Communities. So  he  started  his  lab  as an  assistant  professor  in  the  Department of  neurosurgery  at  Emory, just  across  the  town. And  before  coming  to  Emory, he  was  he  did his  post-doc  at  Columbia  with  Randy  burnout. Before  that  he  did  his  PhD. His  PhD  at  Berkeley  with  my  toys  are aware  he  was  focusing  on  auditory  cortex, but  then  during  the  postdoc, he  moved  on  to  somatosensory  cortex. So  he  is  working  on  all  different  sensory  cortices. And  yeah,  so  with  that, I  like  to  yeah. So  let's  give  him  another  big  applause.  Thank  you. All  right. I  hope  everybody  can  hear  me. Okay,  I've  got  my  my  God. We're  sharing  screen  on Zoom  and  let  me  know  if  there's  any  issues. It's  great  to  be  here. Thanks  especially  to  the  Ming for  assisting  me  here  today  and  being  so  welcoming. And  I  see  a  lot  of  familiar  faces  in the  audience  who  have  been  really, really  grateful  to  be  part  of  this  community. And  I  felt  very  supported  in my  relatively  short  time  here  in  Atlanta, looking  forward  to  working  together  with everybody  for  a  long  time. And  then  again,  my  name  is  Chris  Rogers. I'm  in  the  Department  of  neurosurgery at  Emory  University. We'll  start  out  by  saying, my  central  scientific  question  is  this. How  can  we  control  the  motion  of  our  bodies  in  order to  better  learn  about  and perceive  and  explore  the  world  around  us. You  can  see  an  example  of  that  in this  picture  that  I  have  on  my  cover  slide  of a  cat  clearly  very  interested  in something,  almost  certainly  food. And  you  can  tell  that  the  reason, the  reason  you  can  tell  it's  soaking  interested is  that  it's  looking  up  at  that  object. It's  directing  its  eyes  and  ears  and even  its  whiskers  towards that  thing  that  it  wants  to  learn  more  about. I  think  this  presents  the  brain with  twin  computational  challenges. The  first  is  the  motor  control  problem. How  do  you  move  these  sensors,  eyes,  ears, and  whiskers  to  efficiently  sample  the  space. And  second  is  the  signal  processing  problem. How  do  you  make  sense  of  that  sensory input  that's  coming  in? And  importantly,  how  do  you  use  it  to  direct your  sensors  at  the  next  time  step,  where  you  wanna  go. And  I'm  gonna,  I'm  gonna  try  to  argue  today  that  both  of these  challenges  I  think  are  really important  for  understanding  natural  behavior. And  they  also  provide  potentially a  new  and  interesting  way  to  think  about  brain  disorders. In  this  classic  experiments that  I'm  going  to  tell  you  about  his  motivation, Alfred  Yarbus  asked  humans subject  to  look  at  this  painting, the  painting  that  you  see  on  the  left. In  order  to  give  those  subjects  and  task, he  asked  him  questions  about the  painting  and  I'll  ask  those  questions  again  now. And  I  encourage  you  to  think about  the  answers  and  think  about what  you  find  yourself  doing as  you  try  to  answer  the  questions. The  first  thing  he  asked  them  was, how  old  are  the  people  in  this  painting? When  you  ask  them  that  these  white  lines show  you  where  the  eye  movements  of  the  subjects  land. So  they  looked  at  the  bases  in  the  painting. That's  how  they  made  that  decision. But  if  you  instead  ask  them, how  much  money  does  this  family  has, then  the  subject  is  found  themselves  looking at  their  clothing  and  their  possessions. That's  how  they  made  that. A  judge. I  like  this  one  when  you  ask  them  how  long  has this  family  been  separated  from  one  another? Then  the  subject's  eyes  follow  the  lines  of gaze  that  connect  the  faces  in  the  painting. What  this  tells  us  is  that  perception  isn't actually  just  about  what's out  there  in  the  external  world. It's  actually  about  what  we  internally  wants  to  know, what  we  choose  to  do  at any  given  time  and  what  we're trying  to  accomplish  in  doing. So. This  is  sometimes  called  active  sensing, but  they  don't  actually  like  that  term too  much  because  it  kind  of  suggests  something niche  are  really  specialized  like echolocation  and  fats  and  cool  things  like  that. Whereas  I  actually  want  to  argue that  almost  everything  that  you  do  is  you  go about  your  daily  business every  single  day  is  a  form  of  active  perception. So  for  instance,  let's  say you  come  down  to  your  kitchen  in the  morning  and  you're  trying  to  make  coffee, or  are  you  going  to  go  about  that? Well,  maybe  the  first  thing  you  do  is you  open  up  your  pantry  and  you scan  all  the  items  in there  to  find  the  bag  of  coffee  beans. So  that's  you  physically  moving  your  eyeballs  in  search of  a  particular  visual  stimulus  that  you're  looking  for. Then  you  see  the  coffee  grinder  on  the  shelf. You  want  to  pick  it  up,  what do  you  do?  You  reach  out  for  it. And  that's  not  a  pre-programmed  reach  to some  specified  physical  coordinate. What  you're  actually  doing, what  you're  literally  doing  is  you  reach  out until  you  feel  the  touch  of  the  object. And  that's  what  causes  you  to  terminate  their  reach because  you  know  that  you  have  have  gone, gotten  your  hand  to  where  it  needs  to  be. Grant  the  coffee  beans. You  push  the  button  and  you expected  here  that  really  awful  grinding  noise. That  expectation  is  built in  that  you  don't  even  necessarily think  about  it  until  one  day the  coffee  machine  is  broken, you  don't  hear  that  sound. And  that  absence  of  a  stimulus  is actually  quite  off-putting  and  surprising. That's  how  strong  you've  formed  at  auditory  expectation, event  and  motor  action. Then  finally,  if  you're  like  me, what  do  you  do  once  it's  grounds? You  hold  your  nose  up  to  it  and you  inhale  and  you  expect  the  smell. The  deans,  and  I  know  that  you  all  can imagine  that  just  looking  at  this  image. What  would  the  way  that  I  think  about these  things  is  that  you're  really  taking these  actions  because  you're  trying  to achieve  some  kind  of  sensory  result. And  so  on  this  slide  I  want  to  present  to kind  of  standard  typical  models  of  the  brain. So  in  this  first  one, we  tend  to  think  of  the  brain  in this  model  is  an  input-output  machine, a  feed  forward  processing  things. Sensory  stimuli  come  in and  elicit  some  kind  of  motor  response. And  I  think  Wikipedia  has this  low  information  from  the sense  organs  it's  collected  in  the  brain. The  brain  processes  that  raw  data. Finally,  on  the  basis  of  those  results that  generates  motor  response  patterns. So  a  very  kind  of  cereal  view  of  the  brain  works. And  I  think  a  lot  of  the  examples  I gave  on  the  previous  slide don't  really  fit  into  this  kind  of picture  of  what  the  brain  is  doing. But  we  do  see  this  model  very  frequently  in  neuroscience, and  indeed  in  any  paper  that  separates a  trial  structure  into a  sensory  stimulus  delay  period,  and  motor  response. They're  making  the  assumption that  in  this  case  we  have  the  brain. So  instead  I  favor  this  recurrent  or  closed  loop  model. In  this  model,  it's  recognized  that  the  brain  and the  world  are  locked  in  an  eternal  dialogue. They  are  there. The  brain  is  engaging  the  world  through  bodily  action. And  then  the  world  is  impinging  on the  brain  through  sensory  perception. And  neither  of  those  things  is  first, one  causes  the  other. It's  very  much  a  loop.  And  of  course  I'm  not the  first-person  far  from the  first-person  to  make  this  claim. There's  John  Dewey  in  18  96  circuit, the  motor  response  determines that  stimulus  just  as  truly as  the  stimulus  determines  the  movement. I  think  as  engineers,  we  also  recognize  that, that  feedback  of  this  form, perceptual  feedback  in  this  case, is  essential  for  any, any  real  system  to  have smooth  and  coordinated  interaction with  the  world  around  us,  with  a  real-world. So  the  goal  of  my  lab  is  really  to understand  how  computations  in the  brain  might  regulate this  feedback  loop  to  enable  naturally. Alright,  so  here's  what  we're  going  to  talk  about  today. First,  we'll  talk  about  how  mice  can  use  their  sense  of touch  in  order  to  recognize  objects  around  them, as  you  see  the  wrap  here  doing  on the  way  that  results  in  this  part, all  come  from  my  postdoctoral  research  at  Columbia. It's  all  published,  So  I'm  going  to  try to  go  very  quickly  through  it. It  made  me  emphasize  some  of  the  things that  didn't  really  make  it  into  the  paper. Some  of  the  kind  of  unexpected  stuff which  motivates  everything  that  I'm  doing  in  my  own  lab. And  that's  what  we're  going  to  talk  about  in  part to  a  way  of  studying  active  sensing. And  in  this  case,  in  the  auditory  system, which  is  critical  for  both  humans  and  animals. So  let's  start  off  by  discussing  what  I learned  about  active  sensing  in  the  whisker  system. This  is  really  cool. Video  of  an  act  has  touched  lab  that  shows  how  ras  can identify  objects  by  coordinated  motion  of  their  head, their  whiskers,  our  body  onto it  around  these  objects  in  order  to  recognize  them. Super  complex  behavior. We've  gotten  motion  of  multiple  body  parts all  working  together, tightly  regulated  by  sensory  feedback, in  this  case  from  the  whiskers. So  how  