Yeah,  it's  really  nice  to  be  here  with  you  all  today. This  is  the  first  time  I've  given  one  of these  departmental  or  university  level  seminars since  the  pandemic  hit, actually  with  a  full  day  of  meetings  and  stuff  like  that. I  used  to  do  that  a  lot  back  in  the  old  days, but  first  time  since  today, I'm  going  to  walk  you  through a  bit  of  my  own  personal  journey  over the  last  few  years  on  questions  of  how best  to  analyze  neural  data. Over  the  last  six  years, my  group  has  been  involved  in  a  lot  of  different  things. I  describe  my  group  broadly  as  being  interested  in general  principles  of  intelligence  that  apply equally  to  artificial  and  natural  systems. But  one  of  the  things  we  have  to  do  if  we  want to  understand  general  principles  of  intelligence, be  they  in  a  natural  or  artificial  system, is  to  actually  be  able  to  analyze those  systems  and  make  statements about  them  and  test  hypotheses  with  them. And  also,  of  course,  do  downstream  stuff  with  them. And  for  that,  you  really need  good  machine  learning  tools. As  I'm  going  to  argue  to  today, I'm  going  to  start  out  with  what  I view  as  the  ultimate  goal  with  ML  models  for  neural  data. I'll  try  to  leave  some  time at  the  end  of  the  talk  because I  really  do  want to  open  up  a  can  of  worms philosophically  with  some  questions  here  because  I  think there's  some  critical  questions  for  us  as a  community  to  attend  to. Then  I'm  going  to  talk  about  some  of the  first  steps  that  we  engaged  in  in our  lab  to  work  specifically  with  calcium  imaging  data, which  was  a  personal  pet  project of  mine  because  I  had  done some  calcium  imaging  work  in  my  Phd  and found  the  way  that  we  analyzed it  to  be  immensely  frustrating. I'll  then  tell  you  about  the  change  of  heart. I  had  ultimately,  our  work  on  analyzing calcium  data  hit  a  roadblock at  some  point  about  two  years  ago. And  I  was  feeling  really dissatisfied  with  the  whole  endeavor. And  I'm  going  to  be  very  open  with you  about  that  dissatisfaction, but  something  changed  for  me  in  recent  years. Thanks.  In  part,  in  large  part  due to  meeting  Eva  Dyer  and  her  group  here  at  Georgia  Tech. And  also  some  broader  conversations I've  been  having  with  people  at  Mila, the  research  institute  where  I  work. That  led  me  to  a  different  way  of  thinking  about  it. And  thinking  that  in  fact,  the  key  thing  we  have  to think  about  for  analyzing neural  data  is  how  we  tokenize  it. And  I'll  explain  why.  Then  lastly, I'll  articulate  what  I  view  as  the  end  goal  for  us. What  I'd  like  to  see  this  type  of work  achieve  in  the  next  five  to  ten  years. Or  neuroscience  more  broadly. Let's  start  with  the  need for  machine  learning  methods  and  neuroscience. It's  worth  noting  that  we've received  a  lot  of  money  over  the  last  few  years. If  you  tally  up  the  total  amount  of  money spent  by  the  NIH,  various  European  agencies, Japanese  agencies,  Canadian  agencies, It  adds  up  to  over  $100  billion in  investments  in  neuroscience  over  the  last  ten  years. That's  a  lot  of  money  to  spend  on  a  problem. The  result  has  been  that  we've,  of  course, learned  a  lot  about  the  brain  and  we've  developed wonderful  technologies  for  recording  brain  activity. We're  at  the  point  of recording  brain  activity  that  I  only  dreamed of  as  a  graduate  student  not  more  than  18  years  ago. To  be  able  to  record  from  tens  of  thousands, or  even  1  million  neurons  at  once  in a  behaving  awake  animal,  wild  stuff. But  what  are  we  actually  going  to  use  all  this  data  for? I  think  there's  sometimes  a  tendency in  the  neuroscience  community to  engage  in  a  bit  of  what  I call  the  underpants  gnomes  thinking. Most  of  you  are  probably  too  young for  a  South  Park  reference. But  for  those  of  you  who  are  old  enough, I  really  do  think  sometimes  we, in  neuroscience,  think  like the  underpants  gnomes  from  South  Park, where  in  that  episode the  underpants  gnomes  are like  we  steal  all  the  underwear, then  something  happens,  then  profit. In  neuroscience,  there's  a  little  bit  of  this. We  collect  all  the  data, then  something  happens,  then  good  stuff,  trust  us. But  what  happens  is  in  between the  good  stuff  and  the  collecting  of all  the  data  is  rarely  articulated, I  think  actually,  poorly understood  in  the  neuroscience  community. Now  let's  maybe  at  least  try  to  articulate  the  engle. What  could  we  do  like  theoretically, if  we  had  access  to all  this  neural  data  and  ways  of analyzing  it  and  that  something  happened, worked  out,  what  could  we  do? Well,  there's  numerous  potential  applications of  having  good  neural activity  recording  technologies,  right? So  you  could  of  course,  have brain  computer  interfaces  for. Medical  purposes,  to  help  people  who  are  paralyzed, but  even  maybe  for  non  medical  purposes, for  just  helping  operators  of  large  machinery. Or  if  we  get  really  sci  fi, just  all  of  us  to  control  our  personal  electronics. Drug  efficacy  prediction  is  a  huge  thing. People  react  very  differently  to  psychiatric  drugs. We  have  no  way  of  detecting  right now  who's  going  to  respond  well  to  a  given  drug. Currently,  it's  just  trial  and  error, doctors  throw  drugs  at  people  and you  keep  doing  that  until  one  of  them  sticks. But  if  you  could  record  neural  activity  and  say  like,  oh, you've  got  this  neural  dynamics  type, you're  going  to  respond  well  to  this  drug  boom. Great.  Cut  out.  Three  to  four  years  of  misery in  patients  life,  Epilepsy  treatment. We  still  are  cutting  out pieces  of  the  brain  to  treat  epilepsy. It's  nuts,  but  theoretically, if  you  could  record  neural  activity, understand  where  exactly  the  dynamics  are  going  wrong, and  regulate  that  with  stimulation, it  would  be  immensely  helpful. I  could  go  on  and  on  and  on  throughout  all  of  these. I  think  the  core  point  is  that  we  see many  possible  end  games,  right? Like  it  would  be  really  nice  psychiatric  diagnosis. Like  if  someone's  depressed to  be  actually  able  to  say,  yeah, you've  got  a  real  issue  with your  underlying  neural  dynamics versus  you  just  need  to  change  things  in  your  life. These  are  two  very  different  treatment  paths forward  that  theoretically  we could  identify  with  neural  activity. So  these  are  all  the  great  end  games. This  is  where  we're  going  to  get  to  once that  some  stuff  happens  in  between. The  problem  is,  I  think  that  one of  the  reasons  we  have  this  some  stuff  happens in  between  issue  and this  is  honestly  something  you  that  keeps  me  up  at  night, is  there's  this  question. And  this  is  the  philosophical  can  of worms  I'm  here  to  open  up  today. Is  neural  data  even  interpretable? Many  of  us  would  like  to  be able  to  interpret  our  neural  data. This  is  a  word  you  hear  used  constantly,  the  last  cosine. There  was  a  really  fun  little  moment  of  tension  when Mackenzie  Mattis  gave  her  talk on  this  model  cebra  that  her  group  has  built. And  David  Silo  stood  up  and  said, my  question  for  you  is  you  keep  saying  this renders  the  neural  data interpretable.  