Deep learning architecture for gesture recognition

Pós-graduação em Engenharia Mecânica - FEG

The Cuticular Hydrocarbons Profiles in the Colonial Recognition of the Neotropical Eusocial Wasp, Mischocyttarus cassununga (Hymenoptera: Vespidae)

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES)

The growth of glioblastoma orthotopic xenografts in nude mice is directly correlated with impaired object recognition memory

Cognitive dysfunction is found in patients with brain tumors and there is a need to determine whether it can be replicated in an experimental model. In the present study, the object recognition (OR) paradigm was used to investigate cognitive performance in nude mice, which represent one of the most important animal models available to study human tumors in vivo. Mice with orthotopic xenografts of the human U87MG glioblastoma cell line were trained at 9, 14, and 18days (D9, D14, and D18, respectively) after implantation of 5×10(5) cells. At D9, the mice showed normal behavior when tested 90min or 24h after training and compared to control nude mice. Animals at D14 were still able to discriminate between familiar and novel objects, but exhibited a lower performance than animals at D9. Total impairment in the OR memory was observed when animals were evaluated on D18. These alterations were detected earlier than any other clinical symptoms, which were observed only 22-24days after tumor implantation. There was a significant correlation between the discrimination index (d2) and time after tumor implantation as well as between d2 and tumor volume. These data indicate that the OR task is a robust test to identify early behavior alterations caused by glioblastoma in nude mice. In addition, these results suggest that OR task can be a reliable tool to test the efficacy of new therapies against these tumors.

A performance based study on gender recognition in large datasets

[EN]Gender recognition has achieved impressive results based on the face appearance in controlled datasets. Its application in the wild and large datasets is still a challenging task for researchers. In this paper, we make use of classical techniques to analyze their performance in controlled and uncontrolled condition respectively with the LFW and MORPH datasets. For both sets the benchmarking protocol follows the 5-fold cross-validation proposed by the BEFIT challenge.

Exploring the use of local descriptors for fish recognition in LifeCLEF 2015

[EN]This paper summarizes the proposal made by the SIANI team for the LifeCLEF 2015 Fish task. The approach makes use of standard detection techniques, applying a multiclass SVM based classifier on large enough Regions Of Interest (ROIs) automatically extracted from the provided video frames. The selection of the detection and classification modules is based on the best performance achieved for the validation dataset consisting of 20 annotated videos. For that dataset, the best classification achieved for an ideal detection module, reaches an accuracy around 40%.

On the Design of a Boolean-network Robot Swarm: Collective Recognition of Ground Patterns.

