The Gabor-Einstein Wavelet: A Model for the Receptive Fields of V1 to MT Neurons Stephen G. Odaibo1,2 M.S.(Math), M.S.(Comp. Sci.), M.D. 1Quantum Lucid Research Laboratories, 4 1 Computational Visual Neuroscience Group 0 [email protected] 2 n 2Howard University Hospital, Ophthalmology a J Washington,D.C.,U.S.A. 2 2 Abstract ] C Our visual system is astonishingly efficient at detecting moving ob- N jects. Thisprocessismediatedbytheneuronswhichconnecttheprimary visualcortex(V1)tothemiddletemporal(MT)area. Interestingly,since . o Kuffler’s pioneering experiments on retinal ganglion cells, mathematical bi models have been vital for advancing our understanding of the receptive - fieldsofvisualneurons. However,existingmodelswerenotdesignedtode- q scribe the most salient attributes of the highly specialized neurons in the [ V1 to MT motion processing stream; and they have not been able to do 1 so. Here,weintroducetheGabor-Einsteinwavelet,anewfamilyoffunc- v tionsforrepresentingthereceptivefieldsofV1toMTneurons. Weshow 9 thatthewayspaceandtimearemixedinthevisualcortexisanalogousto 8 the way they are mixed in the special theory of relativity (STR). Hence 5 we constrained the Gabor-Einstein model by requiring: (i) relativistic- 5 invariance of the wave carrier, and (ii) the minimum possible number of . 1 parameters. From these two constraints, the sinc function emerged as 0 a natural descriptor of the wave carrier. The particular distribution of 4 lowpasstobandpasstemporalfrequencyfilteringpropertiesofV1toMT 1 neurons (Foster et al 1985; DeAngelis et al 1993b; Hawken et al 1996) : is clearly explained by the Gabor-Einstein basis. Furthermore, it does v i so in a manner innately representative of the motion-processing stream’s X neuronalhierarchy. Ouranalysisandcomputersimulationsshowthatthe r distribution of temporal frequency filtering properties along the motion a processing stream is a direct effect of the way the brain jointly encodes space and time. We uncovered this fundamental link by demonstrating that analogous mathematical structures underlie STR and the joint cor- ticalencodingofspaceandtime. Thislinkwillprovidenewphysiological insights into how the brain represents visual information. Keywords: Visualcortex,V1,MT,Receptivefield,Gabor,Responseproperties 1 1 Introduction 1.1 Aim To introduce a model which innately and efficiently represents the salient prop- erties of receptive fields of neurons in the V1 to MT motion processing stream. 1.1.1 Specific Aim As one traces the neural pathway connecting the input layers of V1 to area MT, neurons become increasingly specialized for motion detection. In ascend- ing order, attributes such as orientation selectivity, direction selectivity, speed tuning,increasingpreferredspeeds,componentspatialfrequencyselectivity,and pattern (plaid) spatial frequency selectivity are sequentially acquired. Further- more, the temporal frequency filtering properties of the neuronal population changes in a particular way along this pathway. Specifically, the proportion of bandpass temporal frequency filtering neurons to lowpass temporal frequency filteringneuronsincreases. Existingreceptivefieldmodelsdonotrepresentthis fundamentalemergentproperty. Ourspecificaimistointroduceanddescribea modelforthereceptivefieldsofV1toMTneuronswhichinnatelyandefficiently represents the aforementioned properties. 1.2 Motivation and Background The detection of moving objects in our environment is both critical to our experience of the world and to our very survival. Imagine a hunter in rapid pursuit of a small agile animal. Both predator and prey detect abrupt changes in velocity occurring over an interval of less than a tenth of a second. The precision of our visual motion detection system is undoubtedly striking. The middle temporal (MT) area of the mammalian brain plays a key role in thevisualprocessingofmotion. TheimportanceofareaMTinmotionprocess- ingissupportedbyagrowingwealthofelectrophysiologicalevidencewhichwas initiatedbyDubnerandZeki’s1971report(DubnerandZeki,1971;Zeki,1974, 1980;VanEssen,Maunsell,andBixby,1981;MaunsellandVanEssen,1983a,b; Felleman and Kaas, 1984; Ungerleider and Desimone, 1986; Maunsell and New- some, 1987; Rodman and Albright, 1989; Movshon and Newsome, 1996; Smith, Majaj,andMovshon,2005). Forpurposesofmotionprocessing,themaininput signals into MT originate in specialized cells of the primary visual cortex (V1). Kuffler’s early studies of retinal ganglion cell response properties led to sim- ilar studies of simple cortical cells by Hubel and Wiesel. Together, their work