Since its discovery in 2013, the phenomenon of adversarial examples has attracted a growing amount of attention from the machine learning community. A deeper understanding of the problem could lead to a better comprehension of how information is processed and encoded in neural networks and, more in general, could help to solve the issue of interpretability in machine learning. Our idea to increase adversarial resilience starts with the observation that artificial neurons can be divided in two broad categories: AND-like neurons and OR-like neurons. Intuitively, the former are characterised by a relatively low number of combinations of input values which trigger neuron activation, while for the latter the opposite is true. Our hypothesis is that the presence in a network of a sufficiently high number of OR-like neurons could lead to classification "brittleness" and increase the network's susceptibility to adversarial attacks. After constructing an operational definition of a neuron AND-like behaviour, we proceed to introduce several measures to increase the proportion of AND-like neurons in the network: L1 norm weight normalisation; application of an input filter; comparison between the neuron output's distribution obtained when the network is fed with the actual data set and the distribution obtained when the network is fed with a randomised version of the former called "scrambled data set". Tests performed on the MNIST data set hint that the proposed measures could represent an interesting direction to explore.
The DSM-1 was published in 1952, contains 128 diagnostic categories, described in 132 pages. The DSM-5 appeared in 2013, contains 541 diagnostic categories, described in 947 pages. The field of psychology is characterised by a steady proliferation of diagnostic models and subcategories, that seems to be inspired by the principle of "divide and inflate". This approach is in contrast with experimental evidence, which suggests on one hand that traumas of various kind are often present in the anamnesis of patients and, on the other, that the gene variants implicated are shared across a wide range of diagnoses. In this work I propose a holistic approach, built with tools borrowed from the field of Artificial Intelligence. My model is based on two pillars. The first one is trauma, which represents the attack to the mind, is psychological in nature and has its origin in the environment. The second pillar is dissociation, which represents the mind defence in both physiological and pathological conditions, and incorporates all other defence mechanisms. Damages to dissociation can be considered as another category of attacks, that are neurobiological in nature and can be of genetic or environmental origin. They include, among other factors, synaptic over-pruning, abuse of drugs and inflammation. These factors concur to weaken the defence, represented by the neural networks that implement the dissociation mechanism in the brain. The model is subsequently used to interpret five mental conditions: PTSD, complex PTSD, dissociative identity disorder, schizophrenia and bipolar disorder. Ideally, this is a first step towards building a model that aims to explain a wider range of psychopathological affections with a single theoretical framework. The last part is dedicated to sketching a new psychotherapy for psychological trauma.
Methods to find correlation among variables are of interest to many disciplines, including statistics, machine learning, (big) data mining and neurosciences. Parameters that measure correlation between two variables are of limited utility when used with multiple variables. In this work, I propose a simple criterion to measure correlation among an arbitrary number of variables, based on a data set. The central idea is to i) design a function of the variables that can take different forms depending on a set of parameters, ii) calculate the difference between a statistics associated to the function computed on the data set and the same statistics computed on a randomised version of the data set, called "scrambled" data set, and iii) optimise the parameters to maximise this difference. Many such functions can be organised in layers, which can in turn be stacked one on top of the other, forming a neural network. The function parameters are searched with an enhanced genetic algortihm called POET and the resulting method is tested on a cancer gene data set. The method may have potential implications for some issues that affect the field of neural networks, such as overfitting, the need to process huge amounts of data for training and the presence of "adversarial examples".
Borderline personality disorder and narcissistic personality disorder are important nosographic entities and have been subject of intensive investigations. The currently prevailing psychodynamic theory for mental disorders is based on the repertoire of defense mechanisms employed. Another line of research is concerned with the study of psychological traumas and dissociation as a defensive response. Both theories can be used to shed light on some aspects of pathological mental functioning, and have many points of contact. This work merges these two psychological theories, and builds a model of mental function in a relational context called Quadripolar Relational Model. The model, which is enriched with ideas borrowed from the field of computer science, leads to a new therapeutic proposal for psychological traumas and personality disorders.