PhD Program in Computer and Control Engineering
PhD Student: Francesco Manigrasso
Context of the research activity
The successes of deep neural networks (DNNs) in the analysis of multi-dimensional data is largely due to their ability in modelling high-dimensional objects (such as pixels in an image) by learning non-linear feature representations, and the availability of simple and scalable supervised learning algorithms. Nonetheless, DNNs are often surprisingly brittle in practice and lack desirable properties including top-down control, transparency and robustness.
There is an increasing interest in complementing representation learning with techniques borrowed from statistical relational reasoning and knowledge representation. Symbolic knowledge representation excels precisely where statistical machine learning is weakest, in terms of transparency, ability to perform inference and reasoning, representation of high-level concepts, integration of prior knowledge and, in general, model of the external worlds.
Recent advances in computationally efficient statistical inference have also significantly expanded the toolbox of probabilistic modeling techniques that can be applied to deep learning models. This has important practical implications in terms of robustness. One of the most profound effects is, for instance, the ability to “know when you don’t know”.
The PhD candidate will study techniques that incorporate deep learning and probabilistic reasoning for the interpretation of multi-dimensional data, such as images.
The overall objective of this research proposal is to explore novel ways to bridge the gap between deep representation learning and symbolic knowledge representation in multi-dimensional data analysis, leveraging recent advances in the field of neural-symbolic integration and probabilistic modeling.
Specifically, this research proposal targets the area of neural-symbolic integration, which can be used to encode symbolic representation techniques, such as fuzzy and/or probabilistic logic, as tensors in a deep neural network. This would allow to extend current DNNs feature learning capabilities in many ways, for instance by imposing a priori information.
Many recent techniques have been proposed to combine deep learning with reasoning, but the applications are still limited compared to their potential. In many of the proposed approaches, either the feature representation or the reasoning layer is frozen and cannot be easily learned, and/or feature extraction and reasoning are disjoint and cannot be trained end-to-end. Probabilistic reasoning is also required to handle uncertainty and exceptions in real-life applications. It allows to quantify uncertainty in the predictions and thus, enable the development of robust machine learning models. Finally, application in large-scale datasets and a variety of tasks beyond classification is still limited.
The candidate will target the application of neural-symbolic integration techniques to solve problems in different image analysis tasks, such as image-level classification, segmentation and object detection. Of particular interest is the possibility to explore the integration of image level-data with other sources of information, for instance the integration of prior (possibly causal) information.
There are a number of theoretical and practical issues to overcome to apply this approach at scale. First, techniques are needed that can achieve both representation learning and relational reasoning, and that can be trained in an end-to-end fashion at scale. Secondly, the above-mentioned goals must be achieved in a probabilistic framework, leveraging recent advances in statistical inference and optimization. During the PhD, new solutions will be studied and developed to address the issues listed above, and finally compared with state-of-the-art deep learning approaches to assess their potential on machine learning benchmarks.
The proposed techniques will also be applied in at least one selected case study, such as medical image analysis and its integration with other clinical data, or problems related to autonomous driving. To this aim, the PhD candidate will benefit from existing collaborations with industrial and clinical partners, as well as within the SmartData@PoliTO research center.
Skills and competencies for the development of the activity
The candidate should have Competences/Experience in one or more of these fields:
- Machine learning
- Deep learning
- Artificial intelligence