generatedgreatinterestinneuronalreceptivefields(Kuffler,1953;Hubel,1957, 1959;HubelandWiesel,1959;Marˇcelja,1980;Daugman,1985). Thefunctional forms of these receptive fields are at once beautifully simple yet enormously complex. Hence mathematical models have been used in step with electrophys- iological studies to advance our understanding of their properties. Initially, fo- cuswaspredominantlyontheirspatialstructure(Marˇcelja,1980;Gabor,1946). Now, however, their temporal structure is increasingly studied in tandem. In particular, for most neurons in the V1 to MT processing stream, it is now ap- preciatedthatspatialandtemporalfeaturescannotbestudiedseparately. They are spatiotemporally inseparable entities. In particular, motion is encoded by 2 orientation in the spectral domain, and spatiotemporally-oriented filters are therefore motion detectors (Watson and Ahumada, 1983, 1985; Heeger, 1987). Therefore at first glance it may seem an easy matter to mathematically model motion detecting neurons. The challenge, however, is to develop physiologi- cally sound receptive field models which reflect the hierarchical structure of the motionprocessingstream. Variousmodelsdoexistwhicharespatiotemporally- oriented filters, and are therefore motion detectors from a mathematical stand- point (Adelson and Bergen, 1985; Qian, 1994; Qian, Andersen, and Adelson, 1994; Qian and Andersen, 1997; Qian and Freeman, 2009). However, they fail to represent one of the most salient characterizing attributes of the motion processing stream: the lowpass to bandpass distribution of temporal frequency filtering properties along the V1 to MT specialization hierarchy. This feature is discussed further in the next subsection. In addition to the above temporal frequency filter property profile, the neu- rons of the V1 to MT motion processing stream have certain characteristics which set them apart and prescribe bounds on models of their spatiotemporal receptive field structure. Namely, they are tuned to specific orientations, di- rections, and speeds (Movshon and Newsome, 1996); and they are more likely than other V1 cells to exhibit “end-stopping” phenomena (Sceniak, Hawken, and Shapley, 2001). Of the V1 cells, the subclass with direct synaptic projec- tions from V1 to MT are the most specialized towards motion and are thought to be the main channels of motion substrates incident into MT neurons. For instance, Churchland et al showed that V1 and MT neurons lose direction se- lectivity for similar values of spatial disparity, suggesting V1 is MT’s source of direction selectivity (Churchland, Priebe, and Lisberger, 2005). Additionally, these V1 to MT neurons cluster motion substrate properties such as strong di- rection selectivity and tuning to high speeds. For example, Foster et al found that highly direction selective neurons in the macaque V1 and V2 were more likely to be tuned to higher temporal frequencies and lower spatial frequencies, i.e. higher speeds (Foster, Gaska, Nagler, and Pollen, 1985). In other words, neurons that are highly direction selective are more likely to also be tuned to higher speeds. Such neurons are higher up in the V1 to MT hierarchy. Simi- larly,McLeanandPalmerstudieddirectionselectivityincatstriatecortex,and foundthatspace-timeinseparablecellsweredirectionselectivewhilespace-time separable cells were not direction selective McLean and Palmer (1989). Again reflecting the hierarchy of the motion processing stream. The distribution of some distinct anatomical and histological features have been shown to correlate with the motion-specialization hierarchy. For instance, Shipp and Zeki’s retrograde tracer studies showed that the majority of direct V1 to MT projecting neurons originate within layer 4B (Shipp and Zeki, 1989). Based on their results, they suggested that layer 4B contained a functionally subspecialized anatomically segregated group whose members each have direct synaptic connections with MT cells. The exact properties of the V1 to MT projectors are likely to be intermediate between those of cells in V1 layer 4Cα and MT neurons. This is because input into MT appears to be largely con- stituted of magnocellular predominant streams from 4Cα, while parvocellular predominant streams may play a much less role (Livingstone and Hubel, 1988; Maunsell, Nealey, and DePriest, 1990; Yabuta, Sawatari, and Callaway, 2001). Progressive specialization along the V1 to MT motion processing stream is also observable in the morphological characteristics of the neurons. For instance, 3 distinguishing attributes of the V1-MT or V2-MT projectors such as