Learning Nonlinear Visual Processing from Natural Images

The paradigm of computational vision hypothesizes that any visual function -- such as the recognition of your grandparent -- can be replicated by computational processing of the visual input. What are these computations that the brain performs? What should or could they be? Working on the latter question, this dissertation takes the statistical approach, where the suitable computations are attempted to be learned from the natural visual data itself. In particular, we empirically study the computational processing that emerges from the statistical properties of the visual world and the constraints and objectives specified for the learning process. This thesis consists of an introduction and 7 peer-reviewed publications, where the purpose of the introduction is to illustrate the area of study to a reader who is not familiar with computational vision research. In the scope of the introduction, we will briefly overview the primary challenges to visual processing, as well as recall some of the current opinions on visual processing in the early visual systems of animals. Next, we describe the methodology we have used in our research, and discuss the presented results. We have included some additional remarks, speculations and conclusions to this discussion that were not featured in the original publications. We present the following results in the publications of this thesis. First, we empirically demonstrate that luminance and contrast are strongly dependent in natural images, contradicting previous theories suggesting that luminance and contrast were processed separately in natural systems due to their independence in the visual data. Second, we show that simple cell -like receptive fields of the primary visual cortex can be learned in the nonlinear contrast domain by maximization of independence. Further, we provide first-time reports of the emergence of conjunctive (corner-detecting) and subtractive (opponent orientation) processing due to nonlinear projection pursuit with simple objective functions related to sparseness and response energy optimization. Then, we show that attempting to extract independent components of nonlinear histogram statistics of a biologically plausible representation leads to projection directions that appear to differentiate between visual contexts. Such processing might be applicable for priming, \ie the selection and tuning of later visual processing. We continue by showing that a different kind of thresholded low-frequency priming can be learned and used to make object detection faster with little loss in accuracy. Finally, we show that in a computational object detection setting, nonlinearly gain-controlled visual features of medium complexity can be acquired sequentially as images are encountered and discarded. We present two online algorithms to perform this feature selection, and propose the idea that for artificial systems, some processing mechanisms could be selectable from the environment without optimizing the mechanisms themselves. In summary, this thesis explores learning visual processing on several levels. The learning can be understood as interplay of input data, model structures, learning objectives, and estimation algorithms. The presented work adds to the growing body of evidence showing that statistical methods can be used to acquire intuitively meaningful visual processing mechanisms. The work also presents some predictions and ideas regarding biological visual processing.