What  does  that  mean? Mackenzie  was  like  it wasn't  clear  that  there  was  an  answer  to  it. I  think  the  reason  is  like,  okay, intuitively  what  we  mean  by interpretable  all  of  us  is  something  like  this. I  would  argue  that  I  could  get  to  the  point  where  using either  human  language  or  a  very  small  set  of equations  explain  the  algorithms that  are  in  operation  in  the  brain  to  you. That's  roughly  what  we  mean  by  interpretable. Okay,  It's  very  obvious  why  you'd  want  that, but  the  question  is,  is  that even  possible  with  neural  data? There's  an  assumption  that  we make  that  it's  possible  that  we could  get  to  the  point  where  I could  describe  to  you  with  human  language, how  this  brain  region  works. How  you  successfully  pick  up a  cup  or  decide  to  eat  this  food, or  decide  which  assignment to  do  for  your  class,  et  cetera. But  there's  also  the  distinct  possibility  that it's  just  not  interpretable  for  evolution. Had  no  reason  to  produce  neural  dynamics  that  were capable  of  being  interpreted  by the  human  brain  that  was not  something  that  was  selected  for. It  could  very  well  be  that the  algorithms  being  used  by  the  brain  are sufficiently  high  dimensional  that we  can  never  interpret  them  ourselves. And  we  could  never  explain  to  someone  else  how  they  work. That  might  just  be  the  reality  we  have  to live  with  if  that's  the  case. That  means  that  in between  part  where  we  go  from  collecting all  the  data  to  solving  all  the  problems can't  involve  necessarily  us  understanding  things. I  know  that  sounds  really  alarming  to  neuroscientists, but  I'm  putting  it  out  there  as  the  possibility  that maybe  we  have  to  accept  that the  way  we're  going  to  get  to  all  those  wonderful, final  endpoints  is  not necessarily  by  fully  understanding  how  the  brain  works, but  by  allowing  machine  learning tools  to  understand  how  the  brain  works. For  us  to  identify those  high  dimensional  algorithms  for  us. And  do  the  heavy  lifting  of  figuring out  what  will  be  the  response  of  a  person  to  a  drug. How  to  stimulate  their  brain  to  push it  away  from  an  epileptic  seizure. How  to  determine  what  they're  trying  to  think of  when  they're  trying  to  control  a  computer,  et  cetera. And  that  might  just  be  where  we're  going  to  get  to. I'll  be  honest.  I  think  that's  where we're  going  to  find  ourselves. I've  become  more  and  more  convinced  over  time that  that's  just  the  reality  that we  face  as  neuroscientists. It  might  not  be  satisfying in  the  same  way  that  it  would  be satisfying  to  have  a  nice verbal  description  of  how  the  brain  works, but  it  might  just  be  our  Lt. Actually,  I'm  going  to  skip  this  slide. I've  already  pointed  out  it. Machine  learning  can  help  us  to identify  these  relevant  patterns  in  neural  activity. It's  done  it  for  a  lot  of  other  data  to  be  clear. Like  that's  why  it's  the  promise  here,  right? Like  it's  very  hard  for  human  beings to  understand  things  like  protein  folding. But  it  turns  out  ML  systems  are perfectly  capable  of  understanding  protein  folding. Why  not  in  neural  activity? That  brings  me  to  the  point  now where  I've  set  up  what  the  goal  is. I  want  to  see  machine  learning  tools  that  help  bridge that  gap  of  understanding  that  we  might be  incapable  of  possessing  ourselves. When  I  first  entered  into  this  mindset, this  is  what  I'm  going  to  talk  about  now where  I  was  six  years  ago, which  was  using  deep, unsupervised  techniques  to  extract  latent  information. Well,  six  years  ago  I  started thinking  about  this  thanks  to  Chapman's  work. Then  about  four  years  ago, we  started  working  on  it  for  calcium  data. And  that's  where  I'm  going  to  bring  you  to  now. So.  Okay.  So  indeed  Heathan  and  David's  work  was  for  me, a  real  game  changer. Mentally,  I  remember  it  was  2018  cosine. I  saw  David  sitting  there  chatting  with  Tim  Lily  rap. It  was  the  opening  night and  I  came  up  to  them  and  I was  like,  oh  guys,  what's  going  on? Let's  go  party  And  like  David  was  like, I'm  having  a  very  serious  conversation  here. I'm  going  explaining  something  to  Tim  I'm  like,  okay. Okay.  So  I  sat  down  and  listened  to  what  he was  explaining  to  Tim  and  he  was explaining  alphads  to  Tim. Or  laden  factor  analysis  Dy  of  dynamical  systems. I'm  not  going  to  go  on  at  length about  this  because  probably  a  lot of  people  here  are  familiar  with  it. But  I  want  to  describe  the  core  idea  because  it  ended up  being  the  underpinning  for  our  work  for  a  long  time. Like  I  said,  for  me,  it  was  a  real  mental  shift  chief. And  if  at  any  point  I say  something  that  you  disagree  with, please  feel  free  to  interject  the  shift  mentally  for  me. As  I  was  sitting  there  listening  to David  explain  this  system  to  Tim, is  he  said,  look  like  you  get  your  spikes. They're  totally  just  the  expression  of  a  few  small  cells, like  a  snapshot,  a  slice  through this  computational  machinery  that  you're  doing. But  there  should  be  some underlying  neural  dynamics  that  are the  core  computation  that's actually  being  performed  by  the  circuit. What  we  would  want  to  do  from recording  this  neural  spiking  data  is somehow  extract  what  those  underlying  dynamics  are  like. What  is  the  actual  computation  being  performed  that these  spikes  are  just  giving us  a  tiny  little  window  onto. And  the  way  that  these  guys  decided  to tackle  this  problem  is  using a  deep  variational  auto  encoder. The  core  idea  is  you're  going  to  feed your  spikes  into  a  deep  neural  network. And  the  deep  neural  networks  job  is  going  to  be  to recreate  those  spikes  according  to  Bayes, optimal  principles  courtesy  of variational  auto  encoder  math. And  the  idea  was  you have  an  encoding  RNN  and  a  generator  RNN. Where  the  generator  RNN  is  then  going to  spit  out  for  you  in  a  series  of  factors, the  underlying  core  dynamics  that the  system  is  ultimately  engaged  in  computing. And  why  would  running  all of  your  data  through  a  big  variational  auto  encoder. So  like  it  runs  it  through  this  recurrent  neural  network, it  runs  it  through  this  recurrent  neural  network. This  spits  out  a  set  of  initial  states. This  then  spits  out  the  dynamics. You  get  your  latent  factors, those  get  turned  into  firing  rates  for  your  neurons, and  then  you  compare  to  the  actual  spikes  and  you  see, according  to  Polson  distribution, whether  you  are  predicting the  spikes  correctly  with  this  model. Now,  why  would  this  possibly  work? Well,  it's  because  variational  auto  encoders are  designed  to  be a  technique  for  inferring  latent  variables that  are  actually  responsible  for generating  the  data  that  you  receive. According  to  the  core  principles of  what  a  variational  auto  encoder  should do  if  you  run your  system  through  a  variational  auto  encoder  like  this. And  there's  some  tricks  here  that  I haven't  talked  about  in  variational  auto  encoders, but  it's  not  worth  going  into  in  this  talk. But  if  you  run  data  through  this  a  system, theoretically  what  you're  going  to get  out  of  this  system  is an  identification  of  the  underlying  latent  process that  generated  that  data, and  that  should  be  the  dynamics of  your  underlying  computation. And  indeed,  it  did  seem  to  work. Alphad  at  its  core  was able  to  pick  out  some  really  clear  dynamics. Possibly  the  easiest  way  to  see  that  was with  its  ability  to  decode  things. Now  this  is  going  to  be  a  running  theme  throughout my  talk  and  it  was  funnally  enough  in  the  end Mackenzie's  answer  to  David  at  Cosine where  she  was  like  by  interpretable,  I  mean  decodable. Now  of  course,  that's  not  what  most  people mean  by  interpretable,  but I  do  think  that  decodability  is at  least  a  good  metric  for  whether  