size, ar- borization patterns, terminal bouton morphology, and distribution have been observed (Rockland, 1989, 1995; Anderson, Binzegger, Martin, and Rockland, 1998;AndersonandMartin,2002). SincichandHortondemonstratedthatlayer 4BneuronsprojectingtoareaMTweregenerallylargerthanthoseprojectingto layerV2. Overall,neuronsintheV1toMTmotionprocessingstreamarehighly andprogressivelyspecialized. Hencedescriptionoftheirreceptivefieldsrequires adequatelyspecializedmodelswhichreflectnotonlytheirindividualattributes, but also the emergent hierarchical properties of the network. Next we discuss one of the most fundamental of such emergent network properties: the partic- ular distribution of temporal frequency filtering types along the stream (Foster et al., 1985; DeAngelis et al., 1993b; Hawken et al., 1996). 1.3 Temporal Frequency Filtering Property Distribution Hawkenetalfoundthatdirection-selectivecellsweremostlybandpasstemporal frequency filters, while cells which were not direction-selective were equally dis- tributedintobandpassandlowpasstemporalfrequencyfilteringtypes(Hawken, Shapley,andGrosof,1996). Fosteretalfoundasimilarphenomenoninmacaque V1 and V2 neurons. V1 neurons were more likely to be lowpass temporal frequency filters, while their more specialized downstream heirs, V2 neurons, were more likely to be bandpass temporal frequency filters (Foster, Gaska, Na- gler,andPollen,1985). Lessspecializedneuronslocatedanatomicallyupstream (lowerdowninthehierarchy)aremorelikelytohavelowpasstemporalfrequency filtercharacteristics,whilemorespecializedneuronslocatedanatomicallydown- stream(higherupinthehierarchy)aremorelikelytodisplaybandpasstemporal frequency filter characteristics. For example, LGN cells are at best only weakly tunedtodirectionandorientation(Xuetal.(2001),FersterandMiller(2000)) and are hence equally distributed into lowpass and bandpass categories. Layer 4B V1 cells and MT cells on the other hand, are almost all direction selective (Maunsell and Van Essen, 1983b) and hence are mostly bandpass temporal fre- quency filters. We firmly believe this classification is not arbitrary, but instead isadirectmanifestationoftheparticularspatiotemporalstructureoftheV1to MTmotionprocessingstream. Consequently,thereceptivefieldmodelmustre- flect this increased tendency for bandpass-ness with ascension up the hierarchy. In other words, the representation scheme must be one that is inherently more likely to deliver bandpass-ness to a more specialized cell (such as in layer 4B or MT) and lowpass-ness to a less specialized cell (such as in layer 4A, 4Cα, or 4Cβ). However,noneoftheexistingreceptivefieldmodelsreflectthisunderlying spatiotemporal structure. In contrast, as will be seen in Section (2) below, the Gabor-Einstein wavelet’s wave carrier is a sinc function which directly confers the aforementioned salient property. Individual Gabor-Einstein basis elements are lowpass temporal frequency filters; and bandpass temporal frequency filters can only be obtained by combinations of basis elements. A study by DeAn- gelis and colleagues also corroborates the above. They found that temporally monophasicV1cellsinthecatwerealmostalwayslowpasstemporalfrequency filters, while temporally biphasic or multiphasic V1 cells were almost always bandpasstemporalfrequencyfilters(DeAngelis,Ohzawa,andFreeman,1993b). The explanation for this finding is inherent and explicit in the Gabor-Einstein waveletbasis,wherebandpasstemporalcharacternecessarilyresultsfrombipha- 4 sicormultiphasiccombination. Allmonophasicelementsontheotherhand,are lowpass temporal frequency filters. However, we will see that according to the model, the converse is not true: i.e. not all lowpass temporal frequency filters are monophasic, and not all multiphasic combinations yield bandpass temporal frequency filters. 1.4 Analogy to Special Relativity Perceptualandelectrophysiologicalstudiesrevealintriguingsimilaritiesbetween howspaceandtimearemixedinthevisualcortexandhowtheyaremixedinthe specialtheoryofrelativity(STR).Forinstance, withincertain limits, binocular neurons cannot distinguish a temporal delay from a spatial difference. As a result, inducing a monocular time delay by placing a neutral density filter over oneeyebutnottheother,causesapendulumswingingina2Dplanarspacetobe perceived as swinging in 3D depth space. This is known as the Pulfrich effect. Also, most neurons in the V1 to MT stream are maximally excited only by stimuli moving at that neuron’s preferred speed. They are speed tuned. In this subsection,webrieflyreviewandsummarizethespecialtheoryofrelativity. We identify the stroboscopic pulfrich effect and speed tuning as cortical analogues of STR’s joint encoding of space and time. And finally, we explain how STR is used in (and motivates) the design of the Gabor-Einstein wavelet. 