Deep neural models for color classification and color constancy

Color constancy is our ability to perceive constant colors across varying illuminations. Here, we trained deep neural networks to be color constant and evaluated their performance with varying cues. Inputs to the networks consisted of two-dimensional images of simulated cone excitations derived from three-dimensional (3D) rendered scenes of 2,115 different 3D shapes, with spectral reflectances of 1,600 different Munsell chips, illuminated under 278 different natural illuminations. The models were trained to classify the reflectance of the objects. Testing was done with four new illuminations with equally spaced CIEL*a*b* chromaticities, two along the daylight locus and two orthogonal to it. High levels of color constancy were achieved with different deep neural networks, and constancy was higher along the daylight locus. When gradually removing cues from the scene, constancy decreased. Both ResNets and classical ConvNets of varying degrees of complexity performed well. However, DeepCC, our simplest sequential convolutional network, represented colors along the three color dimensions of human color vision, while ResNets showed a more complex representation.     

Relearning Auditory Spectral Cues for Locations Inside and Outside the Visual Field

Previous research has demonstrated that, over a period of weeks, the auditory system accommodates to changes in the monaural spectral cues for sound locations within the frontal region of space. We were interested to determine if similar accommodation could occur for locations in the posterior regions of space, i.e. in the absence of contemporaneous visual information that indicates any mismatch between the perceived and actual location of a sound source. To distort the normal spectral cues to sound location, eight listeners wore small moulds in each ear. HRTF recordings confirmed that while the moulds substantially altered the monaural spectral cues, sufficient residual cues were retained to provide a basis for relearning. Compared to control measures, sound localization performance initially decreased significantly, with a sevenfold increase in front–back confusions and elevation errors more than doubled. Subjects wore the moulds continuously for a period of up to 60 days (median 38 days), over which time performance improved but remained significantly poorer than control levels. Sound localization performance for frontal locations (audio-visual field) was compared with that for posterior space (audio-only field), and there was no significant difference between regions in either the extent or rate of accommodation. This suggests a common mechanism for both regions of space that does not rely on contemporaneous visual information as a teacher signal for recalibration of the auditory system to modified spectral cues.     

Soft constraints in nonlinear spectral fitting with regularized lineshape deconvolution

This article presents a novel method for incorporating a priori knowledge into regularized nonlinear spectral fitting as soft constraints. Regularization was recently introduced to lineshape deconvolution as a method for correcting spectral distortions. Here, the deconvoluted lineshape was described by a new type of lineshape model and applied to spectral fitting. The nonlinear spectral fitting was carried out in two steps that were subject to hard constraints and soft constraints, respectively. The hard constraints step provided a starting point and, therefore, only the changes of the relevant variables were constrained in the soft constraints step and incorporated into the linear substeps of the Levenberg‐Marquardt algorithm. The method was demonstrated using localized averaged echo time point resolved spectroscopy proton spectroscopy of human brains. Magn Reson Med 69:912–919, 2013. © 2012 Wiley Periodicals, Inc.     

Turning back from the brink: Detecting an impending regime shift in time to avert it

Ecological regime shifts are large, abrupt, long-lasting changes in ecosystems that often have considerable impacts on human economies and societies. Avoiding unintentional regime shifts is widely regarded as desirable, but prediction of ecological regime shifts is notoriously difficult. Recent research indicates that changes in ecological time series (e.g., increased variability and autocorrelation) could potentially serve as early warning indicators of impending shifts. A critical question, however, is whether such indicators provide sufficient warning to adapt management to avert regime shifts. We examine this question using a fisheries model, with regime shifts driven by angling (amenable to rapid reduction) or shoreline development (only gradual restoration is possible). The model represents key features of a broad class of ecological regime shifts. We find that if drivers can only be manipulated gradually management action is needed substantially before a regime shift to avert it; if drivers can be rapidly altered aversive action may be delayed until a shift is underway. Large increases in the indicators only occur once a regime shift is initiated, often too late for management to avert a shift. To improve usefulness in averting regime shifts, we suggest that research focus on defining critical indicator levels rather than detecting change in the indicators. Ideally, critical indicator levels should be related to switches in ecosystem attractors; we present a new spectral density ratio indicator to this end. Averting ecological regime shifts is also dependent on developing policy processes that enable society to respond more rapidly to information about impending regime shifts.     