Uno dei principali ambiti di ricerca dell’intelligenza artificiale concerne la realizzazione di agenti (in particolare, robot) in grado di aiutare o sostituire l’uomo nell’esecuzione di determinate attività. A tal fine, è possibile procedere seguendo due diversi metodi di progettazione: la progettazione manuale e la progettazione automatica. Quest’ultima può essere preferita alla prima nei contesti in cui occorra tenere in considerazione requisiti quali flessibilità e adattamento, spesso essenziali per lo svolgimento di compiti non banali in contesti reali. La progettazione automatica prende in considerazione un modello col quale rappresentare il comportamento dell’agente e una tecnica di ricerca (oppure di apprendimento) che iterativamente modifica il modello al fine di renderlo il più adatto possibile al compito in esame. In questo lavoro, il modello utilizzato per la rappresentazione del comportamento del robot è una rete booleana (Boolean network o Kauffman network). La scelta di tale modello deriva dal fatto che possiede una semplice struttura che rende agevolmente studiabili le dinamiche tuttavia complesse che si manifestano al suo interno. Inoltre, la letteratura recente mostra che i modelli a rete, quali ad esempio le reti neuronali artificiali, si sono dimostrati efficaci nella programmazione di robot. La metodologia per l’evoluzione di tale modello riguarda l’uso di tecniche di ricerca meta-euristiche in grado di trovare buone soluzioni in tempi contenuti, nonostante i grandi spazi di ricerca. Lavori precedenti hanno gia dimostrato l’applicabilità e investigato la metodologia su un singolo robot. Lo scopo di questo lavoro è quello di fornire prova di principio relativa a un insieme di robot, aprendo nuove strade per la progettazione in swarm robotics. In questo scenario, semplici agenti autonomi, interagendo fra loro, portano all’emergere di un comportamento coordinato adempiendo a task impossibili per la singola unità. Questo lavoro fornisce utili ed interessanti opportunità anche per lo studio delle interazioni fra reti booleane. Infatti, ogni robot è controllato da una rete booleana che determina l’output in funzione della propria configurazione interna ma anche dagli input ricevuti dai robot vicini. In questo lavoro definiamo un task in cui lo swarm deve discriminare due diversi pattern sul pavimento dell’arena utilizzando solo informazioni scambiate localmente. Dopo una prima serie di esperimenti preliminari che hanno permesso di identificare i parametri e il migliore algoritmo di ricerca, abbiamo semplificato l’istanza del problema per meglio investigare i criteri che possono influire sulle prestazioni. E’ stata così identificata una particolare combinazione di informazione che, scambiata localmente fra robot, porta al miglioramento delle prestazioni. L’ipotesi è stata confermata applicando successivamente questo risultato ad un’istanza più difficile del problema. Il lavoro si conclude suggerendo nuovi strumenti per lo studio dei fenomeni emergenti in contesti in cui le reti booleane interagiscono fra loro.

Role of nitric oxide and endocannabinoids in synaptic plasticity in the perirhinal cortex and in visual recognition memory

Introduction and aims of the research Nitric oxide (NO) and endocannabinoids (eCBs) are major retrograde messengers, involved in synaptic plasticity (long-term potentiation, LTP, and long-term depression, LTD) in many brain areas (including hippocampus and neocortex), as well as in learning and memory processes. NO is synthesized by NO synthase (NOS) in response to increased cytosolic Ca2+ and mainly exerts its functions through soluble guanylate cyclase (sGC) and cGMP production. The main target of cGMP is the cGMP-dependent protein kinase (PKG). Activity-dependent release of eCBs in the CNS leads to the activation of the Gαi/o-coupled cannabinoid receptor 1 (CB1) at both glutamatergic and inhibitory synapses. The perirhinal cortex (Prh) is a multimodal associative cortex of the temporal lobe, critically involved in visual recognition memory. LTD is proposed to be the cellular correlate underlying this form of memory. Cholinergic neurotransmission has been shown to play a critical role in both visual recognition memory and LTD in Prh. Moreover, visual recognition memory is one of the main cognitive functions impaired in the early stages of Alzheimer’s disease. The main aim of my research was to investigate the role of NO and ECBs in synaptic plasticity in rat Prh and in visual recognition memory. Part of this research was dedicated to the study of synaptic transmission and plasticity in a murine model (Tg2576) of Alzheimer’s disease. Methods Field potential recordings. Extracellular field potential recordings were carried out in horizontal Prh slices from Sprague-Dawley or Dark Agouti juvenile (p21-35) rats. LTD was induced with a single train of 3000 pulses delivered at 5 Hz (10 min), or via bath application of carbachol (Cch; 50 μM) for 10 min. LTP was induced by theta-burst stimulation (TBS). In addition, input/output curves and 5Hz-LTD were carried out in Prh slices from 3 month-old Tg2576 mice and littermate controls. Behavioural experiments. The spontaneous novel object exploration task was performed in intra-Prh bilaterally cannulated adult Dark Agouti rats. Drugs or vehicle (saline) were directly infused into the Prh 15 min before training to verify the role of nNOS and CB1 in visual recognition memory acquisition. Object recognition memory was tested at 20 min and 24h after the end of the training phase. Results Electrophysiological experiments in Prh slices from juvenile rats showed that 5Hz-LTD is due to the activation of the NOS/sGC/PKG pathway, whereas Cch-LTD relies on NOS/sGC but not PKG activation. By contrast, NO does not appear to be involved in LTP in this preparation. Furthermore, I found that eCBs are involved in LTP induction, but not in basal synaptic transmission, 5Hz-LTD and Cch-LTD. Behavioural experiments demonstrated that the blockade of nNOS impairs rat visual recognition memory tested at 24 hours, but not at 20 min; however, the blockade of CB1 did not affect visual recognition memory acquisition tested at both time points specified. In three month-old Tg2576 mice, deficits in basal synaptic transmission and 5Hz-LTD were observed compared to littermate controls. Conclusions The results obtained in Prh slices from juvenile rats indicate that NO and CB1 play a role in the induction of LTD and LTP, respectively. These results are confirmed by the observation that nNOS, but not CB1, is involved in visual recognition memory acquisition. The preliminary results obtained in the murine model of Alzheimer’s disease indicate that deficits in synaptic transmission and plasticity occur very early in Prh; further investigations are required to characterize the molecular mechanisms underlying these deficits.