you've got  a  good  hold  of  the  true  latents. Because  if  you  have  a  good  hold  of  the  true  latents, then  indeed  they  should  contain  all  of  the  information necessary  to  actually  generate the  data  that  those  latents  are  controlling, in  this  case  behavior. Here  they  were  testing  it  on  a  maze  reaching  task, where  an  animal  is  doing  a  maze  reaching  task. And  you  can  look  at  what  happens  if  you  train a  decoder  on  the  raw  data  that's  just  been  smoothed, or  maybe  run  through  a  Gaussian  process or  run  through  alphads. The  answer  is  you  can  do  a  much  better  job  of  decoding the  actual  animals  reaches  if  you  run  it  through  alphads, That  really  suggests  you're  picking up  on  the  true  latents  in  the  system. It's  a  very  clear  advance  over the  previous  systems  that  existed. And  yeah,  for  me  I  was  just like  super  excited  when  I  saw  this  stuff. I  was  like,  okay,  boom, So  we  can  start  to  identify the  true  underlying  latents  of  a  neural  circuit. This  might  actually  be  a  push towards  understanding  the  real  dynamics. Getting  some  kind  of  interpretability  if  we  wanted  that. Or  even  just  doing downstream  useful  things  like  decoding,  right? If  you  can  improve  your  decoding, you  can  do  a  better  brain  computer  interface. Boom,  done  great.  I'm  happy. What  about  other  modalities? For  me,  I  was  really into  calcium  imaging  because  I  had  done a  bunch  of  it  in  my  Phd  and  we also  had  a  collaboration  with  the  Allen  Institute. I'll  show  you,  I  think calcium  imaging  is  some  of the  most  beautiful  data  in  the  world  here. This  is  data  that  I'll  describe. Again,  recorded  from  primary  visual  cortex of  mice  watching  some  stimuli. We  did  recordings  in the  apical  dendrites  of both  subgranular  and  supergranular  cells. It's  this  beautiful,  beautiful,  rich  data. I  really  wanted  to  be  able  to take  advantage  of  this  data. Now,  Lp  ads  had  been  designed  for  spiking. We  thought,  well,  how  can  we  modify this  to  make  it  appropriate  for  calcium  imaging? The,  the  way  that  we  approached  this  is  we reasoned  that  this  was  really a  hierarchical  dynamics  situation. That's  what  we  were  dealing  with. The  core  idea  behind  L  Fads  is  that  you have  this  underlying  dynamical  system, that  is  your  computational  dynamics. This  is  what  you're  interested  in. You  have  some  raw  data  which  in the  initial  L  Fads  was  spikes. And  they  presumably  reflect a  fairly  simple  point  process  plus on  generation  based  on  these  dynamics. But  in  calcium  imaging  you  have this  intervening  set  of dynamics  which  are  the  calcium  dynamics. In  addition  to  the  underlying computational  dynamics  that  are  occurring, you're  observing  data  that's  being  mediated  through a  fairly  complicated  dynamical  system  that  is determining  calcium  influx  and  calcium  release and  all  these  other  things. We  reasoned  that  if  we  could incorporate  appropriate  inductive  biases into  a  deep  neural  network, a  modification  of  Al  Fads, then  we  could  get  it  to  learn both  the  latent  dynamics  and the  calcium  dynamics  that  we  see  in  this  data. The  approach  we  took  for  this  was  what's  called a  variational  ladder  auto  encoder, Hence  the  name  of  our  system,  the  cute  Valpaca. The  structure  of  this  neural  network  is  as  follows. You  feed  in  your  raw  calcium  fluorescence  traces. Those  get  first  fed  into  a  network whose  job  is  basically  to infer  the  calcium  dynamics  themselves. It's  like  a  little  mini  or the  calcium  dynamics  in  and  of  themselves. It's  going  to  do  that  variational  basian  inference  to understand  those  underlying  calcium  dynamics and  recreate  those  via  Gaussian  likelihood. Here  we  imagine  the  Gaussian  likelihood was  appropriate  because  we're  assuming  that the  experimental  noise  in fluorescence  imaging  is  Gaussian  distributed. But  then  this  is  why  it's  called  a  ladder  auto  encoder. You  also  feed  some  of your  inferred  activity  up to  another  variational  auto  encoder. So  this  is  a  hierarchy  of  variational  auto  encoders. And  this  variational  auto  encoders  job  is to  infer  those  computational  dynamics. What  we  do  then  is  we  tie  these together  by  doing  a  plusson  likelihood  comparison  between the  rates  this  model  infers  should  be  going  on  and the  underlying  rates  that  we  use to  generate  our  calcium  data. This  was  the  structure  of  Valpaca. We  thought  it  was  rather  clever. It  did  seem  to  work  here. We  did  some  analysis  on  synthetic  data. We  generated  data  using  a  Lorenz  chaotic  attractor. We  ran  that  through  a  model  of  calcium  dynamics. Then  we  added  some  noise  to  it  to  get out  a  fake  calcium  signal. What  we  found  was  that  if  we  look  at  our  ability  to recreate  the  Lorenz  State  or  to recreate  the  underlying  firing  rates that  created  that  data, we  were  doing  better  than  if  we did  a  calcium  deconvolution to  get  estimates  of  spikes out  and  apply  the  alphads  to  that. Or  simply  modify  the  cost  function  of  alphads  to  have a  distribution  at  the  output. We  were  ultimately  able  to capture  those  Lorenz  state  variables  and the  firing  rates  better  than these  other  more  straightforward approaches  that  you  might  take. That  was  encouraging.  Then  Mark  Churchland and  others  everyone  sheath. And  they  were  all  very  friendly  about this  and  it  was  great. They  also  shared  some  data from  monkeys  performing  a  reaching  task. So  we  got  spiking  data, we  created  semisynthetic  data  by  running those  spikes  through  a  calcium  dynamics  model to  generate  calcium  fluorescence  traces. And  then  we  examined  what happened  when  we  ran  our  system  on  those. Previously,  it  had  been  shown  via L  Fads  and  other  systems  that you  get  these  beautiful  rotational  dynamics when  animals  are  doing  these  reaching  tasks. We  showed  that  if  you  just  applied L  Fads  to  deconvolve  thlorescence  trace, it  didn't  seem  to  work  very  well. But  applying  val  Paca  to  this gave  you  these  really  nice  rotational  dynamics  again. Then  lastly  again,  decoding. We  had  some  data  from  the  Allen  Institute  where mice  were  sitting  on  a  wheel  watching  some  stimuli. They  were  getting  recorded  in  primary  visual  cortex. And  they  were  watching  these  frames  of  Gabors, where  the  animals  would  see  sequences  of Gabors  that  all  had  roughly  the  same  orientation. And  then  occasionally,  the  fourth  frame in  the  sequence  would  have  a  very  different  orientation, what  we  called  an  unexpected  frame. We  looked  at  our  ability  here  to  decode, whether  or  not  we  were  looking at  an  unexpected  frame  or  an  expected  frame. From  this  here  you're  looking  at  features  that  you get  out  of  either  running raw  PCA  on  your  data  or  Val  Paca. You  get  traces  that  are  better,  ultimately, at  identifying  the  changes  that occur  for  those  unexpected  frames. This  is  quantified  here  in terms  of  our  ability  to  decode  it. If  we  just  decode  off  of  the  features  that  we  get from  raw  PCA  we  don't  do as  well  as  if  we  decode  the  orientation  of that  frame  from  Alpaca. Okay,  this  all  looked  really  good. We  were  excited  about  this.  