Substrates of auditory frequency integration in a nucleus of the lateral lemniscus

In the intermediate nucleus of the lateral lemniscus (INLL), some neurons display a form of spectral integration in which excitatory responses to sounds at their best frequency are inhibited by sounds within a frequency band at least one octave lower. Previous work showed that this response property depends on low-frequency-tuned glycinergic input. To identify all sources of inputs to these INLL neurons, and in particular the low-frequency glycinergic input, we combined retrograde tracing with immunohistochemistry for the neurotransmitter glycine. We deposited a retrograde tracer at recording sites displaying either high best frequencies (>75 kHz) in conjunction with combination-sensitive inhibition, or at sites displaying low best frequencies (23–30 kHz). Most retrogradely labeled cells were located in the ipsilateral medial nucleus of the trapezoid body (MNTB) and contralateral anteroventral cochlear nucleus. Consistent labeling, but in fewer numbers, was observed in the ipsilateral lateral nucleus of the trapezoid body (LNTB), contralateral posteroventral cochlear nucleus, and a few other brainstem nuclei. When tracer deposits were combined with glycine immunohistochemistry, most double-labeled cells were observed in the ipsilateral MNTB (84%), with fewer in LNTB (13%). After tracer deposits at combination-sensitive recording sites, a striking result was that MNTB labeling occurred in both medial and lateral regions. This labeling appeared to overlap the MNTB labeling that resulted from tracer deposits in low-frequency recording sites of INLL. These findings suggest that MNTB is the most likely source of low-frequency glycinergic input to INLL neurons with high best frequencies and combination-sensitive inhibition. This work establishes an anatomical basis for frequency integration in the auditory brainstem.     

Extraction of Evoked Potentials Based on Borel Spectral Measure for Less Trial Mixtures in Stable Distribution Environments

Stable process can better model the EEG noises found in the impact acceleration experiment. A new method for identifying the independent components of evoked potentials for less trial mixtures is proposed. The method is based on an estimate of the discrete Borel spectral measure for the characteristic function of an alpha stable random vector. Simulations demonstrate that the proposed method can identify independent components and the basis vectors of mixing matrix in the so-called less trial mixtures case of more sources.     

The frequency range of TMJ sounds

There are conflicting opinions about the frequency range of temporomandibular joint (TMJ) sounds. Some authors claim that the upper limit is about 650 Hz. The aim was to test the hypothesis that TMJ sounds may contain frequencies well above 650 Hz but that significant amounts of their energy are lost if the vibrations are recorded using contact sensors and/or travel far through the head tissues. Time-frequency distributions of 172 TMJ clickings (three subjects) were compared between recordings with one microphone in the ear canal and a skin contact transducer above the clicking joint and between recordings from two microphones, one in each ear canal. The energy peaks of the clickings recorded with a microphone in the ear canal on the clicking side were often well above 650 Hz and always in a significantly higher area (range 117-1922 Hz, P < 0.05 or lower) than in recordings obtained with contact sensors (range 47-375 Hz) or in microphone recordings from the opposite ear canal (range 141-703 Hz). Future studies are required to establish normative frequency range values of TMJ sounds but need methods also capable of recording the high frequency vibrations.     

Effects of water on cortical excitability in humans

The effects of water on cortical excitability, measured using magnetoencephalographic recordings, were investigated in a sample of 19 healthy volunteers in a double-blind, placebo experiment comparing water with saline solution. Spontaneous magnetoencephalogram as well as auditory-evoked magnetic fields were recorded before and after the drinking of 750 mL water (9 subjects) or saline solution (10 subjects) and during and after hyperventilation following the drinking conditions. Hyperventilation was used to enhance the hypothesized synchronizing effect of water on spontaneous magnetoencephalographic activity. In addition, the magnetic fields were measured during a dichotic listening task under attended and unattended conditions. The prediction, that intake of water, because of induced cell swelling, will increase neuronal excitability and lead to an increased synchronization of the spontaneous magnetoencephalogram during hyperventilation was confirmed. Hyperventilation induced an increase of spectral power in all frequency bands particularly theta and delta power after water drinking. Furthermore, there was an increase of magnetic mismatch negativity (MMNm) amplitude in attended conditions and a simultaneous decrease in unattended conditions after water drinking. N1m (magnetic N1 wave) revealed significant changes during experimental conditions: increase after drinking and decrease after hyperventilation in both groups. MMNm for attended conditions showed a high positive correlation with osmolality changes (difference in the mol solute per kg water before and after drinking); N1m and PNm (magnetic processing negativity) as well as MMNm for unattended conditions showed significant correlations with subjective ratings of thirst and mood state.