Search

Transfer Learning is a technique that reduces the time taken for training and improves performance by reusing the weights of a previously trained source neural network. This poses a question of how the source network must be chosen, which is still an unsolved problem. In this work, we have built a framework that automatically performs transfer learning by selecting the source neural network based on an estimate of dataset classification difficulty. The framework designates the neural network of the dataset that is closest in difficulty as the source. The framework is evaluated on 7 datasets namely SVHN, Cifar10, Cifar100, GTSRB, MNIST, Flowers and Linnaeus5, and experimental results suggest that in most cases this type of source selection gives the highest improvement in accuracy. The framework provides an average improvement in accuracy of 6.7% for ResNet than when training from scratch, and achieves it within the first 10 epochs in some cases.

The availability and widespread use of social media has made it the preferred medium for companies wanting to spread product information, create public opinion and gain followers. To this end, social media influencers act as a dynamic third-party endorser to spread a brand's message to vast audience across the world. Consumers who exhibit a positive attitude towards the social media credibility are attracted through social media advertisements. The present study examines the perceptions of agencies for opting social media influencers and their role in customer engagement and brand awareness. For this purpose, the study also assesses different methods adopted by these influencers for influencing customers using a systematic review. The findings of the study indicate customer's perception and attitude are much influenced via these influencers since they are more capable of communicating to a niche segment. As compared to traditional advertising strategies, this new technological means of influencers pave way to new competitive strength to the agencies in engaging customers and creating brand awareness.

Approximately one-third of epileptic patients have inadequate seizure control through anticonvulsant medications. For these patients the unforeseeable way in which seizures occur represents the most disabling and dangerous aspect of their disease. Even if seizures seem to occur suddenly and without warning in most patients, there has been a longstanding discussion on whether seizures may be building up slowly in the hours or minutes before the clinical event (Le Van Quyen et al., 1999; Lehnertz et al., 1999; Litt et al., 2001). Some patients describe unspecific premonitory symptoms hours before the seizure (Schulze-Bonhage et al., 2006; Haut et al., 2007), and several studies detected electroencephalography (EEG) changes during the preictal period (for review see Mormann et al., 2007). However, the presence and time-frame of consistent EEG changes prior to seizures remain uncertain and no method of seizure prediction can consistently predict seizures in different patients (Schelter et al., 2006; Mormann et al., 2007). The conventional range of EEG analysis involves frequencies below 100 Hz, but studies over the last decade suggest that localized higher frequencies may also be important. Frequencies above 100 Hz have been extensively characterized in human epileptic mesial temporal structures using depth microelectrodes. Whereas 100–200 Hz oscillations appear related to physiologic memory processing, higher frequencies, between 200 and 500 Hz, are associated with epileptogenic tissue (Bragin et al., 1999). These high frequency oscillations (HFOs) have been termed ripples (80–250 Hz) and fast ripples (250–500 Hz). Recently depth macroelectrodes and spectral and visual analysis techniques have also revealed focal HFOs in humans during interictal (Urrestarazu et al., 2006, 2007; Jacobs et al., 2008; Worrell et al., 2008) and ictal recordings (Jirsch et al., 2006; Ochi et al., 2007; Ramachandrannair et al., 2008). Discrete HFOs occurred mainly in regions of seizure onset and rarely in regions of secondary spread in mesial temporal as well as neocortical seizures. Moreover, no ictal high frequency activities occurred in the seizures of patients with poor localization (Jirsch et al., 2006). In animal studies, a clear relationship between the presence of HFOs and their degree of activity with spontaneous seizures could be shown (Bragin et al., 2004). In an in vitro model of low-Mg2+ seizures, an increase of HFOs preceded seizure activity (Khosravani et al., 2005). In rats, an increase of ripple and fast ripple bands could be observed within the dentate gyrus 1 s before seizure onset (Bragin et al., 2005). The behavior of HFOs in the preictal period in epilepsy patients has been evaluated in only one study (Khosravani et al., 2008), which found changes in the seconds preceding seizures. The possibility that HFOs may reflect basic epileptogenic processes and changes during the preictal period bears scientific and clinical interests. For instance, potential therapies involving EEG-triggered anticonvulsant injections or electrical stimulation could result in seizure control. Such antiseizure therapies may have greater chance of success if seizures could be predicted from the interictal EEG (Elger, 2001; Osorio et al., 2005). We hypothesized that HFOs in the range of 80–500 Hz measured using depth macroelectrodes in patients with intractable temporal and extratemporal epilepsy change in frequency of occurrence during the preictal state. Visual and spectral analyses techniques were used to assess HFO rates and band power in the 15 min prior to seizure onset. These two methods were chosen to detect changes in high frequency power in general but also in distinct HFOs. Distinct HFOs are very short events and, therefore, a change in their rate may not be detectable with spectral analysis.

