Deep learning for computer vision: a comparison between convolutional neural networks and hierarchical temporal memories on object recognition tasks

In recent years, Deep Learning techniques have shown to perform well on a large variety of problems both in Computer Vision and Natural Language Processing, reaching and often surpassing the state of the art on many tasks. The rise of deep learning is also revolutionizing the entire field of Machine Learning and Pattern Recognition pushing forward the concepts of automatic feature extraction and unsupervised learning in general. However, despite the strong success both in science and business, deep learning has its own limitations. It is often questioned if such techniques are only some kind of brute-force statistical approaches and if they can only work in the context of High Performance Computing with tons of data. Another important question is whether they are really biologically inspired, as claimed in certain cases, and if they can scale well in terms of "intelligence". The dissertation is focused on trying to answer these key questions in the context of Computer Vision and, in particular, Object Recognition, a task that has been heavily revolutionized by recent advances in the field. Practically speaking, these answers are based on an exhaustive comparison between two, very different, deep learning techniques on the aforementioned task: Convolutional Neural Network (CNN) and Hierarchical Temporal memory (HTM). They stand for two different approaches and points of view within the big hat of deep learning and are the best choices to understand and point out strengths and weaknesses of each of them. CNN is considered one of the most classic and powerful supervised methods used today in machine learning and pattern recognition, especially in object recognition. CNNs are well received and accepted by the scientific community and are already deployed in large corporation like Google and Facebook for solving face recognition and image auto-tagging problems. HTM, on the other hand, is known as a new emerging paradigm and a new meanly-unsupervised method, that is more biologically inspired. It tries to gain more insights from the computational neuroscience community in order to incorporate concepts like time, context and attention during the learning process which are typical of the human brain. In the end, the thesis is supposed to prove that in certain cases, with a lower quantity of data, HTM can outperform CNN.

Description and Prediction of Freezing of Gait in Parkinson's Disease from Wearable Inertial Measurements Units

Parkinson's disease (PD) is a neuro-degenerative disorder, the second most common after Alzheimer's disease. After diagnosis, treatments can help to relieve the symptoms, but there is no known cure for PD. PD is characterized by a combination of motor and no-motor dysfunctions. Among the motor symptoms there is the so called Freezing of Gait (FoG). The FoG is a phenomenon in PD patients in which the feet stock to the floor and is difficult for the patient to initiate movement. FoG is a severe problem, since it is associated with falls, anxiety, loss of mobility, accidents, mortality and it has substantial clinical and social consequences decreasing the quality of life in PD patients. Medicine can be very successful in controlling movements disorders and dealing with some of the PD symptoms. However, the relationship between medication and the development of FoG remains unclear. Several studies have demonstrated that visual or auditory rhythmical cuing allows PD patients to improve their motor abilities. Rhythmic auditory stimulation (RAS) was shown to be particularly effective at improving gait, specially with patients that manifest FoG. While RAS allows to reduce the time and the effects of FoGs occurrence in PD patients after the FoG is detected, it can not avoid the episode due to the latency of detection. An improvement of the system would be the prediction of the FoG. This thesis was developed following two main objectives: (1) the finding of specifics properties during pre FoG periods different from normal walking context and other walking events like turns and stops using the information provided by the inertial measurements units (IMUs) and (2) the formulation of a model for automatically detect the pre FoG patterns in order to completely avoid the upcoming freezing event in PD patients. The first part focuses on the analysis of different methods for feature extraction which might lead in the FoG occurrence.

Antarctic cloud spectral emission from ground-based measurements, a focus on far infrared signatures

The present work belongs to the PRANA project, the first extensive field campaign of observation of atmospheric emission spectra covering the Far InfraRed spectral region, for more than two years. The principal deployed instrument is REFIR-PAD, a Fourier transform spectrometer used by us to study Antarctic cloud properties. A dataset covering the whole 2013 has been analyzed and, firstly, a selection of good quality spectra is performed, using, as thresholds, radiance values in few chosen spectral regions. These spectra are described in a synthetic way averaging radiances in selected intervals, converting them into BTs and finally considering the differences between each pair of them. A supervised feature selection algorithm is implemented with the purpose to select the features really informative about the presence, the phase and the type of cloud. Hence, training and test sets are collected, by means of Lidar quick-looks. The supervised classification step of the overall monthly datasets is performed using a SVM. On the base of this classification and with the help of Lidar observations, 29 non-precipitating ice cloud case studies are selected. A single spectrum, or at most an average over two or three spectra, is processed by means of the retrieval algorithm RT-RET, exploiting some main IR window channels, in order to extract cloud properties. Retrieved effective radii and optical depths are analyzed, to compare them with literature studies and to evaluate possible seasonal trends. Finally, retrieval output atmospheric profiles are used as inputs for simulations, assuming two different crystal habits, with the aim to examine our ability to reproduce radiances in the FIR. Substantial mis-estimations are found for FIR micro-windows: a high variability is observed in the spectral pattern of simulation deviations from measured spectra and an effort to link these deviations to cloud parameters has been performed.