1.4.1 Review of Special Relativity In 1905, Albert Einstein proposed the special theory of relativity (Einstein, 1905). He was motivated by one thing: a firm belief that Maxwell’s equations of electromagnetism are laws of nature. Maxwell’s equations are a classical description of light (Maxwell, 1861a,b, 1862a,b). According to Maxwell, light is the propagation of electric and magnetic waves interweaving in a particular way perpendicular to each other. The speed of propagation in a vacuum is the constant c = 299,792,458 m/s. The other governing constraint on STR is the relativity principle. It has been around and generally accepted since the time of Galileo, and it states that the laws of nature are true and the same for all non-accelerating observers. Einstein believed Maxwell’s equations to be laws of nature,andthereforetosatisfytherelativityprinciple. Inotherwords,Einstein believed the speed of light must be the same to all non-accelerating observers regardless of their velocity relative to each other. For the speed of light to be fixed, something(s) had to yield and be unfixed. What had to yield were the components of the definition of speed, i.e. space and time. Space and time had tobecomefunctionsofrelativevelocityandofeachother,toensurethespeedof lightremainsconstanttoallnon-acceleratingobservers. Thetransformationrule —of space and time coordinates between two observers moving with constant velocity relative to each other— is called the Lorentz equations. It specifies how space and time are mixed in STR. More specifically, the Lorentz equations are the transformation rule between the (spacetime) coordinates of two inertial referenceframes. Fortwoinertialreferenceframes,RandR(cid:48),movingexclusively in the x direction with constant velocity v relative to each other, the Lorentz transformation is given by, 5 (cid:16) v (cid:17) t(cid:48) =γ t− x , c2 (1) x(cid:48) =γ(x−vt), where (t(cid:48),x(cid:48)) are the coordinates of R(cid:48), (t,x) are the coordinates of R, c is the speed of light, and γ is the Lorentz factor and is given by, 1 γ = . (2) (cid:113) 1− v2 c2 We demonstrate below that the stroboscopic pulfrich effect is analogous to STRinonerespect,andspeedtuningofneuronsisanalogoustoSTRinanother respect. 1.4.2 The Stroboscopic Pulfrich Effect In the classical pulfrich effect, a monocular temporal delay results in a percep- tion of depth in pendular swing (Pulfrich, 1922). It has a simple geometric explanation. By the time the temporally delayed eye “sees” the pendulum at a given retinal position, the pendulum has advanced to a further position in its trajectory, hence the visual cortex always receives images that are spatially offset between the eyes. The spatial disparity results in stereoscopic depth per- ception. This form of pulfrich phenomena is termed classical to distinguish it from the stroboscopic pulfrich effect. In the stroboscopic pulfrich effect, the pendulumisnotacontinuouslymovingbulb,butinsteadisastrobelightwhich samples the trajectory of the classical pendulum in time and space (Lee, 1970; Morgan, 1976; Burr and Ross, 1979; Morgan, 1979). The stroboscopic pulfrich effectcannotbeexplainedbythesimplegeometricillustrationthatexplainsthe classical pulfrich effect. This is because depth is still perceived in sequences involving a temporal delay in flash between eyes, but particularly lacking an interocular spatial disparity in the flash (Burr and Ross, 1979). In stereoscopic depth perception, a planar (x,y) difference in retinal image position (∆x) transforms into (is perceived as) a displacement along the z di- rection. In the stroboscopic pulfrich effect, a temporal difference (∆t) between eyes in retino-cortical transmission of image signal, results in a planar pen- dulum trajectory being perceived as (transforming into) depth. In STR, the transformation is between two inertial reference frames moving relative to each other. By analogy, we assert that for the stroboscopic pulfrich effect, there is a transformationbetweentheinterocularretinalspace(x ,t )andtheperceptual r r depthspace(x ,t ). Weidentifythevariables(x ,t )oftheinterocularspaceas z z r r the difference between eyes of the corresponding variables in the retino-cortical space i.e. x is the difference between the left and right eye retinal positions r of an image; and t is the difference between the right and left eyes in