can  we  go  about  understanding something  that's  still  complex? So  here's  the  approach  that  I  took  to understanding  sensory  processing  by  whiskers. Now,  this  is  a  work  that  I  did  as a  postdoc  and  Randy  Bruno  is  lab  at  Columbia. We  started  out  by  developing a  new  behavioral  task  for  mice. These  mice  are  head  fixed. Do  you  see  their  head  fixed  in  this  little  house  here? And  their  job  is  to  use  their  whiskers to  identify  which  of these  two  orange  colored  shapes  they're  being  presented with  on  any  given  trial,  concave  or  convex. And  the  way  they  indicate  to  us  what  their  decision  is, is  there  going  to  lick  either  the  left  windpipe  or the  right-click  pipe  in  front  of their  face  to  register  their  decision. I  want  to  just  quickly  mentioned  that  I  did  this  with the  invaluable  assistance  of two  research  technicians  and  Randy's  lab, Christina  Hill  and  agreement. On  the  right  here  will  be  a  quick  low  resolution video  of  what  this  behavior  looks  like  in  action. This  is  all  done  in  the  dark so  that  the  nice  happy  is  there  whiskers, you're  seeing  infrared  illumination here  which  the  mice,  cats. On  any  given  trial,  we  rotate either  a  concave  or  convex  shape  into  position. We  bring  it  into  the  mouse  is  Vickery  field. And  as  you  see  here,  the  mice  actively  and voluntarily  move  their  whiskers  back-and-forth  to make  contact  with  the  shape  in  order to  discriminate  what  is  curvature. So  because  this  mouse's  head  fixed, which  is  an  important  design  decision  that  I made  that  has  repercussions  we'll  talk  about  later. But  first  let's  talk  about the  advantages  of  a  head  fixed  mouth so  we  can  get  really  precise  video  of  the  whiskers. In  this  case,  we're  very  zoomed  in  now the  mouse's  head  is  out  of  frame  down  here. We  take  this  video  at  200  frames  per  second. And  what  we're  able  to  do  is  to track  which  whisker  is  interacting  with  the  shape. Now,  normally  mice  have so  many  Westerners  that  they  would  be overlapping  and  a  view  like  this  and there'd  be  no  way  to  tell  which  one  is  which. But  what  I've  done  is  to  trim  off  all  the  whiskers other  than  the  ones  in  the  middle  row,  the  C  row. These  whiskers  are  in  the  same  and  all  nice. And  so  they  actually  have  names. C1  is  the  lung  blue  one  here, C2  is  tracked  in  green  and  C3 is  the  sharp  one  track  and  read  in  the  front. And  the  trimming  off the  whiskers  makes  it  more  difficult for  the  mice  to  do  this  task. But  it  also  enables  me  to  precisely  identifying which  whisker  is  which  and every  single  touch  that  they  make  on  the  shape. And  that's  what's  shown  in  this  video. This  is  an  example  clip  of high-speed  video  during  the  shape  discrimination  task. There's  been  slowed  down  seven  times  from  real  life. So  this  is  a  really  fast  behavior would  be  happening  seven  times  faster. And  I  built  some  custom  software  tools  to  help  us  do a  little  bit  better  job  tracking the  whiskers  as  you  see  here. So  we  can  track  them  all  the  way  out to  their  very  fine  tips, even  when  they  are  rapidly  moving or  partially  obscured  are  in  contact  with  the  shape. And  that  turns  out  to  be  really  important for  the  analysis  that  I'm  going  to  show  you because  we  need  to  be  able  to detect  every  contact  so  that  they  make, even  though  those  contacts  are  very  brief on  very  quick  and  just  with  a  fine  tip,  the  whisker. But  in  doing  so  that  then  allows  me  to  ask,  okay, how  do  mice  tingle  of  these  sensory  inputs, all  these  individual  contacts, and  put  them  together  in  order  to  make a  decision  about  what  shaken  it. So  it  makes  sense  of  these  data. We  developed  a  new  analytical  technique that  I  call  behavioral  decoding. And  I  did  this  in  collaboration  with  hormone  yoga,  yoga, who  is  a  post-doc  in  the  theory  Center  at Columbia  and  now  on  the  faculty  job  market. What  I  did  briefly  was  to measure  everything  that  I  could  run  that  video. So  the  position  of  the  whiskers, the  time  of  the  contacts, the  position  of  the  contacts,  you  know, the  angle  between  the  whiskers, the  duration  of  everything that  you  can  measure  from  video. Take  all  of  that  data  and  put  it  into  these  decoders. Decoders  job  is  to  predict what  was  the  stimulus  on  the  trial. And  importantly,  what  stimulus  did the  mouse  think  that  it  was  concave  or  convex? So  the  reason  to  do  that  is  that  we  can  now tell  exactly  which  of  those  features  that  I  measured, duration,  velocity,  whatever,  which  of  those  features actually  mattered  for  the  mouse's  choice in  which  didn't  matter  at  all. And  just  as  a  quick  spoiler, it  turned  out  that  very  little  of  the  things  that could've  used  actually  matter  to  their  decision. And  what  they  chose  to  pay attention  to  was  in  fact  quite  simple. You  go,  alright,  so  I'll  just  cut  right  to the  punch  line  and  tell  you  how that  makes  discriminated  shape. It  turns  out  that  the  key  variable  is  which  whisker touched  the  shape  on  any  given  trial,  and  how  many  times? So  the  speed  is the  contact  or  the  angle  or  anything  else. Here's  how  we  measure  that. So  what  I'm  plotting  here  is  weights  that  the  decoder assigns  to  each  whisker  making  contact  on  the  shape. So  positive  weights  means  that  the  decoder  that  the  shape is  convex  and  negative  weights  means  that  the  decoder, thanks  for  the  shape  is  concentrated. And  so  even  though  all  three  of the  whiskers  can  patch  both  of  the  shapes, there  are  subtle  differences in  the  relative  proportion  of those  contacts  and  that's  what the  decoder  is  picking  up  on. It  says,  all  things  being  equal, each  additional  C1  contact means  the  shape  is  more  likely  to  be  convex. Traditional  C3  contact  means the  shape  is  more  likely  to  be. More  abstractly  what  that  means, it's  the  mice  must  be  comparing  information across  whiskers  in  order  to  make  their  decision. Comparing  the  number  of  C1  contacts,  the  number  of  C3. Now  this  slide  describes  the  dataset  that  I  collected  and barrel  cortex  of  mice  were  performing  this  task. That  is  the  brain  region  that  processes input  from  the  whiskers  and  the  cortex. On  the  left  is  a  cross-sectional  view of  the  layers  of  cortex. The  green  thing  is  the  electrode  that  I  used  64  sites, so  we're  able  to  record  30  or  so  neurons  simultaneously. The  way  these  experiments  work  is  that  I  can  put  it  into whatever  cortical  column  I  want  to  record  from  that  day. Cerebral  cortex  is  well-known  for  having one  vertical  column  for  each  whisker  on  the  face. And  so  what  I  was  able  to  do  is  record  first  from  C1, then  from  C2,  then  from  S3,  and  so  on. And  we  recorded  about  1,000  neurons  total  in  this  way, not  all  in  the  same  session,  but  combined  overall. Alright,  so  here's  an  example  of  what the  data  look  like  using  this  technique. So  on  the  left  you'll  see  a  video like  what  you've  seen  before. And  on  the  right  is  the  spike  trains from  each  simultaneously  recorded. All  right,  so  I'll  go  ahead  and  start  that.  Now. The  audio  track  will  be  playing the  spikes  from  one  example  neuron. In  this  particular  neuron  I've chosen  because  it  mainly  responds to  contacts  made  by  the  C1  or  green  whisker. In  another  trial  for  the  same  neuron. Then  I'm  going  to  show  you  a  different  neuron. And  this  one  really  responds  more  to  whisker  motion. It  can  be  hard  to  tell  what  the  neurons  are responding  to  just  from  a  video, like  it's  an  author  you  from  analyses  I  did. But  I  really  like  to  show  this  video because  I  loved  the  spin  of  data, because  of  the  fine  structure that's  clearly  evident  in  this, in  these  patterns  of  spikes. I  think  that  everything  that  we feel  and  do  and  think  about must  be  included  at  some  level  in patterns  of  spikes  in  our  brains. But  at  the  same  time  it's  a, it's  a  big  challenge  because  I  think  a  lot  of us  are  facing  this  in  modern  neuroscience. How  do  we  take  this  kind of  fine-scale  behavioral  data  and relate  it  to  this  kind  of  high-dimensional  neural  data that  we  take  at  the  same  time,  it's  a  big  challenge. So  let  me  show  you  one  simple  analysis  that  we  did, which  is  kind  of  the  punchline  of  the  story  in  the  paper. And  then  we  focused  here  on  one  feature that  I  think  matters  the  most  for  shape  discrimination. And  that's  again  which  whisker made  contact  with  the  shape. So  this  is  data  from  one  example  neuron. And  I'm  showing  you  the  spikes  that  it  emitted  in response  to  contacts  made  by  the  C1  whisker, the  C2  whisker,  or  the  S3  here. And  just  by  eye  you  can  see  that  this  neuron  fires  a  lot more  when  the  C1  whisker  touches the  shape  than  it  does  when  the  C2, C3  whisker  Tetris  shapes  then  affect  it's  pretty similar  regardless  of  the  identity of  the  shape  that  it's  actually  touching. It  turns  out  this  is  not  just  a  feature  of this  one  example  on  their  own  that  I've  chosen. This  is  actually  true  in  a  broad  swath  of  neurons  that I've  recorded  across  all  different  layers, excitatory  and  inhibitory,  and  even  in  different  parts of  the  barrel  cortex  that  are supposed  to  be  tuned  for  other  whiskers. We  quite  commonly  found  that neurons  way  more  often  than  you  would  expect where  tuned  to  respond  to  contacts  made  by  the  C1  whisker instead  of  the  C2  received three  whisker  even  when  they  were  located  in, let's  say  the  C2,  C3  columns. Now,  as  a  control  for  this, we  did  a  lot  of  controls  in  the  paper, but  I  think  the  most  compelling  one  is  this. We  trained  mice  on  a  different  task using  the  same  shapes. These  other  mice  were  exposed  to  the  same  exact  shapes, but  they  didn't  have  to  care  about  whether  they were  concave  or  convex. We  call  that  a  detection,  test. It  in  those  mice  we  see  exactly  what  we  would  expect, which  is  that  neurons  on  average  across the  population  respond  equally  to  c1,  c2,  and  c3. Whiskers. On  their  rate  is  what  we  would  expect  from  him. Decades  of  knowledge  by  the  barrel  cortex. This  on  the  left  was  a  very unexpected  finding  on  the  Bureau  predicts is  really  Nazca  Plate  favorites  like  this  and  prefer, prefer  one  whisker  over another  to  show  this  kind  of  bias  corresponds. What  does  that  mean?  