Post on 11-Jan-2016
description
transcript
1
Segment I Past work: impact
Segment II Modeling the Auditory Pathway
Segment III future.cs@purdue: a personal view
Segment IV Q&A
Aditya P. MathurCS Department Colloquium
March 26, 2007
2
Research: Impact Coverage principle and the saturation effect
[Horgan.Mathur96] Microsoft quality gate criteria. Pioneered by Praerit Garg [MS’95] Guidant test quality assessment for medical devices
[recommendation accepted; yet to be implemented]
Software reliability estimation [Chen.Mathur.Rego 95; Krishnamurthy.Mathur 97]
Led to new approaches to software reliability modeling. [Gokhale.Trivedi 98; Singpurwalla.Wilson 99; Goševa-Popstojanova.Trivedi 01; Yacoub et al. 99; Cortellessa et al. 02; Mao.Deng 04]
3
Research firsts with ~No impact (so far!)
Testing on SIMD, Vector, MIMD architectures [joint with Choi, Galiano, Krauser, Rego. 88--92]
Feedback control of software test processes [joint with Cangussu, DeCarlo, Miller. 00--06]
LSL: A language for the specification of program auralization [Boardman.Mathur 94, 94-04]
4
Education: Impact Introduction to Microprocessors [80, 85, 89]
Drove curricula in almost every engineering college in India (including all the IITs).
Continues to be recommended mostly as a reference text in many Indian universities.
Over 100,000 students benefited from this book.
Foundations of Software Testing, Vol 1 [07], Vol 2 [08]
First comprehensive (text) book to present software testing and reliability as an integrated discipline with algorithms for test generation, assessment, and enhancement. Is driving testing curricula in CS/ECE departments.
5
Service: Impact Educational Information Processing System [BITS, Pilani 85]
Led a team of four faculty to design, develop, and deploy from scratch. In use even now(‘06) (code changed from Fortran IV to C!)
Purdue University Research Expertise (PURE) database [06]
Original idea: Dean Vitter. My contribution: Requirements analysis, design, testing, and management; interaction with all 10 colleges.
Over 85% of Purdue (WL) faculty in PURE. Expansion planned to other state universities; enhancement of feature set [with Luo Si]
Software Engineering Research Center (SERC) [94-00]
Started by Conte/Demillo ‘86-87. Led SERC recovery from six industrial members to 13 and
from two university members to four. Over $1.5 Million in research funds awarded to faculty.
6
Segment I Past work: impact
Segment II Modeling the Auditory Pathway
Segment III future.cs@purdue: a personal view
Segment IV Q&A
Aditya P. MathurCS Department Colloquium
March 26, 2007
7
Modeling the Auditory Pathway
Principle Investigator
Aditya Mathur
Graduate Student
Alok Bakshi, Industrial Engineering
Sponsor: National Science Foundation
Collaborators:
Nina Kraus: Hugh Knowles Professor
Sumit Dhar: Assistant Professor,
Department of Neurobiology and Physiology, Northwestern
Michael Heinz: Assistant Professor,
Speech, Language, and Hearing Sciences and Biomedical Engineering, Purdue
8
Objective
To construct and validate a model of the auditory
pathway that enables us to understand the impact of defects and auditory plasticity along the pathway in children with learning disabilities.
9
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
What is Brainstem Auditory Evoked Potential (BAEP)?
BAEP and children with learning disabilities
Existing modeling approaches versus our approach
Progress so far and the future
What is auditory pathway?
Trail
10http://www.iurc.montp.inserm.fr/cric/audition/english/audiometry/ex_ptw/e_pea2_ok.gif
http://www.iurc.montp.inserm.fr/cric/audition/english/audiometry/ex_ptw/voies_potentiel.jpg
Pitch discrimination (VCN)
Transport frequency, intensityInformation; rate encoding/temporal encoding
Azimuth, integration from both ears;ITD and ILD computation
Range,timing, intervals
Spatial map?, Spectral analysis
Sensory integration(e.g. head movement)
Comparison across sounds
What is (ascending) auditory pathway?
Medial geniculate body
Input for sound localization
Onset neurons
Gateway for AC
11
What is Brainstem Auditory Evoked Potential (BAEP)?
ABR [1.5-15ms]: Brainstem
MLR [25-50ms]: Upper brainstem and/or Auditory Cortex
ABR: Auditory Brainstem ResponseMLR: Middle Latency Response
Source: http://www.audiospeech.ubc.ca/haplab/aep.htm
Q: What is the effect of learning disability on ABR?