retino- r cortical image transmission times. The (x ,t ) → (x ,t ) transformation may r r z z take the form x = α(x ±v˜t ), were v˜ is proportional to pendulum velocity z r r and α is a function of v˜. Note the similarity between x and the equation for z x(cid:48) in the Lorentz transformation, x(cid:48) = γ(x−vt). This analogy suggests that similar mathematical structures underlie STR and the joint encoding of space and time in the visual cortex. The stroboscopic pulfrich effect is not by itself direct evidence of joint encoding of spatial and temporal disparities (Read and 6 (a) (b) Figure 1: The pulfrich effect: Top figure, (a), shows planar 2D pendular swing. Bottom figure, (b), shows perception of 3D pendular swing (depth) resulting from interocular temporal delay. 7 Cumming, 2005a,b,c), however direct electrophysiological evidence does exist for such linkage in cortical encoding (Carney, Paradiso, and Freeman, 1989; Anzai, Ohzawa, and Freeman, 2001; Pack, Born, and Livingstone, 2003). This mathematical similarity influences our design of the Gabor-Einstein wavelet in a particular way which we describe in Section (1.5) below. But first, we discuss another neuronal phenomenon, speed tuning, which is analogous to STR in a certain physical sense. 1.4.3 Speed Tuning Most motion processing stream neurons in V1 or area MT are speed tuned. They have a preferred speed, which is the speed of moving stimuli to which they maximally respond. Other speeds also elicit a response, but with a lower rateofactionpotential. Whenpresentedmovingsinewavegratingstimuli,truly speed tuned neurons respond maximally to their preferred speed independent of the spatial or temporal frequency of the stimuli. The velocity, ξ, of a moving wave is, ω ξ = , (3) u where ω is temporal frequency and u is spatial frequency. In higher dimen- sions, ω =uξ +vξ , (4) x y whereuisthespatialfrequencyinthexdirection, v isthespatialfrequency in the y direction, ξ is the x component of velocity, and ξ is the y compo- x y nent of velocity. From both equations above, it is clear that for a speed tuned neuron, achangeinthespatialfrequencyofthestimuliwouldnecessaryrequire a complementary change in the temporal frequency. Hence a broad range of stimuli with widely varying spatial and/or temporal frequencies can maximally excite the neuron, so long as the above equations are satisfied for the neuron’s preferred speed ξ . This is analogous to how the special theory of relativity is 0 motivated by fixing the speed of light. In the case of STR the variables which necessarilychangeincomplementaryfashionarethespaceandtimecoordinates as described in the Lorentz equations. 1.5 Design of the Gabor-Einstein Wavelet WeconstrainedtheGabor-Einsteinmodelbyrequiringthefollowingproperties: 1. The minimum possible number of parameters 2. Relativistic-invariance of the wave carrier 1.5.1 Minimal Number of Parameters In addition to orientation in the spatiotemporal frequency domain, localiza- tion in the space, time, spatial frequency, and temporal frequency domains are the essential features of the V1 to MT neuron receptive field. These are the most basic attributes of the receptive field. As such, they correspond with the minimum number of model parameters needed. Specifically, speed tuning, the 8 spatial and the temporal localization envelopes, the amplitude modulation pa- rameter, and the spatial and temporal frequencies must all be represented. If oneadoptsthenotionofagaussianspatiallocalizer(Marˇcelja,1980;Daugman, 1985), the minimum number of parameters add up to eight: four for the spa- tial envelope and four for the wave carrier. The four necessary spatial envelope parameters are: amplitude factor, A; variance in the x direction, σ ; variance x in the y direction, σ ; and spatial orientation of envelope, θ . The four nec- y e essary wave carrier parameters are: temporal frequency, ω; spatial frequency in the x direction, u; spatial frequency in the y direction v, and the phase, φ. The orientation of the wave carrier θ =tan−1(v/u) is not an independent pa- s rameter. We use it in computer simulations below only for convenience. The essential feature that is not yet explicitly accounted for in the above parameter tally is the temporal envelope of the receptive field. This can be implemented either through the gaussian envelope or through the wave carrier, and in either case may conceivably involve an additional parameter. The next design crite- ria, relativistic-invariance of the wave carrier, resolves this ambiguity without adding another parameter. 