And  why  would  they  do  that? And  why  would  it  be  only  during this  one  particular  tasks  that  we've  trained? And  I  think  that  we  can  start  to  answer that  question  if  you  think  about the  behavioral  decoding  results. So  we  now  behavioral  decoder didn't  know  anything  about  spikes, didn't  know  anything  about  neurons. And  its  job  is  just  to  tell  convex  and  concave  shapes. The  way  that  it  did  that  was  to compare  the  number  of  contacts made  by  C1  whisker  versus  C3  whisker. What  we  see  in  the  neural  data  is  compatible  with that  relative  enhancement  of responses  to  the  C1  whisker  and a  relative  suppression  of  responses  to  the  S3  list  here. So  one  way  of  thinking  about this  is  that  the  weight  identify convex  shapes  is  to  look  for C1  contacts  and  subtract  out  s3  contact. We  think  that's  what  the  neurons  have been  through  the  process  of  learning  this  task. That's  how  the  neurons  have  been reconfigured  or  reformat  it  in  order  to emphasize  that  feature  that's important  for  detecting  convex  shapes. Alright,  so  that  was  the  punchline  of  the  paper. My  son,  a  specific  strategy  for  telling  shapes  apart. That  rule  wasn't  obvious  in  advance. We  were  able  to. In  fact,  it's  not  at  all  the  rule  that  we thought  the  mice  would  do  like  at  all. But  we  were  able  to  quantify  and  discover  that  with this  behavioral  decoding  approach that  they  were  comparing  across  my  scarce. And  then  also  we  found  a  result  in the  somatosensory  cortex  that  was  at least  compatible  with  that  kind  of  strategy, that  kind  of  computation  comparison  between  the  whiskers. So  there's  a  lot  more  controls  and  analyses  and experiments  in  the  paper  that  talks  about  these  findings. But  I  just  want  to  tell  you  today about  some  of  the  things  that  I  was  actually  pretty confused  about  after  the  end  of  all  this  analysis, which  is,  first  of  all, why  did  this  strategy  even  work  for  them? I  said  all  like,  why  is  this the  way  that  they  chose  to  do  the  task? We  actually  design  this  experiment  thinking  it  would  be something  very  different  than  mice  would,  say, compare  the  timing  of  the  contacts  across  whiskers  or compare  the  relative  angles of  the  context  that  they're  making. In  fact,  we  tried  to  set  it  up so  that  all  three  whiskers  can hit  the  shapes  equally  as  you  see  in  these  images. So  how  are  the  mice  somehow  like hitting  one  shape  more  often  with  one  whisker? How's  that  even  possible? So  here's  what  I  think. This  isn't  a  sensory  task  at  all, even  though  that's  how  we  designed  it  to  be. The  mice  have,  are  treating  this  as  a  motor  task. Mice  have  learned  some  really  clever  way of  moving  their  whiskers, something  that's  really  tuned for  these  particular  shapes. Such  that  with  this particular  motor  strategy  that  they've  learned, the  sensory  readout  becomes  really  easy. It's  just  comparing  C1  contacts  versus  C3  contracts. So  in  other  words,  they've  moved  to the  complete  computational  burden  from  the  sensory  side, which  is  what  we  expect  it  to  the, to  the  motor  side. That's  what  we  think  is  happening. They'd  learned  a  complex  motor  policy  that permits  simple  sensory  or  beat  up. You  could  have  imagined  they  would  do  the  opposite  thing, and  that  was  our  original  hypothesis. They  would  use  a  simple  motor  policy, let's  just  say  moving  all  the  whiskers together  and  a  stereotype  fashion. And  then  do  some  really  complicated  signal  processing on  the  input  that  comes  in  to  pull  out  like, let's  say  the  differences  in  timing  or  angle  between the  contexts.  I  think  that's  it. That's  a  completely  plausible  way  that they  could  have  done  the  task. But  it  turns  out  that  none  of  the  mice  did  that  at  all. What  I  thought  about  this  more, I  think  this  is  actually  kind  of  similar  to the  way  that  we  tell  shapes  too. So  when  I  want  to  know  what  shape, let's  say  a  mug  is. The  way  that  I  can  tell  that  shape is  reaching  out  and  grasping  it. So  I'm  not  scanning  this  mug  like  lidar  and  building  up some  3D  model  of  like exactly  the  location  of  where  all  my  fingertips  are. No,  I'm  just  grasping  it. And  then  I  can  introspect  about what  position  is  my  hand  in  right  now  as  I'm  holding  it, as  generally  curves,  therefore, this  must  be  a  gently  curved  objects. So  I  think  of  this  as  a  kind of  an  embodied  computation  and  a  way  of changing  interrogation  about  the  world to  introspection  about  the  body. What  position  is  my  hand? And  when  I  grasp  of  this  object,  the  thing, the  reason  that's  useful  is  because not  only  does  it  tell  us  what  the  shape  of  the  object  is, but  it  also  tells  us  how  we  can  interact  with  bodies. So  here's  a  couple  of  pieces  of  evidence  that support  that  the  fact  that  that might  be  what  the  mice  are  doing. First  of  all,  my  smart  or  really  precise, skilled  and  repeatable  motor strategy  for  investigating  shapes. But  I'm  showing  you  here  is  the  example whisking  on  11  different  trials  from 11  different  mice  with  time  on  the  x-axis and  the  mean  angular  position  of the  whiskers  on  the  y-axis. You  can  see  that  even  though  these  mice  are  all  doing the  same  tasks  on  these  trials, like  the  whisking  is  quite  different. Now  if  I  show  you  more  data  from  the  same  mice, such  that  all  the  data  within each  column  is  data  from  one  particular  mouse. You  can  see  that  there's  a  lot of  consistency  within  mice, even  though  there's  a  lot  of  variability  between  groups. So  for  instance,  mouse  11, the  whisker  laugh  early  on in  the  trial  and  it  tapers  off. Whereas  mouse  one  is  the  opposite. It  starts  up,  it  doesn't  whisk  it  all  in the  beginning  and  it  lists  aggressively  later  on. It's  like  each  mouse  has  learned this  really  precise  exploratory  motion. And  then  it  chooses  deploy  that  over  and  over  and  over again  on  every  trial  in  order to  get  the  information  that  it  needs. The  other  thing  that  we  found  in the  data  is  a  neural  observations. So  I  told  you  about  the  response  to  contacts, but  we  have  a  lot  more  data  than  just  that. So  we  were  able  to  figure out  how  each  one  of  the  neurons  we recorded  fired  as  a  function of  the  sensory  information  that  it, that  is  the  contacts  made  by multiple  whiskers  potentially  in  complex  patterns. Also  the  position  of  the  whiskers  at  any  given  time. And  then  what  I  call  a  cognitive  variables. So  for  instance,  the  choice  of  the  animal  is  going  to make  the  history  of  outcomes  and  rewards. We  can  put  all  of  this  information into  a  generalized  linear  model, which  is  basically  a  form  of  regression. And  ask  how  that  affects  the  firing at  each  individual  neuron. In  our  dataset,  we  found  a  lot  of  stuff  here. But  the  thing  that  I  want  to  emphasize is  that  the  neurons  in somatosensory  cortex  really  seem  to  care  a  lot more  about  what  I  would  call motor  signals  than  anything  else. So  they  strongly  encoded  a  position  of the  whiskers  on  a  moment  by  moment  basis. And  they  also  encoded  some  of  these  cognitive  variables. And  if  anything,  the  variable  that  is  sort  of the  classical  feature  that  we would  expect  to  drive  neurons  in  this  brain  area, which  is  whisker  touch. If  anything,  that  was  like  the  least  represented  out of  all  the  variables  that  we  considered  here. So  motor  signals  as many  other  people  now  and  have  shown  as  well, it  seemed  to  be  really  the  primary  signal  that  is driving  this  sensory  areas,  somatosensory  cortex. So  just  to  summarize  this  part  of  part  one, we  showed  that  these  points  are  what  the  paper's  about, how  mice  discriminate  shape  and  how  neurons  reconfigure their  responses  to  support  that  computation. But  some  of  these  observations  still kinda  really  not  any  after  that,  the  project  was  over. Unexpectedly  find  that  