In many patients with epilepsy, the area responsible for seizure generation, or the seizure onset zone (SOZ), may be difficult to define. Even in patients with identifiable lesions on brain MR imaging, noncongruent clinical and laboratory studies often indicate poor localization of the SOZ, which precludes a surgical approach in those patients with refractory epilepsy. Invasive intracranial EEG evaluation may help in those cases to finally obtain a good definition of the SOZ (Diehl & Lueders, 2000). These methods, however, do not always result in the finding of one clear SOZ, as seizures might be originating from more than one brain area and also because intracranial techniques are always spatially limited. For instance, seizures originating from areas not covered by electrodes but propagating to the actual electrode positions might lead to misinterpretation. Additionally, interictal epileptiform discharges (IED) or spikes, which define the irritative zone (IZ), might be seen in several electrode positions outside the SOZ; it remains unclear how these discharges are related to epileptogenesis (Hufnagel et al., 2000; Rosenow & Lueders, 2000) and how much importance one should put on them to define the epileptogenic area. It is known that patients with spikes generated in multiple brain areas are less likely to become seizure free after surgery than patients with well-localized spikes (Bautista et al., 1999). This apparent multiplicity of epileptic generators may represent widespread disease or might be related to secondary epileptogenesis (Rosenow & Lueders, 2000). Therefore, new measurements for epileptogenicity are required additionally to spikes and the SOZ in stereo EEG (SEEG). Traditionally, EEG oscillations are believed to be relevant up to frequencies in the gamma band (40–100 Hz) but recent findings suggest that high-frequency oscillations (HFOs) ranging between 100 and 500 Hz might be closely linked to epileptogenesis. They were studied in rodents as well as in humans using microwires (with a surface contact of 70 μm2) (Bragin et al., 1999; Staba et al., 2002) and more recently detected also in humans by using macroelecrode contacts, with a surface contact of 0.8 mm2 (Jirsch et al., 2006; Urrestarazu et al., 2006, 2007). HFOs ranging from 100 to 250 Hz and described as ripples were seen in the hippocampus (Hc) and entorhinal cortex of normal rats (Chrobak & Buzsaki, 1996; Chrobak et al., 2000). In humans, similar oscillations generally showed a lower frequency range, between 80 and 160 Hz, and were found within epileptic tissue and outside in the less affected Hc of epileptic patients studied with bilateral temporal lobe intracerebral depth electrodes (Bragin et al., 1999; Staba et al., 2004). It was hypothesized that they are physiological rhythms closely linked to memory consolidation (Draguhn et al., 2000), but additional “pathological” ripples could be observed as well (Bragin et al., 2004). The differentiation between pathological and physiological events remains unclear (Le Van Quyen et al., 2006). HFOs between 250 and 500 Hz, called fast ripples, have been recorded from normal rodent and human brains (Curio et al., 1994, 1997; Jones et al., 2000; Ikeda et al., 2005). It was hypothesized that they are related to somatosensory stimulation and sensory information processing (Curio et al., 1997; Gobbele et al., 2004). Nevertheless, fast ripples in mesial temporal structures were more clearly linked to epileptogenesis than ripples (Staba et al., 2002). Ripples and fast ripples occur frequently during IEDs and may reflect pathological hypersynchronous events. In general, these HFOs are seen more frequently during non-rapid-eye movement sleep (NREM) sleep compared to rapid-eye movement sleep (REM) sleep or wakefulness (Staba et al., 2004). Recently, it has been shown that HFOs may not exclusively be recorded from microwires but also from intracranial macroelectrodes (Jirsch et al., 2006; Urrestarazu et al., 2006). During ictal recordings, HFOs could be identified and occurred mostly in the region of primary epileptogenesis and less frequently in areas of secondary spread (Jirsch et al., 2006). Analysis of interictal recordings showed a clear relation between spikes and fast ripples and again a close relation between HFOs and the SOZ (Urrestarazu et al., 2007). While most fast ripples were noted during spikes, others could be found independently. Fast ripples were more restricted to the presumed SOZ than ripples. In the studies with microwires, the relationship between HFOs and spikes was not specifically addressed but most examples of fast ripples show themoccurring during spikes. The exact relationship between interictal discharges and HFOs, so far, remains unclear. All previous studies focused on channels that showed spiking activity in the first place. It is therefore unknown whether HFOs in general are produced by the same generators as IEDs or derive from independent structures. HFOs have never been measured in interictally inactive channels of epileptic patients, and their presence in these channels might provide additional information on epileptogenesis. Furthermore, HFOs occurring within spiking channels but independently of spikes need to be investigated more closely to gain information about the value of HFOs for the identification of the SOZ independently of spikes. In this study we analyzed HFOs in spiking and nonspiking channels, looking at interictal intracerebral SEEG recordings of 10 patients with lesional focal epilepsy. Differences between HFOs occurring at the same time or independent of spikes, as well as the exact relationship between spikes and HFOs were investigated. We also compared the capacity of the limbic and neocortical (Nc) structures to generate fast oscillations, and, finally, we determined if HFOs could help identify the SOZ. We hypothesize that HFOs occur to a large extent independently of spikes and can add additional localizing information to the SEEG investigation.