We  wrote up  a  paper  and  put  it  on  archive. And  then  courtesy  of  the  whole  review  process  in  science, which  I  oscillate  on, thinking  is  an  amazing  thing  and  a  terrible  thing. This  was  maybe  one  of  those  instances where  it  was  actually  good  for  scientists. We  kept  hitting  two  walls. Reviewers  kept  asking  us  for various  things  that  were  perfectly  reasonable. Asks  in  particular, one  of  the  things  that  they  kept  asking  us  for  is,  well, tell  us  about  your  hyperparameters, how  they're  being  selected, how  robust  is  this  to hyperparameter  selection,  et  cetera. The  honest  answer,  once  we  really forced  ourselves  to  explore  this  more  thoroughly, was  that  it  was  not  robust  to  hyperparameter  selection. It  was  super  sensitive  to  hyperparameter  selection. You  would  get  radically  different  answers  depending upon  the  hyperparameters  that  you  selected. And  the  places  where  you  got  good  results  were  tiny, tiny  dinky  bits  of hyperparameter  space,  which  is  not  good. It  means  that  people  aren't  going  to  be  able  to practically  use  this  system  in  reality. The  other  thing  is  we  weren't  getting good  transfer  across  sessions. So  another  problem  that  I  think  we  face, and  I'm  going  to  come  to this  in  the  final  part  of  the  talk, is  if  we  really  want machine  learning  tools  to  carry  some  of  the  heavy  lifting for  us  in  terms  of  our  inability to  understand  how  neural  systems  work. We  want  to  be  able  to  apply  these  machine learning  systems  to  all  individuals,  right? It  shouldn't  be  that  I  have a  different  model  for  every  person  and  every  animal. I  want  a  single  model  for  everyone. Because  I  want  to  put  that  neural  activity  into a  universal  latent  space  that  tells  me something  that's  true  about  any  brain, not  just  a  brain. So  we  really  wanted  to  apply Val  Paca  to  many  different  animals. And  we  tried  and  we  tried,  and  we  tried, and  it  just  always got  worse  when  we  applied  it  to  multiple  animals. Every  time  we  put  in  another  animal  into the  pipeline  with  a  different  tweak to  the  way  we  were  inputting  the  data. It  still  just  got  worse  in  frustration. At  some  point,  my  postdoc  had gone  off  and  gotten  a  job  at  Graph  Core. And  we  were  in this  review  process  and  at  some  point we  were  just  like,  okay,  suck  it. We're  not  doing  this  anymore. We're  not  going  to  try  to  get this  paper  published.  It's  on  archive. A  part  of  it  was  even  like,  should  we retract  it  But  like,  I  don't  know, it's  not  quite  a  retraction  territory  because  it's  not that  anything  we  said  in  the  paper  was  false. It's  just  that  it  doesn't  like  articulate the  ways  in  which  this  isn't  quite  a  satisfying  system. So  we're  just  going  to  leave  it  there. But  we  decided  to  abandon  this  project  essentially. Now  I  should  note,  Cheathens  Group  has  also  done some  work  on  extending  alphads  to  calcium  imaging. Data  radical  looks  really  cool  in  many  ways, but  I  don't  know  if  you  guys  maybe  find  some  of the  same  problems  that  we  found  here with  sensitivity  to  hyperparameters and  inability  to  transfer. I  think  it  might  be  unfortunately  like a  core  problem  with the  variational  auto  encoder  approach  to  these  issues. I  think  VA's  might  be very  sensitive  to  some  of  this  stuff, but  I  don't  know,  That's  speculation.  Okay. One  of  the  things  that  I  thought could  also  be  a  problem  here, and  this  might  actually  be  more  of  the  problem, is  an  inability  to  scale. This  gets  to  the  transfer  issue, machine  learning  systems. I  think  the  most  important  lesson  in  AI over  the  last  15  years  has  been what  Rich  Sutton  calls  the  Bitter  Less  if  you're unfamiliar  with  the  Bitter  bitter  lesson  is  as  follows. Over  the  course  of  the  progression of  artificial  intelligence, what  researchers  have  repeatedly  encountered  is that  though  it  is important  to  be  clever about  the  way  that  you  engineer  your  systems. Though  it  is  important  to  have neural  networks  or machine  learning  systems  that  are  designed  well. At  the  end  of  the  day,  any  system  will  only  show  you what  it's  truly  capable  of  when  you  scale up  the  amount  of  compute  that  you  throw  at  the  problem, when  you  scale  up  the  size  of the  model  and  you  scale  up  the  size of  the  dataset  that  you  feed  into  it  in  AI, over  the  last  five  years, almost  every  major  advance  has  actually been  just  because  of  increasing  the  scale  of  the  models. This  has  been  what  open  AI  specializes  in,  right? Like  they  didn't  invent  an  to  be  clear,  right? Like  open  AI  did  not  invent  transformers. They  didn't  invent  a  new  loss  function, they  did,  there  are  a  couple  of  architectural  tweaks, but  what  they  do  have  is  the  knowledge, the  know  how  to  increase  the  size  of  the  models, which  is  very  non  trivial,  to  be  clear. I'm  not  here  to  trivialize  that. Actually  building  a  model  with 1  trillion  parameters  is  insane. It's  really  challenging  to  do,  but  if  you  do  it, if  you  successfully  do  it, you  get  way  more  bang  for your  buck  out  of  your  ML  systems. I  think  the  bitter  lesson  surely applies  to  the  analysis  of  neural  data  as  well. I  am  sure  that  we  will  get much  more  out  of  our  data  if  we can  drastically  increase  the  scale  of our  models  and  train  across  all  the  data  out  there. I  started  asking  myself, how  could  we  maybe  do  this? How  could  we  train  on  literally  every  piece  of  data? So  I  was  thinking  about  the  calcium imaging  problem  and  I  was  like,  well, okay,  the  problem  is  that every  animal  has  different  neurons.  All  right. So  like  I  record  a  bunch  of  neurons  in  one  animal, I  record  a  bunch  of  neurons  in  another  animal. And  who's  to  say  that  those  neurons  relate  to each  other  or  have  anything  to do  with  each  other  necessarily? Like  there  are  various  approaches  for  potentially  doing. Like  David  with  Elphads  originally  said,  okay, well  we're  going  to  do  multi  session  stuff by  feeding  different  sessions  in, through  different  linear  encoders to  get  you  into  the  initial  embedding. That's  great,  but  even  that  assumes  that  all  of the  neurons  are  somehow  just  a linear  transformation  of  each  other. It's  not  clear  that  that's  true. It  means  that  you're  operating, once  you  shift  to  neurons, you're  potentially  operating  in  a  space  where, as  it  were,  the  words being  used  by  the  system  are  different. It's  like  you've  got  different  systems speaking  different  languages. And  you're  trying  to  feed  all  of these  things  into  one  model. And  it  doesn't  really  work,  I  thought. Okay,  what  if  we  just  keep  it  really  native? All  calcium  imaging  receives video  data  or  has  video  data. You  record  a  video  of  the  thing. What  if  we  just  analyzed  the  video  data? We  built  this  system. This  was  basically  a  cookie  idea where  what  we  did  is  we  took raw  calcium  videos  and  we'd  feed them  through  a  three  D  deep  convolutional  neural  network. Then  we  would  use  an  unsupervised  technique  called Contrastive  predictive  coding  to predict  the  next  step  of  the  calcium  videos. Basically,  we  feed  in  a  set  of  calcium  videos. This  a  latent  representation just  generates  a  contextual  representation  up  here, which  generates  a  latent  representation  we compare  to  the  actual  z that  we  get  from  the  data  in  the  next  time  step. And  then  we  do  contrastive  learning, meaning  that  we  push  z  hat  and  z  together  and  we repel  z  from  other  zs from  previous  times  steps  or  future  time  steps. This  form  of  contrastive  learning  works  quite well  then  the  idea  is  you'd  freeze  your  network, feed  your  calcium  imaging  data  into  that  network, and  then  decode  things  or. Diseases  or  whatever.  