mice  are  using this  complicated  motor  policy that  simplify  the  sensory  readout. And  most  of  the  activity  in  sensory  cortex  was seemed  to  be  more  motor  related  than  insensitive. And  in  that  paper  I  wasn't able  to  really  wasn't  set  up  in  this  way  to really  seal  the  deal  about  what  was going  on  here  with  these  motor  signals. That's  why  they  decided  that  it  was  kinda  left  with a  feeling  of  maybe  missing  the  biggest  part  of  the  story. In  my  own  lab,  I  want  to  refocus  the  experiments  that  I'm doing  to  really  directly addressed  these,  these  questions. How  can  the  brain  control  the  body  efficiently, better  perceive  the  world? How  does  sensory  and  motor  brain  regions interact  to  support  something  like  this? And  how  can  we  even  understand this  complicated  system  to  begin  with? So  that  brings  me  to  part  two,  which  is  Rehman. I  tried  to  convince  you  that  we  could  actually  use the  auditory  system  has a  better  way  to  answer  these  kinds  of  questions. Alright,  so  why  auditoriums? I  know  this  audience  is  particularly  visually  focused. And  the  reason  I  say  that  is  because  you're  all  primates. And  so  visual  primacy  is  really built  into  the  way  that  we  think  about  the  world. In  a  way  that  probably  isn't  for  mice  who  are, who  are  much  more  olfactory  driven. But  auditory  when  it  comes  to  auditory, we're  on  equal  footing  with  the  mice. It's  a  really,  may  not  be  our  first  sense  that  we  go  to, but  it's  a  really  a  crucial  sense  for  both  of  us, especially  for  things  like  social  communication, language  and  records,  and  for identifying  the  approaching  danger  from  afar. I  think  about  this  every  time  I'm walking  down  a  street  or biking  and  I  hear  a  car  coming  from  behind  me. That's  how  you  can  first  pelvis  that  is  approaching  you. A  lot  of  people  have  talked  about the  role  of  motor  movements, eye  movements  and  vision, sniffing  and  olfaction,  reaching  out  during  touch. But  the  question  of how  to  learn  movements  influence  hearing, I  think  is  much  less  looked  at. I  mean,  that's  that's  something  that  I  want to  talk  about  today  for  the  rest  of  the  talk. So  here's  a  fun  video  to  show  you  an  example  of  that. This  is  a  red  fox  hunting  mice that  are  hidden  beneath  the  snow. The  mice  were  making little  scratching  sounds  like  that  one. And  the  muscles  are  inside the  fox  is  going  to  listen  very carefully  and  that's  how  it's going  to  identify  where  that  mouse. And  once  it  knows  it  attacks. Landing  is  really  funny. But  I  couldn't  think  that  I  want  to point  out  about  this  video is  what  happens  when  the  fox is  listening  really  carefully. It  doesn't  hold  its  head  perfectly  still, which  is  what  you  might  expect. Instead,  I'll  show  that  part  again. Instead  it  *****  its  head  first  left,  and  then  right. So  it's  actually  using  head  motion  as a  way  to  improve  its  audit  brand  perception. And  I  think  this  might explain  something  that  a  lot  of  people  have  noticed  now. And  we  had  Chris  Neil  visit  at the  seminar  series  recently  in  talk  a  lot  about  this. A  lot  of  people  are  observing  motion  related, movement  related  signals  in areas  that  we've  thought  of  as  sunscreen. And  one  possibility  that  i'd, I'd  like  to  explore  is  that the  reason  that's  the  case  is because  in  natural  behavior, when  you  move  the  head, you're  moving  the  ears  and  the  eyes  and  everything  else. And  so  of  course,  that's  going  to  change the  sensory  signal  that  you  get  as  a  result. And  it  would  make  sense  that  sensory  areas, we'd  need  information  about  that  motor  action  in order  to  correctly  interpret the  information  that's  coming  in. So  I  think  that  I hypothesize  that  this  is  not  just  a  kind  of a  broadcasted  signal  that's  going  everywhere, but  it's  actually  potentially a  specific  signal  that's  giving  these  sensory  areas the  information  they  need  in  order  to  interpret the  inputs  that  they're  receiving  at  that  time. But  all  of  this  might  be  kind  of  mixed  up  in the  head  fixed  system  that  so  many  of  us  have  relied  on because  the  mice  can't  actually move  their  heads  in  that  setup. I  want  to  mention  to  you  ever  since I  started  thinking  about  this, I  noticed  this  again  when  I'm  on  my  bike  or walking  and  I  hear  the  car  approaching  from  behind  me, I  actually  find  myself  turning my  head  a  little  bit  and  that  really  allows me  to  better  identify the  angle  that  the  car  is  approaching it  and  whether  I'm  saved for  whether  I  need  to  jump  out  of  the  way. Even  though  we  don't  necessarily  do exactly  what  that  fox  did  in  this  video, I  think  head  motion  is  really  important  way  for  us to  enhance  our  auditory  perception  as  well. So  here's  what  we're  going  to  look  at  that  I will  not  be  bringing  those beautiful  faxes  into  the  laboratory. That's  not  what  we're  gonna  do  here. I'll  be  studying  this  in  mice  because  of the  extensive  toolkit  available for  interrogating  define  neural circuitry  in  that  species. And  also  the  availability  of  a  lot  of genetic  models  for  different  kinds of  brain  disorders  and  diseases. This  is  the  Hebrew  that  we've  developed  to  study  this, my  vision  is  to  move  away  from  head fixed  tasks  and  set  the  mouse  screen. We  put  the  mouse  in  this  large  octagonal  arena surrounded  by  speakers  and  all  sides. There's  little  plastic  dividers  that  divide  up this  octagon  into  eight  little  chambers. The  task  is  loosely  structured  into  trials. So  on  any  given  trial, what  happens  that  one  of  these  speakers, that  goal  speaker  starts  playing  sounds? The  most  his  job  is  to  figure  out where  those  sounds  are  coming  from. So  over  to  that  speaker,  enter  that  chamber, go  to  that  speaker  and  its  nose  into  the  little  nose  poke below  the  speaker  and  get  a  reward,  right? So  that's  one  trial  and  that  just  repeats for  the  rest  of  the  session. But  of  course,  with  a  different  randomly chosen  goals  speaker  each  time. It's  very  loosely  structured, it's  very  self  paced. The  mouse  doesn't  have  to  come  back  to  the  center and  do  anything  to  initiate  the  trial. Xth  trial  just  starts  whenever the  most  houses  as  long  as  it  wants  to  make  its  decision, no  matter  how  long  it  takes  her, how  many  ports  it  incorrectly  posts  before  it  gets  there. One  last  detail  about  the  way  this  task  works. It's  not  a  continuous  sound, like  it's  a  stream  of  intermittent  noise  bursts. So  each  noise  burst  is  ten  milliseconds  long. And  then  they  are  irregularly  presented. We  can  do  them  slow,  fast  however  you  want. But  it's  usually  in  this  irregular structure,  something  like  that. Alright,  so  here's  the  key  question  about  this  task. What  happens  when  the  sounds  start  playing? How  do  they  know  where  they're  coming  from? What's  speaker  is  the  Golf  where  the  sounds  are  coming. If  you  pop  open  any  neuroscience  textbook, you  can  read  about  the  beautiful  circuitry in  the  brainstem, the  midbrain,  the  animals  have, all  of  us  have  to  identify  the  location  of  sound. So  these  circuits,  which  if  you  haven't  read these  papers  that  I  encourage  you  to do  it  because  it's  very  beautiful  work. These  circuits  are  designed  to  pull  out very  small  differences  in the  exact  timing  or  level  between  the  left  and  right  ear, the  sound  will  arrive  slightly  sooner  at  the  rate, or  maybe  slightly  larger  on the  higher  amplitude  on  the  rate. And  the  circuitry  will compute  those  differences  and  in  that  way, identify  what  the  location  of  the  sound  must  be. If  you  look  at  those  foundational  papers, you'll  see  that  the  authors  had  to  try pretty  hard  to  make  the  animals  hold  still. Doing  these  experiments.  Maybe  they're  anesthetized, maybe  they're  restrained  and  bolted  down. Maybe  they're  trained  to  hold  still. Maybe  they  just  need  the  sounds  so  brief  that  there  was no  possibility  for  the  animal to  move  its  head  during  this  time. If  you  think  about  it,  of  course  they  had  to  do  that because  you  would  never  actually  localize. It's  not  in  this  way.  