12
BAEP for normal and language impaired children
Normal children
Language impaired children
Source: Wible, Nicol, Kraus; Brain 2005.
6.2ms 7.2ms
V: lateral lemniscal input to inferior colliculus
Vn: dendritic processing in the inferior colliculus
Observation: Duration of V-Vn found to be more prolonged for children with learning problems than for normal children. Notice also the difference in the slope of V-Vn.
Stimulus: Synthesized /da/
13
BAEP for normal and language impaired children
Onset and formant structure of speech sounds in children
Normal children
Language impaired children
Source: Wible, Nicol, Kraus; Biological Psychology, 2004.
Stimulus: Train of /da/
FFR
FFR: Frequency Following Response
Observation: Mean V-Vn slope was smaller for children with language-based learning problems.
14
FFR for Musicians and Non-musicians
Source: Wong, Skoe, Russo, Dees, Kraus; Nature Neuroscience, 2007.
Stimulus: /mi1/, /mi2/, /mi3/F0: Stimulus fundamental frequency
Observation: Musicians showed more faithful representation of the f0 contour than non-musicians.
15
Importance of the BAEP
Neural activity in the auditory pathway, measured via the BAEP, seems to be a strong indicator of learning disabilities in children.
Auditory pathway is “tuned” by tonal experience.
16
Why model the auditory pathway?
BAEP is an external measurement (black box) of an internal activity.
Direct observation of internal activity is almost impossible in humans.
A validated model will allow direct observation of (simulated) internal activity and offer insights into the relationship between such activity and the BAEP.
This might lead to better diagnosis. Several other advantages too.
17
Research questions
How can neuro-computational models be used to encode, and mimic, the auditory neural behavior exhibited by children with learning disabilities?
How can such models be used to accurately predict the impact of treatments for learning impairments?
18
Existing approaches
Connectionist models: Surface and deep dyslexia: Hinton.Shallice’91, Plaut.Shallice’93 Spatial firing patterns: Nomoto’79
Phenomenological models [P-models]: Sound localization: Neti.Young.Schneider’93 Response to amplitude modulated tones: Nelson.Carney’04 Cochlear model: Kates’93 Speech recognition: Lee.Kim.Wong.Park’03
Simulation models: External ear to cochlear nucleus: Guérin.Bès.Jeannès.’03
19
Our approach
Simulation, system of systems, holistic, approach. Detailed, cellular. Explicit modeling of inherent anatomical and physiological
parallelism. Functionality used primarily for validation of the simulation
20
Our approach
P-model P-model P-model…….
Equations
Anatomy
Assumptions
Simulation
21
Progress
Auditory Nerve fiber model by Zhang et. al.
•Octopus Cell model by Levy et. al.
•Models of other cells being implemented
22
Bushy Cell (in Anteroventral Cochlear Nucleus)
Bushy Cell
Receives excitatory input from 1-20 AN fibers in the same frequency range
AN spikes
Time
Bushy Cell spikes
TimeLatent period
Preserves timing information for the computation of ITD.
23
Bushy Cell Model
Model [Rothman’93, Spirou’05]
• Has no dendrites and axon
• The soma is equipotential
• Receives 1-20 AN fibers with different characteristic frequency
Soma
24
K+ ion channel
http://personal.tmlp.com/Jimr57/textbook/chapter3/images/pro5.gif
Outside
Iext
IK INa IL
gK gNa gL
VK VNa VL
C
Inside
( At potential V )4ngg KK
hmgg NaNa3 tIVVgVVgVVg
dt
dVC extLLNaNaKK
m, n and h depend on V
Hodgkin Huxley Model
25
Segment I Past work: impact
Segment II Modeling the Auditory Pathway
Segment III future.cs@purdue: a personal view
Segment IV Q&A
Aditya P. MathurCS Department Colloquium
March 26, 2007
26
Vision as in the Strategic Plan [2003]
The faculty will be preeminent in creating and disseminating new knowledge on computing and communication. The department will prepare students to be leaders in computer science and its applications. Multidisciplinary activities that strengthen the impact of computation in other disciplines will play an essential role. …..
27
Vision as in the Strategic Plan [2003]
The department will be known for: Faculty who are recognized worldwide as leaders. They will set and
implement the national agenda for discovery and education in computer science.
A superior and diverse student body learning the values, vision, knowledge, and skills of computer science.
Graduates who go on to be faculty at highly ranked departments, researchers at internationally recognized labs, and leaders and innovators in industry and government.
Involvement and leadership in university institutes and centers that foster multidisciplinary research.
Collaboration with public and private enterprises in Indiana, the nation, and
the world.