Let  me  give  you a  concrete  example  here. We  had  some  data  that  we  did  in  a  pilot  study  with Mark  Brandon's  lab  at  Mcgill. They  had  done  a  bunch  of  calcium imaging  recordings  in  the  CA, one  of  mice  who  were  exploring  mazes  of  different  shapes, they  basically  had  this  modifiable  maze, so  they  could  put  little  blocks into  it  to  create  different  shape  mazes. And  they  ran  the  animals  over  many, many  days,  over  a  month. Just  putting  them  into  these  different  shape  mazes and  collecting  calcium  imaging  data. In  the  CA  one,  we trained  this  neural  network  that I  just  showed  you  on  this  data, we  basically  just  took  the  raw  video, raw  single  photon  videos  from CA  one  fed  it  into  a  big  three  D  convent did  contrast  of  learning  on that  and  then  we  looked  at  our  ability to  decode  which  of  these  environments  the  animal  was  in. What  we  were  after  here  was good  multi  session  training  because  as  I  said, I  think  what  we  really  need  are  models that  apply  across  individuals. I'm  done  with  models  that  only  work  for one  individual.  It  seemed  to  work. They  had  just  recorded  five  mice  for  us. We  looked  at  if  you  just  train  this  model  on  one  mouse, and  you  look  at  the  confusion  matrix  here. Here's  the  ground  truth  of  which  of those  ten  environments  where  the  mouse  was  in. Here's  what  was  predicted. If  you  train  on  just  one  mouse, it  looks  pretty  sloppy. But  if  we  trained  on  four  mice, we  could  then  transfer  to a  new  mice  and  get  pretty  good  decoding  on which  environment  they  were  in  that's  just fed  off  of  the  raw  calcium  data. I  was  pretty  excited  by this.  I  was  like,  okay,  this  is  it. We  can  train  across  every  calcium  imaging  dataset  ever. We're  just  going  to  feed  the  raw  videos  from that  calcium  imaging  into  this  one  model. It'll  be  huge,  but  it's going  to  decode  all  the  stuff  for  us. But  that  didn't  really  work. I  think  the  principle  was  good. I  think  the  principle  was  good. But  we  hit  the  following  roadblocks. First  of  all,  it  was  insanely  compute  intensive. Doing  three  D  comms  on calcium  imaging  was  like  just  killing  our  GPU's. We're  not  open  AI,  we  don't  have thousands  of  InvidioGPU's  to  throw  at  this  problem. At  some  point  my  student  was  like, yeah,  like  this  is  cool  Blake, but  I'm  tired  of  wasting  my  life on  this  watching  the  GPU's  run. I  was  like,  okay,  that's  fair. Then  of  course,  the  other  issue  is  it's  not  going  to necessarily  work  for  other  modalities. Like  maybe  it  would  work  for  other  imaging  modalities. You  could  even  think  about  doing a  version  of  it  with  things  like  neural  pixels, where  you  treat  the  feature  like each  electrode  as  a  pixel  and you  just  treat  it  as  an  image  that  you  work  with. But  again,  it's  going  to  be  super  compute  intensive and  whether  it  would  work  well with  that  system  is  unclear. We  gave  up  on  that  as  well. Let  me  now  bring  you  to  a  bit  of  hopefulness. After  years  of  being frustrated  by  these  various  endeavors, I  think  I've  finally  come  to  what  is a  more  general  approach  and  it's  all  about  being  good, about  how  you  design  tokens  to  feed  into  a  transformer. Let's  talk  about  that.  This  shift  in  mentality  came about  for  me  because,  okay, to  reiterate  our  thing  here  again, what  I  really  want  is  I  want  to  be  able  to train  across  many  different  datasets containing  many  different  subjects. Like  if  I've  got  a  model  of  one  monkey  doing a  reaching  task  and  I  can say  something  about  the  neurodynamics  going  on  here. That's  all  very  well  and  good, but  what  I'd  really  like  to  be  able  to do  is  to  not  just  train  on  this  one  individual, but  I  want  to  be  able  to  zoom out  to  every  possible  individual. I  want  to  train  on  monkeys doing  different  reaching  tasks. Maybe  non  reaching  tasks, I  want  to  train  on  datasets  from  every  lab  in  the  world. Because  I  want  scale PT  was  trained  on  trillions  of  words. I  need  trillions  of  spikes. I'm  not  interested  in  your  dataset  with  five  monkeys. I  want  every  dataset  that's ever  been  collected  in  neuroscience. How  am  I  going  to  possibly  do  that? Well,  the  answer,  luckily, oh,  this  is  okay. So  the  answer,  I  think  lies  in  resolving  this  problem. As  I  already  articulated, the  problem  that  we  face  is  that  the  neurons  that  we record  in  any  two  sessions  are  not  identical. Nor  are  they  necessarily  a  linear  combination of  some  underlying  features,  right? You're  dropping  electrodes  into  a  system with  hundreds  of  millions  or  billions  of  neurons, depending  on  the  species  you  work  on. The  chance  that  you  just  happened  to sample  a  similar  set  of neurons  in  two  individuals  is  like  infinitesimally  small. You  have  to  accept  the  fact  that  you've  gotten different  neurons  in  every  recording  and You  know,  if  you  have  a  vanilla  sequence  model where  you're  feeding  each  neuron  in, as  a  separate  channel  to  your  model, which  is  the  traditional  approach,  right? We  treat  each  neuron  as a  separate  dimension  in  our  input, a  separate  channel,  and  we  feed  that  into  our  model. And  now  I  give  you  a  different  set  of neurons  from  a  different  individual. Which  channel  do  these  go  to? And  how  am  I  going  to  figure  that  out? And  should  there  even  be  a  similar  channel  they  go  to? What  if  they're  totally  different  channels? It  seems  a  very  intractable  problem. And  this  was  why  I  had  wanted  to move  to  raw  imaging  because  I  was  like,  oh,  well, you  don't  have  to  worry  about  it  if you  just  feed  the  raw  image and  let  the  ML  system  discover  it. But  as  I  said,  it  doesn't  scale  well. I  spent  some  time  feeling  frustrated  with  this  problem. And  then  the  answer  came  about  courtesy of  my  colleagues  here  at  Georgia  Tech, Eva  Dyer  and  Mediazabu, their  solution  they  called. When  I  first  saw  this  acronym,  I  was  like,  guys, come  on  But  I've actually  really  come  to  love  it because  it  really  does  describe  the  model. Well,  if  you  don't  know,  I  guess  Polo  is the  sound  that  Kirby  makes  when  he  eats  things. As  I've  articulated  to  you, our  philosophy  here  is  we  want  to  eat  all  the  data. This  model,  Poo,  is a  really  clever  solution  to that  problem  of  different  neurons. Let's  talk  about  this  now. The  core  idea  is  to  tokenize  neural  data  in  the  way  that we  do  language  and transformers  in  transformer  models  and  machine  learning. What  you  do  is  you  have  your  word  and  it's  word  number, whatever  out  of  your  lexicon. You  run  that  through  a  very  simple  embedding  system to  transform  that  word  into a  vector  that  a  neural  network  can  work  with. That  vector  becomes  your  token that  you  do  your  sequence  modeling  on. And  in  language,  of  course, there  are  some  reasonable  and  obvious  ways to  tokenize  things. You  can  do  some  very  simple  clustering on  sequences  of  characters. And  there  are  certain  sequences  of characters  that  are  always  going  to  pop  up and  that  are  consistent  across  text  that  you  receive, no  matter  what  the  text  is. But  this  is  the  core  idea. Then  each  of  your  words  becomes  a  vector  in this  space  and  you  do  sequential modeling  over  these  vectors. What  Eva  and  Meti  said  is, well,  what  if  we  could  do  something  similar  with  neurons? What  if  we  had  a  way  of  embedding  unit through  into  this  fx  and  then  working  with  that. Now  in  practice,  in  principle  that  sounds  good, but  in  practice  the  issue  is  the  one  I  just said  which  is  the  embedding is  somewhat  obvious  for  words, you  know,  it's  clear  like  across  text. Okay,  like  the  word  dog is  going  to  be  the  same  word,  dog. Roughly,  maybe  sometimes  the  underlying  meaning  shifts. But  like,  it's  not  unreasonable to  treat  the  word  dog  like the  same  token  in every  piece  of  text  you  train  your  neural  network  on. But  it's  not  reasonable to  say  that  whatever  happened  to  be unit  911  in  this  recording is  the  same  across  all  of  your  recordings. That  would  be  stupid.  