Hold  perfectly  still. Instead,  you  would  move  your  head, you  would  move  your  body  in  order  to figure  out  where  that  sound  is  coming  from. So  for  instance,  here's  what  we hypothesize  that  the  mice  will  do. If  they  want  to  know  where  the  sound  is  coming  from, they  can  scan  their  head  left  and  right. They  hear  more  sounds  on  the  right. Maybe  they  turn  their  head  to  the  right. Until  the  sounds  are  roughly  equal  on  both  sides. These  a  motor  strategy  like  that. One  cool  thing  that  I  think  about  this  strategy, if  it's,  if  it's  what  we  actually  observe, is  that  it  turns  the  sensory  problem  of auditory  localization  into  something  that's  one  and the  same  as  the  motor  problem  of  orienting  with  it with  the  goals  speaker  in  order to  go  over  and  collect  your  reward. So  it's  potentially  another  example  of  embodied  company. So  I'm  gonna  show  you  a  video  that  the  task, which  is  a  little  bit  unexpected. So  go  ahead  and  start  this. Now,  this  one  is  not  sped  up  or  slowed  down  at  all. This  is  real  time. On  each  trial,  the  mount, one  of  the  speakers  starts  playing  sound as  indicated  by  the  yellow  notes  here. There's  no  audio  in  this  video. And  what  you  can  see  is  that  the  mouse  just  had to  figure  out  where  that's  coming  from  and  go  over  it, stick  its  nose  and  in  order  to  get  reward. Now  a  couple  of  observations  here. So  first  of  all,  behavior  is  really  this  animal  is not  standing  still  collecting  information. And  then  like  eliciting  some,  some  behavioral  response. Know  if  anything,  it's  constantly  in  motion. And  that  motion  is  at  best  modulated  by the  sensory  input  that  the  mouse  is receiving  at  any  given  moment. The  other  thing  is,  what  about the  starting  motion  that  you  see kinda  looks  like  maybe  the  mouse  is  like  sampling  each, each  little  room  in  order  to  know  if  that's  good. The  sound  is  coming  from. When  I  saw  this  and  when  I  read  a  book  by  Jennifer  grow, I  realized  that  with  this  reminds  me  of  is when  you  have  that  low  battery sound  on  your  smoke  alarms  somewhere  in  your  house, but  you  don't  know  what  room  it  is. Again,  you  don't  like  kinda  sit  still  and calculate  exactly  what  angle  that  sound  is  coming  from. You  go,  you  go  around  and  you  stick  your  head  in each  room  line  at  a  time  and  you ask  if  the  sound  is  louder  in  there. That's  a  really  simple  and an  effective  and  adaptive  strategy for  solving  that  problem. We  also  see  mice  use other  kinds  of  strategies  too  that  are  more similar  to  the  head  scanning  motion that  I  mentioned  on  the  previous  slide. There  is  some  variability  in the  strategy  they  use and  it  seems  to  be  highly  effective. But  the  structure  of  this  environment, how  long  these  dividers  are,  things  like  that? The  other  thing  that  I  just  think  it's  kinda cool  is  that  the  mice  really  like to  mess  around  in this  box  that  for  lack  of  a  better  word. So  when  they're  when  they  don't  feel  like  working, they  do  something  that  I  think  can really  only  be  described  as  play. Which  I  think  is  actually the  most  exciting  motor  aspect  of  this  tests. Like  what  they're  doing  here  is  way  more  cool than  what  we've  asked  them  to  do. He's  doing  this  balancing  acts  on  top  of  these  dividers, which  are  kinda  rickety  and  wobbling  around. And  he's  using  his  whiskers  there  to decide  where  to  place  this  next  step. I  think  I  think  he's  also  potentially  using  his  tail as  an  active  stabilizer like  kinda  like  a  fifth  limb  here. And  I  think  that  allowing the  mouse  to  move  freely  really  opens  the  door  to  a  lot of  these  kinds  of  more  naturalistic and  maybe  really  much  more  interesting on  motor  abilities  that  are  quite capable  of  and  maybe better  at  than  some  of  the  laboratory  tasks. Typically  challenge  them  on. Not  something  that  like  to explore  a  little  more  in  the  future. And  especially  in  the  context  of motor  cortex  or  other  motor  regions that  might  be  controlling  this  kind  of  behavior. And  we'll  just  skip  too  many  fun  videos  already. I'll  just  skip  this  last  time, but  that's  just  a  mouse  using another,  another  behavioral  strategy. So  a  couple  of  quantification. So  the  mice  learn  this  task pretty  quickly,  which  is  a  good  thing. So  this  is  the  accession  number  on the  x-axis  sessions  of  training, each  line  as  a  mouse. And  then  the  number  of  trials  increases  with  training  and their  performance  according  to two  different  metrics  increases  over  time  too. So  the  chance  rate  of  performance, one  out  of  eight  is  pretty  low  on  this  task. So  the  nicer  correctly  going  to that  gold  word  before  any  other  one, even  half  that  amount  of  time, it's  significantly  above  chance. Mindful  learned  the  task  which is  great,  unexpected  observation. The  females  tend  to  be  slightly but  consistently  better  than  the  males, which  is  not  something  we  expected  at  all. And  if  anybody's  interested  in  that, I  can  talk  offline  about  why  we  think  that  might  be. Then  I  want  to  mention  also my  lab  manager,  Roman  guard  Guo, who  got  a  lot  of  these  experiments set  up  and  has  had  a  hand in  almost  everything  else  that I'm  Michelle  you  for  the  rest  of  this  talk. So  one  aspect  of  behavior  that  we  can study  with  this  task  is  recovery  from  sensory  loss. Hearing  loss,  probably  the  most common  neurological  disorder  in  the  world. Common  in  the  elderly, also  common  in  kids.  You  have  an  ear  infection. You  can't  hear  as  well  on  one  side versus  the  other  for  a  short  period  of  time. We're  just  gonna  understanding  how  the  brain can  overcome  that. What  motor  strategies  mice  might use  to  compensate  for  sensory  loss. So  for  instance,  if  the  mouse, normally  with  your  son  on  the  right, maybe  it  turns  to  the  right. But  if  we  knock  out  hearing  on  one  side, how  does  that  affect  its  behavior? Is  it  going  to  overshoot  to the  left  or  something  like  that? And  so  what  we'd  like  to  model  here. So  we  developed  a  surgical  approach to  induce  hearing  loss  and  the  mice. We  did  this  in  collaboration  with the  residents  surgeon  and then  your  surgeon  at  Emory  and  Caitlin  Brooks. And  she  taught  me  and  my  undergrad  and Meghan  Jin  to  do  these  really  difficult  surgeries. So  you're  looking  into  the  ear  canal  here, or  way  down  at  the  bottom  there  is  the  malleus, one  of  the  middle  ear  bones. So  one  of  the  smallest  bones  in  the  mouse  body. And  then  Megan  Guy,  quite  good  at  doing the  surgery  and  extracting  the  malleus. Here's  one  after  it's  been  removed. That's  a  really  controlled  way  for  us  to  induce  hearing loss  on  one  or  both  sides  of  the  head. It's  not  a  complete  hearing  loss, it's  about  30  dB  reduction  and  sensory  in. Here's  what  happens  when  we  definitely  mice, the  mice  on  both  sides. That's  the  red,  the  red  lines, I  shouldn't  say  deafness,  so  it's  a  partial  hearing  loss. But  when  we  do  it  on  both  sides, the  mice  do  much  worse  at  the  task. And  they  tend  to  not  really  recover. And  maybe  they  recover  slightly  as  you  can  see  over  time. But  when  we  definitely  my  son  just  one  side. First  of  all,  it's  not  a  partial  reduction. It  is  a  complete  reduction  almost  a chance  with  unilateral  hearing  loss. But  they  retain  this  ability  to  actually recover  almost  to  the  ability  that  they  were  before. So  we're  really  invested understanding  how  that's  working. I  mean,  the  malleus  is  not growing  back  great,  but  there's  no, there's  no  recovery  on  the  sensory  side that's  possible  in  this  condition. And  so  our  hypothesis  is  that  it's  a  motor  recovery. So  maybe  they  learn  to  hold their  head  in  a  different  way. Pointer  healthy  here  at  the  sound,  something  like  that. Some  way  that  they're  able  to  actually  make  up  for that  lack  of  that  loss  of sensory  information  that's  not  coming  back. I  think  that's  maybe  a  cool  way  to  look at  the  plasticity  and  capability  as motor  circuits  and  also  might  have some  implications  for  rehabilitation  paradigm. So  what  parts  of  the  brain  are  involved  in  this  task and  learning  it  and  performing it  and  recovering  from  hearing  loss. This  is  a  view  of  the  entire  mouse  brain.  View. You  probably  never  see  where  anterior  is  on  the  right. Dorsal  is  on  the  top  and  posterior  is  in  the  back. And  I've  chosen  this  view  because  it  puts auditory  cortex  front  and  center. And  that's  really  where  we've done  most  of  our  experiments. So  we're  starting  at  auditory  cortex  or  context. You  can  see  some  other  cortical  regions including  visual  somatosensory  and  motor, and  then  cerebellum  in  the  back. So  a  lot  of  people  were  skeptical, but  auditory  cortex  would  be  involved  in  this  at  all. There's  a  widespread  belief  that  it's  completely dispensable  for  auditory  localization. And  indeed,  there  are  really  beautiful  circuits. As  I  mentioned,  I'm  in  the  colliculus  for  this. But  this  is  a  freely  moving  tasks. There's  a  motor  component  and  we  wanted  to  check for  ourselves  whether  it  was  actually  necessary. So  the  way  we  did  that  first  course  but  effective. So  this  is  a  surgical  lesion  approach. We  didn't  aspiration,  so  we  suck  out  brain  tissue. Your  auditory  cortex  on  both  sides  has been  removed  by  aspiration. This  was  a  surgery  that  was  done by  the  sum  of  the  scene  and Eliana  Palliative  other  undergrads in  the  last  group  of  mastery, this  is  pretty  difficult  surgery. We're  going  to  ask,  how  does  that  affect  their perform  their  ability  to  do  this  task? If  they  are  unable  to  do  this  house, I  would  suggest  that  auditory  cortex  is  really essential  for  the  computations  that  are  taking  place. And  in  order  to  know  where  the  lesions  are, we  aligned  it  with  the  Allen  Brain  Atlas  or  the  blue here  it  shows  you  the  various  auditory  cortical  regions. We  find  the  auditory  cortex. There's  some  variability  across  animals, first  of  all,  which  we'll  get  into. But  we  believe  that  I  was  very  cortex  is generally  necessary  for  this  sound  seeking  task. So  