28
Goals
2. Strengthen interdisciplinary research and educational programs.
3. Improve upon the existing research environment for faculty and students, in particular for tenure-track assistant professors.
4. Meet our implicit obligations to the state and the nation, in particular to our customers.
5. Maintain excellence where it already exists.
1. Offer a broader set of options to our undergraduate students.
29
Undergraduate Education
Tackle the declining enrollment problem: Revisit the undergraduate curriculum: should we change the
core? Should we offer alternate cores for different specializations? Create specializations: such as SE, Visualization, Security. Offer scoping into the MS program.
CPC sponsored undergraduate research projects. Some may lead to MS thesis.
Consider formalizing advisory role for the CPC in undergraduate curriculum design.
Strengthen the CS study abroad program.
Goal: Offer a broader set of options to our undergraduate students. Meet our implicit obligations to our customers.
30
Graduate Education
Enrollment Admissions MS and PhD programs. Interdisciplinary programs
Goal: Meet our implicit obligations to the state and the nation.
31
Faculty: Hiring
Look to the future of CS. Continue support for research in core areas but aim to
establish collaborative groups that are radically different in their perspective and aspirations.
Consider CS as a discipline essential to finding solutions to problems of key significance to humans: cancer and other diseases, large scale information processing, finance, health care, etc.
Aim at creating strengths in new and challenging areas while retaining current strength in core areas.
Goal: Strengthen interdisciplinary research and educational programs.
32
Faculty: Tenure
Reduce the uncertainty for an Assistant Professor. Focus (primarily) on scholarship; identify quantitative and
qualitative indicators of scholarship. Consider “quality” as a multi-dimensional attribute.
Identify and communicate ways of measuring impact/potential impact.
Create a “Tenure card” that aids in (accurate) self assessment.
Strengthen the third year review process.
Goal: Improve upon the existing research environment for faculty and students, in particular for tenure-track assistant professors.
33
Other programs/staff
Outreach programs All staff Facilities Corporate Partners Program Development
Goal: Maintain excellence where it exists.
34
Segment I Past work: impact
Segment II Modeling the Auditory Pathway
Segment III future.cs@purdue: a personal view
Segment IV Q&A
Aditya P. MathurCS Department Colloquium
March 26, 2007
Thanks!
35
Auditory Neuron Model
(Zhang et al., 2001)(Heinz et al., 2001)(Bruce et al., 2003)
36
Cochlear Nucleus
Consist of 13 types of cells Single cell responses differ based on
# of excitatory/inhibitory inputs Input waveform pattern
Onset response
Buildup response
Input tone
37
Octopus Cell
Octopus Cell
Receives excitatory input from 60-120 AN
fibers
AN discharge
rate
Time
Octopus Cell
discharge
rate
TimeLatent period
38
Schematic of a typical Octopus Cell
http://www.ship.edu/~cgboeree/neuron.gif
Representative Cell• Has four dendrites
• Receives 60 AN fibers with 1.4 - 4 kHz CF
•Majority of input from high SA fibers, medium SA fibers denoted
by superscript ‘m’
39
Octopus Cell Model Simplifications
Four dendrites replaced by a single cylinder Active axon lumped into soma Synaptic transmission delay taken as constant 0.5 ms Compartmental model employed with
15 equal length dendritic compartments 2 equal length somatic compartments
40
Octopus Cell Model
2 somatic compartments and 15 dendritic compartments modeled by the same circuit with different parameters
Different number of dendritic compartments depending on number of synapses with AN fibers
Soma Dendrite
41
Octopus Cell - Output
The output of the model implemented by Levy et. al. is compared against our model on the right side of the figure for a tone given at CF in figure A
Same comparison is made in figure B but with a tone of different intensity
42
Fusiform Cell
Fusiform Cell
Receives different inhibitory inputs from
DCN
AN discharge
rate
Time
Fusiform Cell
discharge
rate
TimeLatent period
43
Fusiform Cell Model Exhibit buildup and
pauser response and nonlinear voltage/current relationship
The model simulates the soma of fusiform cell with three K+ and two Na+ voltage dependent ion channels
The model doesn’t take into account the Calcium conductance
Doesn’t model the synaptic inputElectrical model of fusiform
cell
44
Fusiform Cell Model Characteristics
Predicts the electrophysiological properties of the fusiform cell by using basic Hodgkin-Huxley equations
Simulates the pauser and buildup response by virtue of intrinsic membrane properties
Synaptic organization of cells in DCN is not understood presently, so this model doesn’t model synapse and take direct current as the input instead
Doesn’t rule out the possibility of inhibitory inputs as the reason for pauser and buildup response
45
References Hiroyuki M.; Jay T.R.; John A.W. Comparison of algorithms for the
simulation of action potentials with stochastic sodium channels. Annals of Biomedical Engineering, 30:578–587, 2002.