So  how  do  you  deal  with  this? This  is  the  tokenization  of  text. We  receive  our  sentence, we  chunk  it  up  into  these  vectors  that we  receive  into  our  large  language  model. Instead,  what  even  Mei  said  is  they  said, let's  not  tokenize  the  neurons, let's  tokenize  individual  spikes. Let's  do  so  in  a  way  that the  transformer  learn  about the  relationship  between different  neurons  and  different  datasets. Such  that  they  can  be  radically  different  neurons, but  maybe  they  have  some interdependent  relationship  between  them. What  they  did  is  they tokenize  spikes  rather  than  neurons. They  take  a  variable  that  indicates literally  just  the  ID  of the  unit  paired  with  the  recording  ID. You  basically  just  say  like  this  was  neuron 1029  in  recording  516. You  also  add  to  that  timestamp, you're  going  to  take  this  information  and  you're  going  to run  this  through  a  learnable  embedding. In  this  way,  every  spike  of  every  neuron  can become  its  own  custom  word  for  that  dataset. It's  like  you  can  learn  a  separate  lexicon for  every  dataset. This  was  like  for  me,  I  was  like, when  I  saw  this,  Boom. Okay,  I  get  it.  Yes,  that's  going  to  work.  It  does  work. Here's  an  illustration  of  the  process. Every  spike  receives  a  token. And  note  that  now  it  doesn't  matter  if  you've  got different  neurons  in  different  datasets,  right? Because  I  don't  need  a  set  of channels  that  have  to  be  the  same  across  datasets. Every  spike  I  receive, I  just  generate  a  new  token. This  is  this  is  a  spike  from  this  unit  in  this  recording. It  occurred  at  this  time.  That's  it. And  then  you  feed  in  a  sequence  of tokens  into  your  neural  network. The  architecture  looks  like  this. Initially  what  we  do  is  we  take those  timestamp  and  unit  ID  embeddings,  or  rather  ID's. And  then  we  run  those  through a  learnable  embedding  system that  uses  a  cross  attention  mechanism. Where  there's  a  series  of learned  latents  for  that  embedding and  those  get  compared  via cross  attention  with  your  spikes. That  comparison  is  what  allows  the  system to  a  flexible  language  for treating  spikes  from  different  neurons  in different  recordings  as  being somehow  related  to  one  another. Even  if  they're  not  identical  to  each  other. Or  a  linear  combination  of  them. Because  cross  attention  is  a  very  flexible  mechanism. It  allows  the  system  to compare  across  all  the  spikes.  Right? So  that's  the  thing  it's  going  to  do, cross  attention,  and  it's  going  to  look  across,  okay, for  all  these  spikes,  which  spikes  should be  treated  the  same  or  different  or  whatever? And  learn  this  flexible  lexicon. After  running  through  the  cross  attention, you  then  have  a  series  of  self  attention  layers  per a  standard  transformer  model. Then  in  this  case, what  Eva  and  Medi  were  interested in  was  doing  direct  behavioral  decoding. We're  not  doing  unsupervised  latent  space  recovery  here, we're  just  doing  direct  behavioral  decoding. They  also  feed,  in  some  tokens, requesting  information  about  a  particular  behavior. This,  again,  is  important  for  being  flexible  across recording  sessions  because  that  way  you  can  have a  different  behavior  in  every  recording. Run  that  through  one  more  cross  attention  mechanism, before  finally  outputting the  predicted  behavior  that  you  have requested  in  an  initial  set  of  tasks. We  examined  this  across  a  range  of  these  reaching  tasks. The  first,  I  should  just  say, yes,  I  forgot  I  included  this. If  you  just  look  at  a  single  session, even  polio  does  really  well  often  here, this  was  just  training  on  a  maze  reaching  session. If  you  compare  to  a  lot  of  the  other  techniques for  some  good  latent  representations or  doing  direct  decoding, Poo  is  ultimately  better  at extracting  the  behaviors  that are  occurring  in  this  system. But  I  think  for  me, the  even  more  exciting  thing  is  that Poo  works  with  multiple  sessions. As  I  said,  what  you  can  do  is  you  can  treat  each  of these  behaviors  as  a  different  set  of  queries, each  of  the  recording  sessions as  a  different  set  of  tokens. You  batch  these  all  up  together and  feed  them  into  your  transformer. It  actually  works  here  we're  looking  at decoding  in  single  sessions versus  multi  sessions  for  a  series  of different  datasets  collected  by  different  labs  as  well. We're  not  even  in  a  situation, we're  in  a  single  lab. These  are  different  datasets collected  across  different  labs, different  monkeys  doing  different  reaching  tasks. For  me,  the  key  thing  is  that  in  each  of  these  plots, the  multi  dataset  model does  better  than  the  single  session  model. And  that's  exciting,  because that  means  scaling  is  working. That  means  we  can  take  advantage  of  the  bitter  lesson. Because  if  it  gets  better, when  I  add  more  monkeys  in, even  if  they're  doing  different  tasks, even  if  it's  from  a  different  lab, even  in  some  cases  it's  in  a  different  brain  region. That  means  I  can  now  train  on  all  the  data. I  can  just  hover  it  all  up. And  my  model  should  keep getting  better  and  better  and  better. And  we  can  pull  an  open  AI  and  just  go massive  on  scale  and eventually  get  all  the  stuff  we  want. This  is  illustrated  here  also with  respect  to  the  scale  of  the  model. Here  Medi  was  exploring  what  happens when  you  increase  the  scale  of  your  model in  terms  of  the  number  of spikes  that  you're  feeding  into  your training  set  and  the  number of  layers  that  your  model  has. We're  looking  at,  again, the  R  squared  with  the  actual  behavior. So  this  is  our  ability  to  decode  behavior. This  is  what  we  get  from  a  single  session  model on  this  particular  reaching  task. Then  we  have  all of  the  different  random  target  reaching  tasks. And  you  can  see  moving  from a  single  session  to  all  of  the  random  targets, you  get  this  very  clear  increase in  your  ability  to  decode. Now  we  add  a  set  of  center  out  reaching  tasks,  boom. Again,  increase  your  ability  to  decode. Here  you  can  see  you're  increasing  the  number  of  spikes. Theoretically,  this  curve  could  keep  going,  right? We  could  keep  going  on  this out  to  1  billion  spikes,  1  trillion  spikes. And  that  R  square  should  keep  going  up. In  particular,  it  should  keep  going up  if  we  increase  the  size  of  our  model. So  that's  the  other  key  thing  to really  take  advantage  of  that  data  scale. You  need  big  models  and  that's  again, a  key  thing  that  we  saw  in  polio, is  that  when  you  increase  the  number  of  layers, you  also  get  these  additional  benefits. Okay,  it  looks  like  we're  really  in  a  good  place  then. For  taking  to  heart  the  bitter  lesson. Here  even  we  did,  this  is  a  fun  little  example. Transfer  across  species  met and  a  guy  who  works  with  us  at  Mila, they  trained  a  model  on  all  that  monkey  data, decoding  monkeys  doing  the  reaching  tasks. And  then  they  applied  it  to  a  dataset from  will  it  it  all  of  people  who had  Utah  rays  in  their  motor  cortex  and were  asked  to  imagine  writing  characters. And  the  task  is  to  decode what  character  they  were  imagining. If  you  train  a  model  of equivalent  architecture  on just  a  single  session  from  a  human, these  are  the  sort  of  error  rates  that  we get.  