if  you  look  first  at  the  dark  lines, That's  the  animals  that  had a  severe  impairment  after  auditory  cortex  lesion. They've  built  from  doing  pretty  well  with the  task  before  being basically  a  chance  and  they  don't recover  with  additional  training. But  we  also  see  animals,  that's the  gray  solid  lines  here. And  where  we  do,  what  we  think  is  the  same  surgery and  we  get  very  little  effects  in  those, in  those  cases  and  mild  effect. Then  as  a  control,  we  also  have visual  cortex  lesions  of  approximately  the  same  size. We  don  t  think  vision  is  involved  in  this  task. We  don't  think  visual  cortex  is  involved  in  this  task. And  those  mice  don't  show  any  effect  at  all. So  that's  a  control  for  if  there's some  side-effects  from  removing this  brain  tissue,  which  we  don't  see. But  we  do  see  this  variability between  an  auditory  cortex  lesions. And  we  think  we're  still  processing  the  histology  here, but  we  think  that  it  has  to  do  with  the  size and  location  and  the  lesions. So  this  was  work  that  Osama did  with  the  summer  student  careers  and  Oregon, this  is  a  lot  of  work,  so  we've  got hundreds  of  slices  here, is  align  them  all  with  the  Allen  Brain  Atlas  and then  traced  out  the  location  of  the  lesion  and  all of  those  different  slices  and  related  it to  the  cortical  regions  that  we're  looking  at. And  so  these  are  just  all  the,  all  the  brain  slices. Again,  blue  is  auditory  cortex, pink  is  visual  cortex. Wherever  you  see  this  stippled  pattern, that's  lesions,  brain  tissue and  whether  it's  open  and  that's  healthy  brain  tissue. So  what  we  saw  when  we  did  all  this  work  was  that the  lesions  where  he  will  only partially  taken  out  auditory  cortex, That's  the  ones  on  the  left. Those  are  the  mice  that  had  the  smallest  effect. So  maybe  it's  not  sufficient. We  think  they're  just  remove  primary  auditory  cortex. Those  other  auditory  subfields might  be  capable  of  doing  this  task. So  there's  some  redundancy there  where  they  can  cover  for  each  other. But  in  Osama's  case, when  he  was  able  to  lesion  the  entire auditory  cortex  on  both  sides, primary  subfields  and  secondary  subfields. Then  the  mice  show  that  really  large  effect. Then  again  as  a  control, we  don't  see  any  effect  in visual  cortex  lesions  even  of  the  same  size. There  seems  to  be  some  redundancy  redundant  circuits within  auditory  cortex, but  it  does  seem  to  be  involved, if  not  essential,  for  the  behavior  that  we're  looking. And  then  of  course,  future  directions  here, abigail  macro  re-visit  grad rotations  student  in  the  lab  who  developed a  human  genetic  approach  where  we're  able to  have  injected  virus. They  allow  us  to  more reversibly  inactivate  activity  in  this  brain  region. We'd  also  like  to  do  optogenetics  approach so  we  can  get  more  temporal  specificity. It  started  to  look  at  the  role  of different  inputs  and  outputs  from this  brain  region  and understand  the  broader  circuit  is  working  together. That's  it,  That's  a  future  direction. Alright,  so  in  the  last  few  minutes  I  want  to focus  on  what  I'm  actually  the  most  excited  about. This  is  really  kind  of  fresh, fresh  data  that  we've  just  started  collecting. And  it's  what  we're  building  towards here  is  large-scale  recordings  and  freely  moving  mice. We  want  to  understand  how  computations  in auditory  cortex  and  elsewhere  enabled  this. Now,  I  know  you're thinking  I'm  going  to  St.  neuro  pixels. But  the  problem  here  is  that, and  this  is  where  we  pay  the  price  for  moving to  the  freely  moving  animals. There's  only  so  much  weight  that  these, these  little  guys  can  carry  around  on  their  head. They're  like  20  or  30  g.  And neural  pixels  was  designed  essentially  for  rat  brains. And  so  it's  just  too  large  for my  security  or  at  least  to  comfortably  carry  around. Another  moving  into  primates  and  humans  and  stuff. So  I  don't  know  if  they're  ever  going  to  really optimize  for  the  freely  living  mouse  case, but  that  we  really  care  about. Now  some  people  have  been  able  to  do  it  often  by only  using  males  which  are  bigger  and  stronger, male  nice,  and  females. But  of  course  that  introduces  sex  bias. In  any  case,  even  if  you  can  get  them  to  do  it through  weight  training  regimens  and  everything  else. You've  got  all  these  wires  coming  out  of  the  head. Those  wires  actually  put a  small  but  kind  of  annoying  force  on  the  head  of the  animal  as  it's  trying  to  do these  really  sophisticated  head  motions that  we're  trying  to  study. And  so  we  wanted  to  move  away  from  wires and  everything  else  that's  big  and  bulky. And  so  what  we've  done  is  we've  got this  commercially  available. It's  from  a  company  called  white  matter  systems. It's  a  wireless  system  that's  capable of  256  channel  recording. We're  using  tetrads  that  are chronically  implanted  in  auditory  cortex. The  thing  is  about  the  size  of  a  dime, so  it  fits  on  the  top  of  the  mouse's  head. We're  getting  good  data  out  of  it. We  can  get  single  units  and  auditory  cortex  works  well. And  I'll  show  you  a  video  in  the  next  slide. And  it  doesn't  seem  to  affect  their ability  to  move  at  all. What  we  have  in  the  works  is  this  device  is  also  capable of  genetic  manipulation  with  it. So  the  idea  would  be  to  move  to  a  freely  moving  animals. We're  doing  wireless  neural  recording, wireless  optogenetics  and  3D  tracking. I  think  that's  how  we're  really  going  to address  these,  these  questions. This  works  being  led  by  Cedric  bulb  was  an  MD, PhD  student  in  the  lab. Help  from  Jessica  MI  and  other  undergrad. And  Lucas  Williams  Center at  another  rotation  students  who did  the  post  tracking that  I'll  show  you  in  the  next  video. So  here's  the  video  of  the  mouse  performing  that  task. So  you're  hearing  a  sound. I'm  not  doing  that. I  can  watch  that  all  day,  but we  only  have  so  much  time  here. We're  just  starting  to  dig  into  this  data. I  don't  have  a  lot  of  sophisticated  analysis  to  show  you, but  I'm  trying  to  start  also  with  an  open  mind and  just  see  where  the  data  take  me. One  of  the  things  that  I  see  in  this  video  maybe  is that  there's  these  kind  of  pause  and  neural  activity that  seems  to  happen  just  at the  times  when  the  mouse  is  entering  one  of the  chambers  or  turning around  to  exit  the  chamber  or  maybe  moving  its  head. And  I  think  that  pause  might  represent a  time  when  activity  is  briefly synchronized  because  auditory  perception  is  particularly important  at  that  time,  or  something  like  that. And  I've  got  one  quick  analysis  to  show  you, which  is  all  we've  done  so  far, auditory  cortex,  you  want  to  know that  these  neurons  respond  to  sounds. Of  course. This  is  the  PSD  shifts 20  simultaneously  recorded  neurons  in  auditory  cortex. With  respect  to  the  sound  onset. Remember  the  sounds  are  like  seasonally  is  bursts. And  we  see  this  classic  auditory  cortex  response, which  is,  on  the  one  hand  really  big, but  on  the  other  hand,  really  small. Meaning  it's  like  ten  standard  deviations above  the  baseline. But  it's  in  such  a  brief  window  of  time, it's  about  five  to  15  milliseconds after  the  sound  comes  on  that  if  you  do  the  math, it  works  out  to  be  zero  or  one  additional  spikes per,  per  neuron  per  sound. So  thinking  back  to the  data  from  the  first  part  of  the  talk,  I, I  kinda  think  that  the  sound  evoked responses  aren't  going  to  be  the  major  player  in  this, in  this  brain  area,  auditory  cortex. I  think  what  might  be  much  more  important is  the  motion  of  the  animal. The  animal  is  expectations about  what  is  going  to  hear  next  in its  predictions  about  how its  motion  is  going  to  affect  those  sounds. So  what  we'd  like  to  do  to  test  that  as  to record  and  motor  cortex  on  both  sides  of the  brain  and  auditory  cortex on  both  sides  of  the  brain  using this  wireless  system  and  start  to  move towards  understanding  how these  brain  areas  are  working  together. So  we  can  use  optogenetics to  monitor  or  manipulate  these, these  connections  between  these  green  areas. Try  to  understand  how  is  the  action aligned  with  the  sensory  and  how our  interactions  between  these  brain  areas  enabling. This  is  my  last  slide  is  my  overall  model of  how  the  brain engages  with  the  world  through  active  sensing. So  standard  model  is  that sensory  areas  of  the  brain  are  for  encoding sensory  input  and  motor  areas are  for  controlling  motor  output. What  might  be  happening  here  is  slightly  different. I  think  it's  sensory  areas  make the  building  and  internal  model  of  the  world. That  the  purpose  of  this  model  would  be  to  be able  to  predict  the  causes  of  our  sensations, to  predict  what  is  causing  out  there  in  the  world. Because  sensory  input  that  we're  receiving. Then  at  the  same  time, motor  is  my  building  an  internal  model  of  the  body. So  something  that's  capable  of predicting  the  effects  of  our  actions. If  I  move  my  head  in  this  way, how  is  that  going  to  affect  the  sound  that  I  hear? And  of  