1.5.2 Relativistic-invariance of the Wave Carrier Given the analogies of spatiotemporal mixing in the visual cortex to spatiotem- poral mixing in the special theory of relativity, it is reasonable to conjecture that their endpoints are within proximity of each other. Einstein eventually sought a lexicon in which physical laws are expressed the same way in any two inertial reference frames. That lexicon is called relativistic-invariance (or Lorentz-invariance). By analogy, we ultimately seek a lexicon in which physi- cal laws can be expressed the same way in both the interocular and perceptual spaces. Lorentz-invarianceofphysicallawisthespecification(orendpoint)aris- ingoutofSTR.Hence,requiringLorentz-invarianceofthereceptivefield’swave carrier is desirable. Moreover, this immediately resolves the above ambiguity regarding implementation of the temporal envelope profile. The temporal en- velope must be implemented via the wave carrier for relativistic invariance to hold. And since the receptive field amplitude eventually falls, the sinc function emergesasanaturaldescriptor. Furthermore,thesincfunctionhasthedistinct advantage of not introducing any additional parameters. 1.6 Summary Thewayspaceandtimearemixedinthevisualcortexbearsresemblancetothe way they are mixed in the special theory of relativity. We demonstrated that the stroboscopic pulfrich effect and speed tuning are cortical analogues of the spacetime mixing mechanics of STR. In STR, the mixture of space and time is described by the Lorentz transform which relates the spacetime coordinates of inertial reference frames moving with constant velocity relative to each other. In the striate and extrastriate motion processing stream, the analogous trans- formation is between the 3D interocular space (x,y planar inter-retinal space + inter-retinocorticaltime)andthe3Dperceptualspace(x,y,z3Dspace). Though on the surface, STR and cortical spacetime encoding appear to be unrelated processes,weconjectureandhavepartlyshownasharedunderlyingmathemat- ical structure. This structure can be exploited to deepen our understanding of 9 cortical encoding of motion, depth, and more fundamentally, the joint encod- ing of space and time in the visual cortex. Here, we proceeded by requiring relativistic-invariance of the receptive field’s wave carrier. We simultaneously required that the model be constrained to the minimum possible number of parameters. Under these constraints, the sinc function with energy-momentum relationasargumentemergedasanaturaldescriptorofthereceptivefield’swave carrier. Furthermore, the Gabor-Einstein wavelet explains a number of salient physiological attributes of the V1 to MT spectrum. Chief amongst these being the distribution of bandpass to lowpass temporal frequency filter distribution profile;whichwepostulateisafundamentalmanifestationofthewayspaceand time are mixed in the visual cortex. The remainder of this paper is organized as follows: Section (2) introduces the Gabor-Einstein wavelet. Section (3) presents computer simulations. Sec- tion(4)isthediscussion. Itincludesthefollowingsubsections: Subsection(4.1) introduces a simple framework for classifying and naming the components and hierarchical levels of receptive field models. Subsection (4.1) also identifies the Gabor-Einstein’splacewithinthislargerframeworkofreceptivefieldmodeling. Subsection (4.2) briefly discusses related work. Subsection (4.3) discusses some phenomena which cannot be explained by models such as the native Gabor- Einstein wavelet in isolation, i.e. models of single neuron receptive fields early in the visual pathway. Specifically, it discusses some higher order phenomena and non-linearities which necessarily arise from neuronal population and net- work interactions. Section (5) concludes the paper. 2 The Gabor-Einstein Wavelet In this section, we present the Gabor-Einstein wavelet. It has the following properties: • Like the Gabor function, it is a product of a gaussian envelope and a sinusoidal wave carrier. • ItdiffersfromtheGaborfunctioninthatitswavecarrierisarelativistically- invariantsincfunctionwhoseargumentistheenergy-momentumrelation. • Its gaussian envelope contains only spatial arguments. i.e. its spatial envelope is not time-dependent. • It has the minimum possible number of parameters. • Its fourier transform is the product of a mixed frequency gaussian and a temporal frequency step function. • Like the Gabor function, it generates a quasi-orthogonal basis. We define the Gabor-Einstein wavelet as follows, (cid:18) (x−x )2 (y−y )2(cid:19) G(t,x,y)=A exp − 0 − 0 sinc(ω t−u x−v y+φ), (5) 2σ2 2σ2 0 0 0 x y 10