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.

Melody recognition at fast and slow tempos: Effects of age, experience, and familiarity

Eighty-one listeners defined by three age ranges (18–30, 31–59, and over 60 years) and three levels of musical experience performed an immediate recognition task requiring the detection of alterations in melodies. On each trial, a brief melody was presented, followed 5 sec later by a test stimulus that either was identical to the target or had two pitches changed, for a same–different judgment. Each melody pair was presented at 0.6 note/sec, 3.0 notes/sec, or 6.0 notes/sec. Performance was better with familiar melodies than with unfamiliar melodies. Overall performance declined slightly with age and improved substantially with increasing experience, in agreement with earlier results in an identification task. Tempo affected performance on familiar tunes (moderate was best), but not on unfamiliar tunes. We discuss these results in terms of theories of dynamic attending, cognitive slowing, and working memory in aging.

Levels-of-processing effects on "remember" responses in recognition for familiar and unfamiliar tunes

We investigated the effect of level-of-processing manipulations on “remember” and “know” responses in episodic melody recognition (Experiments 1 and 2) and how this effect is modulated by item familiarity (Experiment 2). In Experiment 1, participants performed 2 conceptual and 2 perceptual orienting tasks while listening to familiar melodies: judging the mood, continuing the tune, tracing the pitch contour, and counting long notes. The conceptual mood task led to higher d' rates for “remember” but not “know” responses. In Experiment 2, participants either judged the mood or counted long notes of tunes with high and low familiarity. A level-of-processing effect emerged again in participants’ “remember” d' rates regardless of melody familiarity. Results are discussed within the distinctive processing framework.

Aging and experience in the recognition of musical transpositions

The authors examined the effects of age, musical experience, and characteristics of musical stimuli on a melodic short-term memory task in which participants had to recognize whether a tune was an exact transposition of another tune recently presented. Participants were musicians and nonmusicians between ages 18 and 30 or 60 and 80. In 4 experiments, the authors found that age and experience affected different aspects of the task, with experience becoming more influential when interference was provided during the task. Age and experience interacted only weakly, and neither age nor experience influenced the superiority of tonal over atonal materials. Recognition memory for the sequences did not reflect the same pattern of results as the transposition task. The implications of these results for theories of aging, experience, and music cognition are discussed.

The Role of Features and Context in Recognition of Novel Melodies

WE INVESTIGATED HOW WELL STRUCTURAL FEATURES such as note density or the relative number of changes in the melodic contour could predict success in implicit and explicit memory for unfamiliar melodies. We also analyzed which features are more likely to elicit increasingly confident judgments of "old" in a recognition memory task. An automated analysis program computed structural aspects of melodies, both independent of any context, and also with reference to the other melodies in the testset and the parent corpus of pop music. A few features predicted success in both memory tasks, which points to a shared memory component. However, motivic complexity compared to a large corpus of pop music had different effects on explicit and implicit memory. We also found that just a few features are associated with different rates of "old" judgments, whether the items were old or new. Rarer motives relative to the testset predicted hits and rarer motives relative to the corpus predicted false alarms. This data-driven analysis provides further support for both shared and separable mechanisms in implicit and explicit memory retrieval, as well as the role of distinctiveness in true and false judgments of familiarity.