Kim D.O.; Ghoshal S.; Khant S.L.; Parham K. A computational model with ionic conductances for the fusiform cell of the dorsal cochlear nucleus. The Journal of the Acoustical Society of America, 96:1501–1514, 1994.
Levy K.L.; Kipke D.R. A computational model of the cochlear nucleus octopus cell. The Journal of the Acoustical Society of America, 102:391–402, 1997.
Rothman J.S.; Young E.D.; Manis P.B. Convergence of auditory nerve fibers onto bushy cells in the ventral cochlear nucleus: Implications of a computational model. The Journal of Neurophysiology, 70:2562–2583, 1993.
Zhang X.;Heinz M.G.;Bruce I.C.; Carney L.H. A phenomenological model for the responses of auditory-nerve fibers: 1. nonlinear tuning with compression and suppression. The Journal of the Acoustical Society of America, 109:648–670, 2001.
46
References
• Drawing/image/animation from "Promenade around the cochlea" <www.cochlea.org> EDU website by R. Pujol et al., INSERM and University Montpellier
• Gunter E. and Raymond R. , The central Auditory System’ 1997
• Kraus N. et. al, 1996 Auditory Neurophysiologic Responses and Discrimination Deficits in Children with Learning Problems. Science Vol. 273. no. 5277, pp. 971 – 973
• Purves et al, Neuroscience 3rd edition• P. O. James, An introduction to physiology of hearing 2nd
edition• Tremblay K., 1997 Central auditory system plasticity:
generalization to novel stimuli following listening training. J Acoust Soc Am. 102(6):3762-73
47
Bushy Cell Model Characteristics
As the number and conductance of inputs is varied, the
full range of response seen in VCN Bushy cell are
reproduced
For inputs with low frequency(< 1 kHz), the model
shows stronger phase locking than AN fibers, thus
preserving the precise temporal information about the
acoustic stimuli
The model simulates the spherical bushy cell, but
doesn’t reproduce all characteristics of globular bushy
cell
48
Progress
Cochlea
UnknownConnection
KnownConnection
Nucleus Boundary
AN
Fib
ers
Cochlear Nucleus
49
Progress
Medial Superior Olive
Lateral Superior Olive
Medial Nucleus of the
Trapezoid Body
INFERIOR COLLICULUSNot
Implemented
Not Implemented
SUPERIOR OLIVARY COMPLEX
COCHLEAR NUCLEUS
Pyramidal Cell
Stellate Cell
Inter-Neurons
Bushy Cell
Octopus Cell
Fusiform Cell
Not Implemented
Implemented
50
Cochlear Nucleus
51
Bushy Cell
Slow low threshold potassium conductance
Some constants associated with Bushy cell:
Fast high threshold potassium conductance
Passive leakage conductance
Inhibitory synaptic conductance
52
Bushy Cell
The cell potential (V) is given by:
Where
Reverse potential for corresponding ions
Leakage potential
Membrane potential
53
Bushy Cell Model
Factor to scale rate constants to body temperature
General expression for scaling rate constants to temperature T
The three conductance mentioned earlier are given as:
54
Bushy Cell Model
Here themselves depend on voltage of soma V
Here denotes the arrival time for spike and synaptic
Conductance reaches its peak value of at time
Variation is given as:
Here and are given as:
55
Bushy Cell Model
56
Bushy Cell Model - Output
Response of Bushy cell for different number of input AN fibers (N), and synaptic conductance (A)
Fig. A shows the response of our implemented model for N=1 and A= 9.1, while the output obtained by Rothman et. al. is shown in D for same parameter.
57
Next Step
Implement the ILD circuit and find out the correlation between neural output and sound source (azimuth angle)
Cochlea
Cochlear Nucleus
SBC GBC MNTB MNTB
LSO
Cochlea
Cochlear Nucleus
GBC SBC
LSO
58
Next Step
Implement the ITD circuit and find out the correlation between neural output and sound source (azimuth angle)
Cochlea
Cochlear Nucleus
SBC GBC MNTB MNTB
MSO
Cochlea
Cochlear Nucleus
GBC SBC
MSO
LNTBLNTB
59
Next Step
Implement the dorsal cochlear nucleus neurons and find out the correlation between vertical angle and neural output in DCN region