Not  bad,  okay? You  can  get  an  error  rate  of around  15  to  20%  but if  you  pre  train  polio  first  on  all  the  monkey  data, you  crunch  that  error  down. And  that's  super  exciting  because  that  means  not  only can  we  train  on  all  the  data  from  a  given  species, but  all  the  species. Now  I  can  train  on  all  the  mouse  data, all  the  monkey  data, and  use  it  in  human  beings. For  me,  that  was  like  super  exciting. As  an  aside,  we  ended  up  deciding  to apply  this  to  calcium  imaging  data  as  well. We  came  back  to  this  data  set  and  this  is a  fun  example  of  how  it  can  be  beneficial. This  is  really  hot  off  the  press  data. The  next  couple  of  plots  are  going  to be  shambolic  aesthetically,  But  anyway, we  had  this  data  where  we  had  done recordings  in  both  the  dendrites  and the  somata  of  mice observing  these  sequences  of  these  Gabor  frames. One  of  the  things  we're  interested  in  is, could  we  do  multi  session  training across  dendrites  and  somata. Like  in  principle,  that  sounds  a  little  bit  crazy. Could  you  train  a  single  model  on data  from  both  dendrites  and  somata? Why  the  hell  should  that  work? But  I  think  poo  makes  it  possible. The  poo  philosophy  that  we  adopted  here  is  we  said,  okay, what  we're  going  to  do  is  we're  going  to  notice the  fact  that  in  the  dendrites  versus  the  Somata, you  have  very  different  shapes. To  your  ROI's  right, dendrites  have  long  ROI's with  relatively  little  area  in  fact, whereas  Somata  have  large  area,  nice  round  ROI's. So  let's  say  that  you  provide  this  information  to  Poo, you  don't  just  tell  it  here an  event  occurred  at  this  time. It  was  from  this  neuron,  it  was  from this  recording,  but  you  also  say, and  it  was  from  this  ROI that  has  this  width,  height,  and  area. We  prepare  now  a  set  of additional  features  to  feed  into  Poo, also  the  height,  width, and  area,  and  the  position  within  the  imaging  frame. Now  our  input  to the  Poo  embedding  scheme  becomes  the  spike. Spike  type  variable,  the  value  embedding, this  is  the  size  of  the  calcium  transient. Then  our  ROI  positions  and  ROI  features, these  are  all  concatenated  together. The  spike  ID  gets  added  to  these. And  this  is  what  then  gets  run  through that  cross  attention  mechanism  to develop  your  embeddings  for  the  transformer. If  we  do  that,  we  actually get  good  transfer  across  somatic  and  dendritic  sessions. This  is  a  funny  plot, but  I'll  break  this  down  very  briefly  for  you. One  of  the  things  we  found  with  this  data is  that  an  MLP  by  itself could  actually  do  pretty  well  on  a  single  session. It  was  actually  a  pretty  decent  model. We  were  having  trouble  often  with  Val  Paca,  for  example, to  get  good  decoding  of  the  orientations  that  could beat  just  an  MLP  trained  on  a  single  session. And  we  thought,  oh,  well,  if  we  could  do multi  session  stuff,  we'd  do  better. But  we  kept  struggling  with  that. But  what  we  find  is  that  with  polio, with  these  additional  features  of  the  ROI  is  added  in, now  we  can  get  reasonably  well  past  that  MLP  level. The  other  thing  is  this  is  not sensitive  to  hyperparameters. We've  basically  done  no  hyperparameter  tuning  here. It's  just  taken  from  the  initial  hyperparameters we  had  from  spiking  data. This  just  seems  to  work  out  of  the  box. It's  because  I  think  the  transformer  is  able  to  learn this  flexible  lexicon  for  different  neural  data. And  it's  a  really  powerful  approach,  I  think, for  doing  multi  session  stuff  now. It  means  we  could  train  across  dendrites, across  brain  regions,  across  species,  across  tasks. Boom,  bitter  lesson  can  finally  be  taken  into  account. That  brings  me  to  the  end  goal, and  I  know  I  need  to  finish  up. Where  I  ultimately  want  to  see  us  move towards  is  a  foundation  model  for  neuroscience. There's  a  lot  of  data  out  there. I  just  want  to  emphasize  that  for  everyone, if  we  consider  all  the  different  modalities like  it's  just  insane. How  many  hours  of neuro  data  have  been  recorded  and  freely  released. Almost  none  of  which  is getting  used  in  a  single  model,  right? The  challenge  is  the  fact  that  each  brain  is  unique. They're  not  speaking  the  same  language. We  need  a  universal  translators  for brain  that  can  take account  of  the  ways  in  which  each  neuron  is  different. Each  behavior  is  different, each  stimulus  is  different, and  each  recording  modality  in  species  is  different. I  think  that  if  we  had  a  universal  translator  for  brains, this  would  allow  us  to  finally make  sense  of  all  this  diverse  data. Because  what  it  would  do  is  it  would take  that  neural  activity  data  and  put  it  into a  universal  latent  space  that  we  could  then use  for  decoding  diagnosis, treatment,  prediction,  control  of neural  activity,  et  cetera. And  you  wouldn't  have  to  worry about  the  fact  that  you  didn't  have enough  data  for  any  individual because  this  is  the  key  problem. Theoretically,  this  would  work  on  a  single  individual. If  I  could  record  hundreds  of  thousands of  hours  from  that  individual,  I  give  you  a  Utah  array. And  now  I  want  you  to  spend the  next  ten  years  of  your  life just  giving  me  data  to  train  my  ML  model. That's  unfeasible.  But  what  is  feasible  is  if  I  have a  system  that  is  trained  on all  of  that  other  data  out  there  first, that  I  can  fine  tune  on  it rapidly  with  just  say  an  hour  or two  of  data  from  a  single  human  and  get  good  performance. And  what  I'm  really  articulating  here  is  the  vision  of a  Neuro  foundation  model  in  AI  research. A  foundation  model  refers  to a  large  neural  network  that's  been  pre trained  on  huge  amounts  of diverse  data  and  then  fine  tuned  for  downstream  tasks. And  I  think  that's  really  what  we  want  to  do  here. We  want  to  feed  in  all  of  the  different  types of  data  into  a  single  model. Pretrain  it  in  some  way, possibly  in  a  self  supervised  manner, and  then  have  it  available  for fine  tuning  for  these  various  different  applications. Drug  efficacy  prediction,  disease  risk  prediction, brain  computer  interface,  whatever. Okay,  to  finish  up, I'll  say  the  hope  is  that  if  we  did  that, we  could  then  after  pre  training  the  model, release  it  and  then any  group  around  the  world  could  use  it. This  is  my  ultimate  dream  is  that we  collectively,  as  neuroscientists, will  produce  this  large pretrained  model  that  could  be  used  by everyone  else  for  their  clinical  or  research  purposes. We  can  now  peer  into  brains with  ever  more  sophisticated  technology. It's  incredible  what  we  can  do. Young  me  would  have  been  so impressed  with  the  recording  technologies  we  have, but  I  don't  feel  like  we're  using  the  data  sufficiently yet  and  I  really  think  this  is the  direction  we  need  to  go  in  to  do  so. So  that's  that.  Thank  you  very  much  for  listening. Thank  you  to  all  my  collaborators  and students  and  of  course  my  sugar  daddies  for  funding  me. And  thank  you,  chiefin, for  inviting  me  and  having  me  here  at  Georgia  Tech. Yeah,  yeah.  It  seems  like  over  the  course  of  the  talk, there  is  a  pivot  from interpretability  being  something  to  strike  for which  methods  like  alps  built  into  the  Let's just  build  the  biggest  thing  possible  and get  the  biggest  tar  squared  in  decoding. Yeah.  