course  they  don't  do  these  things  in  isolation. The  interactions  between  these  areas  are  really critical  for  updating  these  internal  models  and allowing  them  to  exchange  information about  the  kinds  of  adaptive  actions  that  would be  informative  or  important  to take  and  what  the  effect  of  those  actions  might  be. I  think  that  the  framework  of  predictive  processing is  could  be  a  really  good  way  to think  about  these  interactions. Making  predictions  about  the  future  is an  essential  way  to  just  survive  and  thrive. They  have  to  be  able  to  predict  what's  going  to  happen when  I  reach  out  and  touch  this  object. What  is  this  person  going  to  say  when  I, when  I  say  whatever, I  think  that's  an  essential  computation than  any  intelligence  system  must, must  be  capable  of. This  framework  of  thinking  about predictions  and  exchange  with  predictions could  be  a  good  way  to  think about  cortical  processing  in  general, and  possibly  neural  processing  and  other  systems  as  well. The  last  thing  I  want  to  say  is  that this  paradigm  could  be  a  good  way  to  think about  how  we  could  use  body  motion  to compensate  for  sensory  and  cognitive  disorders. So  we  talked  already  about  hearing  loss. We're  also  interested  in  looking  at  Alzheimer's. We  have  some  Alzheimer's  mice  in  the  lab. We  think  about  Alzheimer's,  you  think  about the  memory  effects  that  you  see  very, very  late  on  the  progression  of  this  disease. But  they're  actually  preceded  by several  decades  by  sensory  and  motor  changes  as  well. We'd  like  to  think  more  about  what those  sensory  motor  changes  might  be. Maybe  an  opportunity  to  think  about  early  diagnosis  and also  a  way  to  think  about  how  these  early, potentially  subtle  changes  in the  way  we  move our  bodies  during  diseases  like  Alzheimer's might  be  contributing  to  or  relating to  some  of  the  cognitive  deficits  that those  patients  experience  as  well. With  that,  I'll  just  do  my  acknowledgements. I  want  to  mention  I  think  I  mentioned most  of  the  people  in the  lab  as  their  contributions  came  up. The  authors  ever  hear  her  coauthors, colleagues  of  mine  at  Columbia  and  rain  Bruno  is  lab  for the  first  part  of  the  top  whisker  is  part  of the  talk  on  then  of  course, I  want  to  thank  my  funding  sources  and  mentioned  that the  lab  is  actively  recruiting graduate  students  and  postdocs,  so  on. Please  spread  the  word  and  thank  you  very  much  for your  attention. Great  Chris. So  do  you  think  that the  predictions  that  are  the predictive,  Predictive  Coding  idea, do  you  think  that  the  predictions are  precise  predictions  of specific  states  like  imputations  of what's  going  to  happen  or  if  they're  probabilistic. And  what  do  you  think  that? Do  we  have  some  sense  of  the  variability and  we're  comparing  the  variability  of  what's  happening. Or  do  you  think  we're  comparing  in  this, particularly  in  this  auditory  case, maybe  the  precise  expectation to  precisely  what  happens  to  have  a  precise  error. Yeah.  Probably  a  little  bolus  guess. I  mean,  in  general,  I  would like  to  take  a  Bayesian  approach  to  this. So  to  think  about, it  will  be  probable  as  thick, I  think  computations  in  the  brain, it  has  some  other  half  to  be a  probabilistic  because  the world  is  probabilistic  or  at  least  the  way that  we  experience  it  as  her  sensory  apparatus. Including  the  photo  like a  delay  life  processing  isn't  isn't  really  probabilistic? Based  on  the  variability  of  me. Yes,  thank  you.  Good  point. I  do  expect. So. I  do  expect  to  see  explicit  coding  of the  uncertainty  or  probabilistic  nature  of  it. And  apparently  that's  by  design. So  that's  why  I  have the  stimuli  in  this  task  that  they  are  probably, it  is  not  just  if  it  was just  one  sound  and  you  orient  towards  it, then  I  think  that  that  can  be  solved  in, let's  say  like  delay  line  type  of  way, kind  of  like  an  instance  owls,  right? But  that's  because  there's very  little  uncertainty  in  the  stimulus  of  self. And  because  I'm  interested  in  thinking  about more  probablistic  computations  that  I  think  the  cortex might  be  especially  good  for  design  for. Therefore,  I  think  that  the  task  needs  to be  an  inherently  probabilistic  two. So  if  we  have  some  other  versions  of  the  tasks  that really  kinda  pushing  that  direction where  there's  some  spatial  uncertainty. Maybe  multiple  speakers  are  playing  sounds. Sounds  that  are  more  or  less  informative. So  that  the  mice  really  have  to integrate  over  space  and  time in  order  to  make  their  decision  about  where  to  go  next. I  was  really  motivated  by  somewhere  from  being Breton  and  Carlos  Brody  from  awhile  ago  where  they, they  use  these  Poisson  stimuli. And  the  reason  they  do  that  is  because  now  you  have explicit  control  over  how  uncertain  the  stimulus  is. And  you  can  write  down  what  the  best even  an  ideal  observer  might  be  able  to  do. And  I  think  that  I  didn't  have time  to  talk  about  one  of  Lucas's  project  in  the  lab, but  it  was  to  kind  of  build  a  model  that  can  find the  Bayesian  optimal  solution  for a  noisy  and  probabilistic auditory  localization  tasks  like  this. So  I  guess  my  answer  is, I  think  it  will  be  probabilistic, but  also  that's  because  we're  designing the  experiment  to  put  in that  direction  in  the  first  place. I  think  that  the  colliculus  would  be  the  first  place. I  would  look  for  a  more  kind  of deterministic  if  you  want  to  call  it  that  are  more, like  I  said,  I  could  delay  line  system,  hardwired. Choose  the  speaker,  the  question. Yeah.  Yeah.  So  thanks. Great  talk.  I  enjoyed  it  very  much. I  really  like  the  fact  you're covering  different  sensory  modalities  and  kind of  making  us  sort  of  think  outside  of the  way  we  usually  do  things. I'm  curious  both  of  those  sensory  those  senses. They  are  both  sensory  and  motor. But  the  motor  doesn't  actually affect  media  when  you  hear  something  or  when  you, the  whiskers  are  touching. There's  no  real  interaction  with  the  object  right  there. Such  small  forces,  e.g. with  the  whiskers  and  with  auditory  year  just  receiving. And  I'm  wondering  if  you  have  any  thoughts  about. The  example  we  used  was  holding  your  coffee  mug, but  you're  also  picking  it  up. And  that's  in  some  ways  very  different,  right. And  so  I'm  wondering  if,  if, if  you  have  any  thoughts about  that  kind  of  modality  versus  one  where it's  you're  not  really  mechanically interacting  with  the  objects  we're  impacting. Yeah.  I  mean,  I  think  that's  why the  reach  and  grasp  system  is  a  really  interesting  one. Because  not  only  is  it  kind  of  sensory  adaption  action, but  you're  also  like  effecting  the  world  directly. If  you're  changing  the  world, you're  deforming  the  object  or  lifting  it. And  that's  an  opportunity  to  really  think about  control  of  your  own  body, but  also  the  external  world  at  the  same  time. I  think  that's  a  really  fascinating  system. It's  challenging  to  study. If  you  think  about  grasping  an  object, you  can't  really  visualize  what's  happening  between the  point  of  contact is  impossible  to  visualize  at  the  time  of  contact. I  think  it's  more  difficult to  kind  of  get  some  traction  there. So  I've  chosen  to  really focus  on  what  you  might  call  the  not  passive, but  where  all  the  control  problems  have  to  do with  controlling  the  body  rather than  controlling  the  external  world. But  I  think  it  would  be  really  cool  to  think  about. Ultimately,  of  course,  what  we  want  to  do is  engage  the  world  in  some  way. But  the  way  I've  set  this up  is  still  a  little  bit  more  of a  sensory  perspective  where  all  the  control  is  just being  used  to  optimize  signal  collection. And  not  necessarily  to change  the  external  world,  but  that  could  be. Yeah,  oh,  think  more  about  ways to  engage  with  that  because  I  think  that's, I  think  that's  a  really  interesting  future  directions. That  was  awesome,  Chris. Thinking  about  sensory  motor  circuits  in  general. So  the  first  part  of  your  talk, you  mentioned  that  in  barrel  cortex  there  is a  huge  weight  in  the  GLM  for  movement. And  how  do  you  untangle  those? Like  in  a  linear  model,  it  seems  like  with  the  whiskers, every  movement  also  changes the  sensory  stimulus  in  this  context  as  well, wherever  the  mouse  is  orienting  is  going  to  change the  impulse  response  of  its  head  relative  to  the  speaker. So,  are  there,  are there  strategies  you're  thinking  of  for decoupling  the  sensory  and  motor components  in  this  context? Or  do  you  think  your  task  has  some  way  of giving  you  a  really  nice  traction  on  this. Certainly  very  challenging. I  think  what  we  have  to  help us  is  that  motion  is  generally  continuous, but  the  sensory  input  is  temporarily  discrete  units, so  the  contact  is  made  at a  certain  time  or  a  sound  is  played  at  a  certain  time. And  I  showed  that  PSD  H, and  it's  similar  in  URL  cortex, you  get  these  very  transient  kind of  quick  pulsatile  responses. So  that  might  be  a  way we  can  disentangle  it  on  the  neural  level. Whereas  I  would  