Mechanisms for alphavirus nonstructural polyprotein processing

Replication and transcription of the RNA genome of alphaviruses relies on a set of virus-encoded nonstructural proteins. They are synthesized as a long polyprotein precursor, P1234, which is cleaved at three processing sites to yield nonstructural proteins nsP1, nsP2, nsP3 and nsP4. All the four proteins function as constitutive components of the membrane-associated viral replicase. Proteolytic processing of P1234 polyprotein is precisely orchestrated and coordinates the replicase assembly and maturation. The specificity of the replicase is also controlled by proteolytic cleavages. The early replicase is composed of P123 polyprotein intermediate and nsP4. It copies the positive sense RNA genome to complementary minus-strand. Production of new plus-strands requires complete processing of the replicase. The papain-like protease residing in nsP2 is responsible for all three cleavages in P1234. This study addressed the mechanisms of proteolytic processing of the replicase polyprotein in two alphaviruses Semliki Forest virus (SFV) and Sindbis virus (SIN) representing different branches of the genus. The survey highlighted the functional relation of the alphavirus nsP2 protease to the papain-like enzymes. A new structural motif the Cys-His catalytic dyad accompanied with an aromatic residue following the catalytic His was described for nsP2 and a subset of other thiol proteases. Such an architecture of the catalytic center was named the glycine specificity motif since it was implicated in recognition of a specific Gly residue in the substrate. In particular, the presence of the motif in nsP2 makes the appearance of this amino acid at the second position upstream of the scissile bond a necessary condition for the cleavage. On top of that, there were four distinct mechanisms identified, which provide affinity for the protease and specifically direct the enzyme to different sites in the P1234 polyprotein. Three factors RNA, the central domain of nsP3 and the N-terminus of nsP2 were demonstrated to be external modulators of the nsP2 protease. Here I suggest that the basal nsP2 protease specificity is inherited from the ancestral papain-like enzyme and employs the recognition of the upstream amino acid signature in the immediate vicinity of the scissile bond. This mechanism is responsible for the efficient processing of the SFV nsP3/nsP4 junction. I propose that the same mechanism is involved in the cleavage of the nsP1/nsP2 junction of both viruses as well. However, in this case it rather serves to position the substrate, whereas the efficiency of the processing is ensured by the capability of nsP2 to cut its own N-terminus in cis. Both types of cleavages are demonstrated here to be inhibited by RNA, which is interpreted as impairing the basal papain-like recognition of the substrate. In contrast, processing of the SIN nsP3/nsP4 junction was found to be activated by RNA and additionally potentiated by the presence of the central region of nsP3 in the protease. The processing of the nsP2/nsP3 junction in both viruses occurred via another mechanism, requiring the exactly processed N-terminus of nsP2 in the protease and insensitive to RNA addition. Therefore, the three processing events in the replicase polyprotein maturation are performed via three distinct mechanisms in each of two studied alphaviruses. Distinct sets of conditions required for each cleavage ensure sequential maturation of P1234 polyprotein: nsP4 is released first, then the nsP1/nsP2 site is cut in cis, and liberation of the nsP2 N-terminus activates the cleavage of the nsP2/nsP3 junction at last. The first processing event occurs differently in SFV and SIN, whereas the subsequent cleavages are found to be similar in the two viruses and therefore, their mechanisms are suggested to be conserved in the genus. The RNA modulation of the alphavirus nonstructural protease activity, discovered here, implies bidirectional functional interplay between the alphavirus RNA metabolism and protease regulation. The nsP2 protease emerges as a signal transmitting moiety, which senses the replication stage and responds with proteolytic cleavages. A detailed hypothetical model of the alphavirus replicase core was inferred from the data obtained in the study. Similar principles in replicase organization and protease functioning are expected to be employed by other RNA viruses.

Tactile processing in human somatosensory and auditory cortices

Tactile sensation plays an important role in everyday life. While the somatosensory system has been studied extensively, the majority of information has come from studies using animal models. Recent development of high-resolution anatomical and functional imaging techniques has enabled the non-invasive study of human somatosensory cortex and thalamus. This thesis provides new insights into the functional organization of the human brain areas involved in tactile processing using magnetoencephalography (MEG) and functional magnetic resonance imaging (fMRI). The thesis also demonstrates certain optimizations of MEG and fMRI methods. Tactile digit stimulation elicited stimulus-specific responses in a number of brain areas. Contralateral activation was observed in somatosensory thalamus (Study II), primary somatosensory cortex (SI; I, III, IV), and post-auditory belt area (III). Bilateral activation was observed in secondary somatosensory cortex (SII; II, III, IV). Ipsilateral activation was found in the post-central gyrus (area 2 of SI cortex; IV). In addition, phasic deactivation was observed within ipsilateral SI cortex and bilateral primary motor cortex (IV). Detailed investigation of the tactile responses demonstrated that the arrangement of distal-proximal finger representations in area 3b of SI in humans is similar to that found in monkeys (I). An optimized MEG approach was sufficient to resolve such fine detail in functional organization. The SII region appeared to contain double representations for fingers and toes (II). The detection of activations in the SII region and thalamus improved at the individual and group levels when cardiac-gated fMRI was used (II). Better detection of body part representations at the individual level is an important improvement, because identification of individual representations is crucial for studying brain plasticity in somatosensory areas. The posterior auditory belt area demonstrated responses to both auditory and tactile stimuli (III), implicating this area as a physiological substrate for the auditory-tactile interaction observed in earlier psychophysical studies. Comparison of different smoothing parameters (III) demonstrated that proper evaluation of co-activation should be based on individual subject analysis with minimal or no smoothing. Tactile input consistently influenced area 3b of the human ipsilateral SI cortex (IV). The observed phasic negative fMRI response is proposed to result from interhemispheric inhibition via trans-callosal connections. This thesis contributes to a growing body of human data suggesting that processing of tactile stimuli involves multiple brain areas, with different spatial patterns of cortical activation for different stimuli.