The development and evaluation of a computer-based program to test and to teach the recognition of facial affect

Autism is a chronic pervasive neurodevelopmental disorder characterized by the early onset of social and communicative impairments as well as restricted, ritualized, stereotypic behavior. The endophenotype of autism includes neuropsychological deficits, for instance a lack of "Theory of Mind" and problems recognizing facial affect. In this study, we report the development and evaluation of a computer-based program to teach and test the ability to identify basic facially expressed emotions. 10 adolescent or adult subjects with high-functioning autism or Asperger-syndrome were included in the investigation. A priori the facial affect recognition test had shown good psychometric properties in a normative sample (internal consistency: rtt=.91-.95; retest reliability: rtt=.89-.92). In a prepost design, one half of the sample was randomly assigned to receive computer treatment while the other half of the sample served as control group. The training was conducted for five weeks, consisting of two hours training a week. The trained individuals improved significantly on the affect recognition task, but not on any other measure. Results support the usefulness of the program to teach the detection of facial affect. However, the improvement found is limited to a circumscribed area of social-communicative function and generalization is not ensured.

Evaluating retraining rules for semi-supervised learning in neural network based cursive word recognition

Training a system to recognize handwritten words is a task that requires a large amount of data with their correct transcription. However, the creation of such a training set, including the generation of the ground truth, is tedious and costly. One way of reducing the high cost of labeled training data acquisition is to exploit unlabeled data, which can be gathered easily. Making use of both labeled and unlabeled data is known as semi-supervised learning. One of the most general versions of semi-supervised learning is self-training, where a recognizer iteratively retrains itself on its own output on new, unlabeled data. In this paper we propose to apply semi-supervised learning, and in particular self-training, to the problem of cursive, handwritten word recognition. The special focus of the paper is on retraining rules that define what data are actually being used in the retraining phase. In a series of experiments it is shown that the performance of a neural network based recognizer can be significantly improved through the use of unlabeled data and self-training if appropriate retraining rules are applied.

Changes in emotion recognition relate to hypertension status

Background: Emotional processing in essential hypertension beyond self-report questionnaire has hardly been investigated. The aim of this study is to examine associations between hypertension status and recognition of facial affect. Methods: 25 healthy, non-smoking, medication-free men including 13 hypertensive subjects aged between 20 and 65 years completed a computer-based task in order to examine sensitivity of recognition of facial affect. Neutral faces gradually changed to a specific emotion in a pseudo-continuous manner. Slides of the six basic emotions (fear, sadness, disgust, happiness, anger, surprise) were chosen from the „NimStim Set“. Pictures of three female and three male faces were electronically morphed in 1% steps of intensity from 0% to 100% (36 sets of faces with 100 pictures each). Each picture of a set was presented for one second, ranging from 0% to 100% of intensity. Participants were instructed to press a stop button as soon as they recognized the expression of the face. After stopping a forced choice between the six basic emotions was required. As dependent variables, we recorded the emotion intensity at which the presentation was stopped and the number of errors (error rate). Recognition sensitivity was calculated as emotion intensity of correctly identified emotions. Results: Mean arterial pressure was associated with a significantly increased recognition sensitivity of facial affect for the emotion anger (ß = - .43, p = 0.03*, Δ R2= .110). There was no association with the emotions fear, sadness, disgust, happiness, and surprise (p’s > .0.41). Mean arterial pressure did not relate to the mean number of errors for any of the facial emotions. Conclusions: Our findings suggest that an increased blood pressure is associated with increased recognition sensitivity of facial affect for the emotion anger, if a face shows anger. Hypertensives perceive facial anger expression faster than normotensives, if anger is shown.