expect  signals  that  relate  to  motion to  be  smoother  because  the  motion  itself has  smooth,  it's  ongoing. And  so  that  is  essentially the  way  we  were  able  to  disentangle  it. And  then  also  the  way  the  GI, my  first  approach  to  those  kinds  of  questions  is  always regression  just  because  it's simple  and  it's  interpretable. And  so  you  can  just, you  don't  even  have  to  explicitly  asked, but  the  model  essentially  it  tells you  when  you  make  exactly  the  same  motion and  there  isn't  a  sensory  input  on that  whisk  cycle  or  on  that  head  turn. What  is  the  difference  in  the neural  response  that  you  get? That  would  be  one  answer  to  your  question  and  you  can. Disentangling  is  just  kinda  the least  squares  solution  on  it  as  long as  you  have  enough  variability  in  your  feature  set, which  you  always  done  from a  variability  in  this  kind  of  behavior, then  you  can  do  a  pretty  good  job  of  disentangling  it. But  I  think  conceptually  it's  more difficult  than  that  because  they  just  can't  do  it. Emotion  is  taken  for  the  purpose of  filtering  the  sensory  input. So  I  think  there's  gonna  be  conceptual  challenges  ahead. I  probably  can  directly  answer that  until  we  have  the  data  in  hand, until  we  really  start  to  dig  into  it. Hello. So  I  guess  my  question  is really  in  the  same  vein  as  Jeff's, but  I  was  just  wondering. So  the  cortex, the  somatosensory  cortex  seems  to  have  this  signal  of like  a  copy  of  the  motion  in  that  it's  representing. And  I  was  just  wondering,  I  guess, what  the  proprioceptive  component  of  that  was. I  think  it's  really  the  same  question  as  Jeff. Yeah. I  haven't  better  answer  to  that.  10  there  are no  plastic  proprioceptors  in  the  whisker  system. It's  just  too  fast. We  thank  for  proprioceptors  to  make  sense. It  seems  to  be  entirely  efference  copy. That  tells  the  brain  where  the  whiskers are  and  maybe  to  some  extent  skin  stretch, like  the  actual  skin  of  the  face. But  that  would  give  you  a  relatively  coarse  signal about  my  kind  of  like  whisking  or  not. But  it's  not  fast  enough  to  tell  you exactly  where  the  whisker  is  at  any  given  time. Um,  but  I  really just  didn't  think  about  that  in  the  head, in  the  head  direction  signal, which  is  what  we're  actually  working  on  now. So  head  direction  input  could  come  from  efference  copy, it  could  come  from  proprioception. And  absolutely,  it  can  even come  from  reference  like,  you  know, it's  where  your  head  is  pointed  because of  the  visual  input  that  you're getting  or  vestibular  input,  for  instance. And  I,  so  I'm  going  to  try  to  take an  open  mind  to  that  and  think  that  it might  be  any  of  those  things and  probably  is  all  of  those  things  to  some  extent. And  then  yeah,  I guess  we  might  try  it  at  some  point  to  inactivate different  brain  areas  like  the  stimulator  areas versus  visual  areas  are some  things  to  try  to  break  that  apart. There's  a  really  good  paper, awesome  visas  that  first  down  here, I  think  it's  Adrian  pay  rushes  lab  or something  where  they  might  get  navigation and  the  role  of  multi-modal  inputs that  are  important  for  grounding  this  head  direction. Because  that  answer  your  question.  Yeah. Thank  you. Because  that  was  an  amazing  talk. I  am  curious  about  you  talked  a  little  bit  about the  role  of  compensations  in  the  context. The  second  part  of  the  talk  with the  malleus,  remove  fall. Like  I  actually  think  in  the  first  part  of the  talk  with  the  whisker  trimming  there  there  is some  argument  for  there  to  be compensations  and  I'm  curious  whether  or  not you  feel  as  though  There's  kind  of  a  uniform  set  of compensations  that  are  occurring either  on  the  sensory  side  or  the  motor  side, or  whether  this  is  something  that  differs  by  animal. Whether  or  not  you  can  talk  about  the  types  of compensation  swimming  have  observed  on  the  motor  side, or  maybe  things  that  you  might expect  on  the  sensory  side. For  the  hearing  loss  case. I  don't  think  there's  really any  possibility  for  sensory  compensation. Like  once  you've  removed  that, it's  kinda  like  blindingly. I  mean,  it's  not  a  complete  perfect, but  I  just  don't  think  there's  anything  they  can  do  two, on  the  sensory  side. So  I  think  it  has  to  be  entirely  motor. And  our  current  guess  is  that  it  has  either holding  a  healthy  ear  towards  the  sound. But  it  could  be  other  things  too  simple, like  taking  longer  to  make  your  decision, integrating  more  evidence  before  you  decide  what  to  do. Maybe  the  re-learning  is  just  the  case. It  could  be  a  cognitive  compensation,  I  guess, sort  of  more  aware  that  you  have this  limitation  and  then  you  compensate  in  other  ways, which  I  think  maybe  it's interesting  for  the  Alzheimer's  cases  to  actually. So  I  would  guess  that  in  our  case it's  a  it's  never  a  sensory  compensation. Now,  the  whisker  case  though, because  that  is  a  reversible, the  whiskers  fall  and  they  grow  back  in  all  the  time. That's  definitely  a  case  where  I  would  imagine more  of  a  role  for  sensory  compensation. Like  you  said,  it's  only  a  70%  hearing  loss,  right? There  potentially  could  be some  sense  this  true  Actually,  yeah. Okay,  That's  true.  That's  a  great  point. So  we  want  to  first  measure  that  more  explicitly. They  have  to  develop  and  it's  kinda  like  the  best. So  you  have  to  develop  his  way  of  measuring the  brainstem  response  to  the  auditory  stimulus. So  we  can  see  to  what  extent that  the  changes  have  been  very  first  couple  of synapses  are  are  either  constant  over  time, which  I  guess  is  what  I'm  assuming, or  whether  there  is  some  some  sensory  recovery. But  I  think  that  even  if  it  is, there  is  some  recovery  like  it's  gonna  be  hard, like  you've  removed  so  much  that  the  signal that  there's  only  so  much  you  can  really  do  on  that  side. There's  only  so  much  they  filtering  can  really  do to  help  you  because  it's such  a  disordered  signal  that's  coming in  and  it  gives  you  the  hearing  loss. And  so  I  think  that I  tend  to  think  of  it  more  as  a  motor  problem, but  but  that's  a  good  point. I  guess  we  won't  really  know  until  we do  those  ABR  screener. I'm  curious  if  you  look  at  this  task  from the  perspective  of  an  accumulation, that  basically  the  noises  not  from  this  sensory  input, but  it's  probably  wear  the  mask  looks  right. If  they  go  to  the  wrong  direction, then  it  goes  down. The  evidence  is  going  down. And  as  they  move  closer, it's  like  it's  building  up. So  I'm  curious  if  that's  how  your  new  Canon, how  they  solve  this  task  and that's  what  you  choose  is  1  s. Yeah,  yeah. So  I  am  definitely  want  to  look  at  that  more. I  do  think  something  like  that  is  happening. Lucas  and  rotation  projects, purely  theoretical  simulation-based, built  the  little  agent  to  solve  this  task. And  I  think  that  that  could  be  a  good  way to  start  to  think  about  that  problem. We  haven't  really  gone  any further  with  that  and  they're  real  data, but  I  want  to  weaken  the  weekend, design  it  such  that  we  have  total  control over  the  amount  of  noise  in  the  sensory. So  presumably  remaining  sources of  noise  or  motor  information  going  the  wrong  direction, he's  not  really  collecting  the  right  and  information. There's  also  a  lot  of  like, I  guess  you'd  call  it  cognitive  noise. So  when  the  mice  get  confused, they  don't  just  kinda  like  sit  there  in the  middle  and  stop  working. They  resort  to  this  really  simple  strategy which  is  not  effective, which  is  just  going  around  and checking  all  of  the  parks  and  work. And  so  a  lot  of that  when  they're  doing  badly  at  the  task, I  don't  think  it's  necessarily  just sensory  or  motor,  but  it's  actually, they've  chosen  to  adopt this  very  simple  strategy  that  is  chance  level. So  it's  really  more  of a  behavioral  stages  that  they  go  through. And  I  think  that  ends  up  being the  major  limitation  on  how  well  they  can  do  it. But  I  guess  what  I  was  trying  to like  one  is  that  there's  this  picture  that  you  have  here, that  is  it  that  we're  trying  to  see  if auditory  cortex  is  performing  some  decision-making  task. Here  is  what  you're showing  here  that  they  were  trying  to  do some  printing  different  processing  computation. And  when  you  say  cause, I  guess  by  costuming  then  location  here,  right? So  I  have  a  little  bit  of  difficulty also  complete  the  mapping. This  to  your  question. Yeah.  Yeah. I  mean,  certainly  this  is  not  very  fleshed  out, so  this  is  kinda  like  a  roadmap for  where  I'd  like  to  go  to  think  about  things. I  mean,  I  definitely  think  that  it's both  parties  involved  both in  decision-making  and  in making  predictions  about  the  world. Like  making  predictions  about  the  world  for the  purpose  of  coming  up  with  the  decision. But  I  think  that  too,  yeah, we're  going  to  need  the  data  before  I can  say  anything  more  rigorous  about  that. Alright,  if  there's  any  more  questions, feel  free  to  approach  the  podium, but  for  now,  we'll  thank  the  speaker. Thanks  so  much,  Chris,  for  a  grade.