deep learning, demonstrated through prior projects or publications, is preferred.