Development of a new flexor tendon repair technique performed with bioabsorbable poly-L/D-lactide (PLDLA) suture

Sormen koukistajajännevamman korjauksen jälkeisen aktiivisen mobilisaation on todettu johtavan parempaan toiminnalliseen lopputulokseen kuin nykyisin yleisesti käytetyn dynaamisen mobilisaation. Aktiivisen mobilisaation ongelma on jännekorjauksen pettämisriskin lisääntyminen nykyisten ommeltekniikoiden riittämättömän vahvuuden vuoksi. Jännekorjauksen lujuutta on parannettu kehittämällä monisäieommeltekniikoita, joissa jänteeseen tehdään useita rinnakkaisia ydinompeleita. Niiden kliinistä käyttöä rajoittaa kuitenkin monimutkainen ja aikaa vievä tekninen suoritus. Käden koukistajajännekorjauksessa käytetään yleisesti sulamattomia ommelmateriaaleja. Nykyiset käytössä olevat biohajoavat langat heikkenevät liian nopeasti jänteen paranemiseen nähden. Biohajoavan laktidistereokopolymeeri (PLDLA) 96/4 – langan vetolujuuden puoliintumisajan sekä kudosominaisuuksien on aiemmin todettu soveltuvan koukistajajännekorjaukseen. Tutkimuksen tavoitteena oli kehittää välittömän aktiivisen mobilisaation kestävä ja toteutukseltaan yksinkertainen käden koukistajajännekorjausmenetelmä biohajoavaa PLDLA 96/4 –materiaalia käyttäen. Tutkimuksessa analysoitiin viiden eri yleisesti käytetyn koukistajajänneompeleen biomekaanisia ominaisuuksia staattisessa vetolujuustestauksessa ydinompeleen rakenteellisten ominaisuuksien – 1) säikeiden (lankojen) lukumäärän, 2) langan paksuuden ja 3) ompeleen konfiguraation – vaikutuksen selvittämiseksi jännekorjauksen pettämiseen ja vahvuuteen. Jännekorjausten näkyvän avautumisen todettiin alkavan perifeerisen ompeleen pettäessä voima-venymäkäyrän myötöpisteessä. Ydinompeleen lankojen lukumäärän lisääminen paransi ompeleen pitokykyä jänteessä ja suurensi korjauksen myötövoimaa. Sen sijaan paksumman (vahvemman) langan käyttäminen tai ompeleen konfiguraatio eivät vaikuttaneet myötövoimaan. Tulosten perusteella tutkittiin mahdollisuuksia lisätä ompeleen pitokykyä jänteestä yksinkertaisella monisäieompeleella, jossa ydinommel tehtiin kolmen säikeen polyesterilangalla tai nauhamaisen rakenteen omaavalla kolmen säikeen polyesterilangalla. Nauhamainen rakenne lisäsi merkitsevästi ompeleen pitokykyä jänteessä parantaen myötövoimaa sekä maksimivoimaa. Korjauksen vahvuus ylitti aktiivisen mobilisaation jännekorjaukseen kohdistaman kuormitustason. PLDLA 96/4 –langan soveltuvuutta koukistajajännekorjaukseen selvitettiin tutkimalla langan biomekaanisia ominaisuuksia ja solmujen pito-ominaisuuksia staattisessa vetolujuustestauksessa verrattuna yleisimmin jännekorjauksessa käytettävään punottuun polyesterilankaan (Ticron®). PLDLA –langan todettiin soveltuvan hyvin koukistajajännekorjaukseen, sillä se on polyesterilankaa venymättömämpi ja solmujen pitävyys on parempi. Viimeisessä vaiheessa tutkittiin PLDLA 96/4 –langasta valmistetulla kolmisäikeisellä, nauhamaisella jännekorjausvälineellä tehdyn jännekorjauksen kestävyyttä staattisessa vetolujuustestauksessa sekä syklisessä kuormituksessa, joka simuloi staattista testausta paremmin mobilisaation toistuvaa kuormitusta. PLDLA-korjauksen vahvuus ylitti sekä staattisessa että syklisessä kuormituksessa aktiivisen mobilisaation edellyttämän vahvuuden. Nauhamaista litteää ommelmateriaalia ei aiemmin ole tutkittu tai käytetty käden koukistajajännekorjauksessa. Tässä tutkimuksessa ommelmateriaalin nauhamainen rakenne paransi merkitsevästi jännekorjauksen vahvuutta, minkä arvioidaan johtuvan lisääntyneestä kontaktipinnasta jänteen ja ommelmateriaalin välillä estäen ompeleen läpileikkautumista jänteessä. Tutkimuksessa biohajoavasta PLDLA –materiaalista valmistetulla rakenteeltaan nauhamaisella kolmisäikeisellä langalla tehdyn jännekorjauksen vahvuus saavutti aktiivisen mobilisaation edellyttämän tason. Lisäksi uusi menetelmä on helppokäyttöinen ja sillä vältetään perinteisten monisäieompeleiden tekniseen suoritukseen liittyvät ongelmat.