So  is  interpretability  just  forget  about  it? No,  no.  Yeah.  You  very accurately  observed  that  trajectory. Because  that  has  been  my  personal  trajectory over  the  last  six  years. I  will  admit,  possibly to  the  chagrin  of  some  people  here. I  don't  know.  I've  given  up  on  interpretability. As  I  said  in  the  intro, I'm  not  convinced  that  interpretability  is possible  for  the  brain  is  basically  what  I've  come  to. And  I  think  there's  some  interesting  questions there  and  we  can  discuss.  Yeah. Because  I  think,  yeah,  so  in  some  of those  tasks  like  alphas  does  pretty  well, almost  as  good  as  polio, but  arguably  has  interpretable  latency. Absolutely.  We  sacrifice  it  all  for  like  4%  R  squared. Right?  Right,  absolutely. So  the  question  is whether  all  neural  circuits  are  going  to  be nicely  interpretable  like  some  of those  neural  circuits  that  we've applied  alpha  to  successfully. I  think  the  challenge  is  like  when  it  comes  to, so  a  lot  of  the  work, and  this  was  the  case  for  our  paper  on  Valpaca  as  well. You  take  a  system  where  you  already  know  the  end  answer you  want  and  then  you  show  you  can  derive  it  from. Right,  great.  But  what  if  you don't  know  what  that  end  answer  should  look  like? What  does  it  mean  to  be  interpretable? I  think  there's  some  real  challenges  there  now. Heathen  and  Nick  are thinking  hard  about  this  and  I  think  they've  got the  right  approach  for  starting  to  ask  this  question and  whether  it's  even  possible  For  the  record  guys, I  hope  your  competition  proves  me  wrong, that  in  fact  interpretability  is  possible. And  I'd  really  like  to  see  that  happen. And  maybe  it  is  for  the  record, like  this  is  just  a  sort  like  intuition. I  very  much  so  have  pivoted  to, you  know  what,  forget  about  interpretability, I  want  that  boost  of  4%  on  my  T, at  least  then  I  can  have  something  that might  help  people  in  the  real  world. And  that's  my  goal  to  some  extent. Yeah.  Sorry.  I  saw  a  very  quick  hand  shoot  out  there. But  Nick,  I'm  sure,  has  a  burning  answer  to  this. So  you're  on  the  side  of the  underpants  sno  to  switch  your  answer. Well,  almost  I  am  on  the  side of  the  underpants  gnomes  that  have  been  hired  at  open  AI, which  is  to  say  I  have  yes,  drunk,  the  cooled, and  now  believe  that  the  answer  can  be  as  simple as  you  need  to  pick  a  good  tokenization  regime, and  you  need  to  have  some  reasonable  model  designs, but  then  scale  should  be  able  to  bring  you  a of  the  way  You  got  it.  Sorry. Yeah,  you  had  a  question,  okay. And  you  motivate  your  talk,  right, With  all  the  amazing  implications that  possible  with  the  kind  of  approach. And  so  from  that  perspective, from  getting  practical  outcomes, it  makes  a  lot  of  sense. Same  thing  for  language  model.  They  can  do  things. Yeah.  Right.  But  neuroscientists, they  have  presumably  also  like a  basic  science  goal  of  the  still  in the  fundamental  principles  of  how  to  systems  work. Right.  Which  is  directly  related  to  interpretability. Are  those  separate  for  you? Do  you  just  like  not  care  about  the  basic  science  quest or  whatever  like  the  theory  of everything  in  the  brain,  Right? My  answer  to  that  is  that if  you  had  asked  me  that  question  four  years  ago, I  would  have  said,  oh  like  absolutely I  care  about  the  basic  science  quest. That  is  what  I  care  most  about, all  this  other  stuff  is  fluff. But  I  suppose  I  have  been  changing  philosophically, the  reason  is  twofold. One  is  my  fear  that  maybe  we might  just  be  upshitz  creek  in  neuroscience  and maybe  it's  the  case  that we're  not  going  to  be  able  to  fully get  an  easy  to understand  description  of  how  the  brain  works, but  the  other  desire  or the  realization  rather  that in  many  applications  in  this  world, many  of  the  technologies  that  have the  biggest  impact  on  people's  day  to  day  lives, people  did  it  without  the  basic  science  understanding. This  came  about  for  me  because  I heard  interview  after  interview  after interview  with  neuroscientists  when  we  were  talking  about like  brain  computer  interfaces  or diagnoses  or  drug  treatment,  et  cetera. And  they  would  say,  well,  they'd  say  to  the  journalist, the  problem  is  we  don't  understand how  the  brain  works  first. We  have  to  understand  how  the  brain  works. At  first,  I  was  always  going  along  with  that. I'm  like,  yeah,  you're  right,  you're  right,  you're  right. And  then  I  started  looking at  the  history  of  other  sciences, chemistry,  physics,  optics. These  people  would  do  stuff without  really  knowing  why  it  worked. But  it  worked  and  it  had  huge  impacts  on  the  world. This  is  a  very  cynical  take,  but  here  it  goes. If  we  want  that  tap  to  stay  open  for neuroscience  is,  no,  I'm  serious. If  you  want  to  keep  receiving those  hundreds  of  billions  of  dollars  of research  money  to  get to  that  basic  science  understanding, it's  time  to  start  delivering some  stuff  that  actually  helps  people. If  we  don't,  we're  not  going  to  keep  getting  the  funding. Guaranteed  politicians  are  going  to  pull  it  back. Even  though  some  part  of  me  really  likes  the  vision, I  want  to  see  that  basic  science  understanding, and  I  hope  I'm  wrong  about  interpretability,  I  really  do. I  think  we  still  need  to  get  on  with this  job  anyway,  right  now, because  we  need  to  start  producing some  concrete  results  for  the  funders. Yeah.  As  you're  trying  to  put more  and  more  data  types  is your  like  token  vector  just  can  get longer  and  longer  and  keep  adding  more  like modality  specific  features.  How  do  you  envision  that? Well,  you  don't  have  to  keep  increasing  the  length  of your  token  sequences  necessarily as  you  add  more  modalities, because  you  can  just  view  that  as almost  additional  text  that  you're  feeding  in. Right?  So  you  can  just start  to  take  in  more  modalities  and feed  those  in  as  if  they  were  additional  texts into  the  model  and  keep  your  context  window  the  same. The  only  thing  I'd  say  with  that  though is  the  mentionality  of  the  token  itself. Oh,  the  dimensionality  of  the  token  itself. No,  I  don't  think  that's  going  to  be a  big  bottleneck  depending  upon  the  tokens  you  design. And  this  is  where  token design  is  going  to  be  the  key  to  this. This  is  going  to  be  the  sort  of clever  engineering  thing  over  the  next  few  years. How  are  we  going  to  design  our  tokens? So  if  I  take  FMRI  data  or  EEG  data, what  should  the  token  look  like? You  want  it  to  be  a  token  that's somehow  commensurate  to  a  spike  of  some  sort. And  that's  how  you  want  to  start  to think  about  it  and  maybe  that's  going to  have  to  have  a  different  form. But  I  think  the  core  principle  is  there, Maybe  all  you  need  is  you  need  to  say, where  was  the  electrode  on  the  skull? Maybe  give  it  a  breakdown  of  some  of the  frequency  composition  at  that  moment  in  time. And  then  some  other  information  about  like it  was  this  subject  and  this  task,  et  cetera. And  you  feed  that  into  tokens  of probably  the  same  dimensionality. I  suspect  it  would  work  because  there's  probably enough  dimensionality  in  one of  these  tokens  to  handle  it. A  few  hundred  dimensions  is  enough  for that  level  of  information  would  be  my  guess. But  we'll  see.  I  think  we're  done. Okay,  everybody,  let's  think  quick.  Yeah.