High Frequency Oscillations (HFOs) in the EEG are a promising biomarker of epileptogenic tissue. Given that the visual marking of HFOs is highly time-consuming and subjective, automatic detectors are necessary. In this study, we present a novel automatic detector that detects HFOs by incorporating information of previously detected baselines. The detector was trained on 72 channels and tested on 278, achieving a mean sensitivity of 96.8% with a mean false positive rate of 4.86%. This low rate is reasonable since only visually marked baseline segments were considered as the true negatives. This detector could be useful for the systematic study of HFOs and for their eventual clinical application.

Interictal fast oscillations between 100 and 500 Hz have been reported in signals recorded from implanted microelectrodes in epileptic patients and experimental rat models. Oscillations between 250 and 500 Hz, or fast ripples (FR), appeared related to the epileptic focus whereas ripples (80-200 Hz) were not. We report high-frequency oscillations recorded with intracranial macroelectrodes in seven patients with refractory focal epilepsy during slow-wave sleep. We characterize the relation of fast oscillations to the seizure focus and quantify their concordance with epileptiform transients, with which they are strongly associated. The patients were selected because interictal spikes were found within and outside the seizure onset zone. Visual inspection was used to identify and classify the ripples and FRs according to their relation to epileptiform spikes. Continuous-time wavelet analysis was used to compute their power. Ripples were present in all patients while FRs where found in five of the seven patients. Most ripples and FRs occurred at the same time as epileptiform transients. The rate of occurrence of ripples was higher within the seizure onset zone than outside in four of seven patients. The rate of FRs was much higher within the seizure onset zone than outside in four of the five patients with FRs (in these four patients, FRs were almost inexistent outside the seizure onset zone). The power of ripples and FRs tended to be higher in the electrodes where their rate was also higher. These results indicate that FRs were more restricted to the electrodes located within the seizure onset zone, especially to the hippocampus, than ripples. In only one patient, FRs were more frequent outside the seizure onset zone; this patient was the only one with cortical dysplasia and the electrode with a high rate of FRs was inside the lesion. This study demonstrates that interictal ripples and FRs can be recorded with depth macroelectrodes in patients. Most occur at the time of epileptiform spikes but some are isolated. Ripples do not show a clear differentiation between the seizure onset zone and remote areas, whereas FRs have a higher rate and higher power in the seizure onset zone. Our results also suggest a special capacity of the abnormal hippocampus to generate FRs, although they were also recorded in other structures.

ArtiSynth is a modular, component‐based system for performing dynamic 3D simulations of the human vocal tract and face. It provides a test bed for research in areas such as speech synthesis, linguistics, medicine, and dentistry. ArtiSynths framework enables researchers to construct, refine, and exchange models of all parts of the vocal tract and surrounding structures. ArtiSynth introduces a probe concept to unify input and output data flow, which allows control of and access to models with time varying data series. ArtiSynth supports interconnected heterogeneous models, such as rigid body, mass‐spring, and parametric, using a point‐set connection method, called markers, for constraint satisfaction. Using ArtiSynth, we created a muscle‐driven rigid body jaw model, a parametric principle component tongue model from MRI images, a parametric lip model, and mass‐spring face tissue model. We combined them in various ways. Data from medical imaging (MRI, CT, and ultrasound) and other technologies such as optical tracking can be used to drive ArtiSynth models. We are currently developing an acoustical rendering framework supporting source‐filter models and other advanced methods. The system incorporates a powerful scripting interface as well as an easy‐to‐use graphical interface. [Work supported by NSERC Canada and ATR Japan.]