Diffusion of Be, Co and Mn impurities in compound semiconductors and glassy carbon studied by the modified radiotracer technique

The main method of modifying properties of semiconductors is to introduce small amount of impurities inside the material. This is used to control magnetic and optical properties of materials and to realize p- and n-type semiconductors out of intrinsic material in order to manufacture fundamental components such as diodes. As diffusion can be described as random mixing of material due to thermal movement of atoms, it is essential to know the diffusion behavior of the impurities in order to manufacture working components. In modified radiotracer technique diffusion is studied using radioactive isotopes of elements as tracers. The technique is called modified as atoms are deployed inside the material by ion beam implantation. With ion implantation, a distinct distribution of impurities can be deployed inside the sample surface with good con- trol over the amount of implanted atoms. As electromagnetic radiation and other nuclear decay products emitted by radioactive materials can be easily detected, only very low amount of impurities can be used. This makes it possible to study diffusion in pure materials without essentially modifying the initial properties by doping. In this thesis a modified radiotracer technique is used to study the diffusion of beryllium in GaN, ZnO, SiGe and glassy carbon. GaN, ZnO and SiGe are of great interest to the semiconductor industry and beryllium as a small and possibly rapid dopant hasn t been studied previously using the technique. Glassy carbon has been added to demonstrate the feasibility of the technique. In addition, the diffusion of magnetic impurities, Mn and Co, has been studied in GaAs and ZnO (respectively) with spintronic applications in mind.

Measurement of the WW+WZ production cross section using a matrix element technique in lepton+jets events

We present a measurement of the $WW+WZ$ production cross section observed in a final state consisting of an identified electron or muon, two jets, and missing transverse energy. The measurement is carried out in a data sample corresponding to up to 4.6~fb$^{-1}$ of integrated luminosity at $\sqrt{s} = 1.96$ TeV collected by the CDF II detector. Matrix element calculations are used to separate the diboson signal from the large backgrounds. The $WW+WZ$ cross section is measured to be $17.4\pm3.3$~pb, in agreement with standard model predictions. A fit to the dijet invariant mass spectrum yields a compatible cross section measurement.

Top quark mass measurement in the tt̅ all hadronic channel using a matrix element technique in pp̅ collisions at √s=1.96 TeV

We present a measurement of the top quark mass in the all-hadronic channel (\tt $\to$ \bb$q_{1}\bar{q_{2}}q_{3}\bar{q_{4}}$) using 943 pb$^{-1}$ of \ppbar collisions at $\sqrt {s} = 1.96$ TeV collected at the CDF II detector at Fermilab (CDF). We apply the standard model production and decay matrix-element (ME) to $\ttbar$ candidate events. We calculate per-event probability densities according to the ME calculation and construct template models of signal and background. The scale of the jet energy is calibrated using additional templates formed with the invariant mass of pairs of jets. These templates form an overall likelihood function that depends on the top quark mass and on the jet energy scale (JES). We estimate both by maximizing this function. Given 72 observed events, we measure a top quark mass of 171.1 $\pm$ 3.7 (stat.+JES) $\pm$ 2.1 (syst.) GeV/$c^{2}$. The combined uncertainty on the top quark mass is 4.3 GeV/$c^{2}$.

Two bracketing schemes for the Penn Treebank

The trees in the Penn Treebank have a standard representation that involves complete balanced bracketing. In this article, an alternative for this standard representation of the tree bank is proposed. The proposed representation for the trees is loss-less, but it reduces the total number of brackets by 28%. This is possible by omitting the redundant pairs of special brackets that encode initial and final embedding, using a technique proposed by Krauwer and des Tombe (1981). In terms of the paired brackets, the maximum nesting depth in sentences decreases by 78%. The 99.9% coverage is achieved with only five non-top levels of paired brackets. The observed shallowness of the reduced bracketing suggests that finite-state based methods for parsing and searching could be a feasible option for tree bank processing.