>> IT'S MY PLEASURE TO INTRODUCE LARRY ABBOTT, LARRY IS THE CO DIRECTOR FOR THE CENTER FOR THEORETICAL NEURAL BIOLOGY AT WILLIAM, AND HE'S ALSO THEORETICALS IN HIS TITLE BECAUSE HE'S CLEARLY A THEORETICAL SCIENTIST. HE'S ALSO ON THE MULTICOUNCIL WORKING GROUP FOR THE BRAIN INITIATIVE WHICH TOM IS CHAIRING WITH STORY LANDIS, SO I THINK LARRY IS ONE OF THE FIRST GENERATION NEURAL SCIENTISTS AND HE'S HAD A BROAD IMPACT ON THE FIELD TO THE EXAT THE PRESENT TIME THAT PEOPLE THAT HAVE TRAINED THAT BROUGHT HIM BACK SO HE'S CONTRIBUTED A LOT OF THE FUNDAMENTAL IDEAS WE HAVE ABOUT HOW NETWORKS OF NEURONS IN THE BRAIN CAN FUNCTION. HE STARTED OUT IN THEORETICAL PHYSICS AND THEN HE STARTED WORKING WITH EVE MARTYR IN THE EARLY 90S TO BRING THESE PHYSICAL MODELS, TO THE DATA SHE HAD TO TRY TO GENERATE SOME INSIGHT INTO REALLY WHY SHE WAS OBSERVING THE KIND OF DYNAMICS SHE WAS OBSERVING, SO, YOU BE REALLY TO USE THESE THEORETICAL TOOLS FROM PHYSICS TO HELP EXPLAIN WHAT SHE WAS OBSERVING, HE MOVED ON AFTER THAT OR HE DIDN'T MOVE ON, HE CONTINUED TO DO THAT FOR A LONG TIME BUT HE THEN WORKED ON POPULATION CODING WITH AMELIA SALINASSPACE AND PETER DIANE AND THIS WORK HAD A HEAVY INFLUENCE ON ME IN THE LATE 90S, I WAS WORKING WITH THE MULTIELECTRICCAL RECORDINGS IN MONKEYS AND THE PAPER PROTHAT HE CONTRIBUTED WITH DIANE WAS A FORMAL DIALOGUE ABOUT HOW CORRELATIONS AMONG NEURONS CAN EFFECT BEHAVIOR AND FROM MY VIEW, I WENT TO WORK ON THE PROBLEM FOR ANOTHER SIX OR SEVEN YEARS BUT THAT EARLY PAPER PROVIED ONE OF THE STANDARD MODELS WE HAVE FOR HOW TO THINK ABOUT THESE THINGS. HE'S ALSO WORKED ON LOCAL PLASTICITY RULES LIKE TIMING DEPENDENT PLASTICITY TO UNDERSTAND HOW THESE LOCAL RULES CAN GIVE RISE TO AGAIN DYNAMICS IN NETWORKS, AND MORE RECENTLY, HE'S BEEN DOING WORK IN TRYING TO UNDERSTAND HOW NETWORKS TRAIN UP TO DO LET'S SAY COGNITIVE TASKS, CAN GIVE RISE TO THE KIND OF MEASURES IN THESE AREAS IN THE PREFRONTAL CORTEX. SO AN EARLY PIONEER IN TRYING TO GENERATE EXPLANATIONS FOR COMPLEX BIOLOGICAL PHENOMENON, REALLY EXPLANATIONS FOR THE KIND OF DATA WE MEASURE IN NETWORKS IN THE BRAIN. AND AND TODAY HE'S GOING TO TALK SOMETHING ABOUT LEARNING IN FLIES AND FISH. SO, LARRY? >> [ APPLAUSE ] NK YOU FOR INVITING ME AND THANKS FOR THE INTRODUCTION, THE NICE THING ABOUT NEURON SYSTEM YOU CAN STICK AROUND. HAVE YOU DON'T HAVE TO HAVE ONE PREPARATION AND YOU DON'T HAVE TO HAVE A LAB THAT DOES ONE PREPARATION AND I WILL TAKE ADVANTAGE OF THESE TODAY AND DISCUSS THESE TWO RATHER SCARY LOOKING CREATURES. >> I GUESS AND TRY TO CONTRAST THE ADVANTAGE OF BEING ABLE TO DO THAT AS YOU CAN TAKE THE LESSONS FROM ONE PREPARATION AND CARRY IT TO ANOTHER AND I HOPE TO DO THAT TODAY, AT THE RISK OF TRYING TO JAM AN AWFUL LOT INTO YOUR HEAD. SO I APOLOGIZE FOR THAT BUT I HOPE TO THE COMPARISONS WILL COME THROUGH AND THE SIMILARITIES AND SO I'M GOING TO START WITH THIS GUY HERE, OBVIOUSLY A FISH, BUT LET ME DO THE PEOPLE BEFORE I DO THE FISH. I'M GOING TO TALK ABOUT A LOT OF DATA TODAY DESPITE THE FACT THAT I'M A THEORIST, AND THAT'S NOT DUE TO ME. IN THIS CASE IT'S DUE TO NATE SAWTELLs, WITH THE DATA I'M ABOUT TO DESCRIBE AND TALKING ABOUT THE PEOPLE WORK N OTHING--WORKING IN NATE'S LAB AND GETTING AN IDEA OF WORK NOTHING THEORY AND EXPERIMENT OF BOTH OF THESE STUDENTS WERE MINE AND GREG AND ANN, AND GREG NOW WORKS AT DEEP MIND WHICH IS A DIVISION OF GOOGLE AND WORKS FOR DAVID ANDERSON AT CALTECH AND NOW IT'S A POST DOC AND PATRICK IS A MIXED THEORY EXPERIMENTAL STUDENT WHO WORKS WITH THE AND HE'S HE'S STILL A GRADUATE STUDENT BUT HE WORKS WITH TELLA AND ME AND, HE DID THE PAPER WITH NATE, SO THAT'S DISSOLVING OF THE BORDERS AND WHAT WE TRY TO DO. OKAY, SO, THIS FISH IS AN ELECTRIC FISH, CALLED AN ELEPHANT NOSE FISH, YOU CAN BE THE SEE TOO WELL IN THIS PICTURE BUT YOU'LL SEE WHY IN A SECOND AND ITS JOB IS TO FIND VARIOUS BUGS, WORMING IN THE WATER AND EAT THEM. AND O I'M GOING TO TALK ABOUT THE SYSTEM, THE TWO SYSTEMS THEY USE TO DO THIS. SO FIRST OF ALL, YOU CAN SEE THE LITTLE TRUNK WHICH IS FULL OF ELECTRORECEPTORS, AND THERE'S TWO SYSTEMS, VERY IMPORTANT FOR MY TALK THAT YOU APPRECIATE THAT THERE ARE TWO SENSORY SYSTEMS, AND ELECTRY SENSORY SYSTEMS THAT WE WILL TALK ABOUT. SO ALL OVER THE SKIN OF THIS FISH ARE SENSITIVE ELECTRORECEPTORS THAT ALSO PRESINENT NORMAL ELECTRIC FISH AND SHARKS FOR EXAMPLE AND THEY ARE SENSITIVE ENOUGH TO ACTUALLY DETECT THE ELECTRIC FIELDS GENERATED BY THE MUSCLE ACTIVITY OF A PREY IN THE WATER. SO VERY SENSITIVE RECEPTORS. NOW THE THING THAT MAKE THIS IS FISH AN ELECTRIC FISH IS THAT IT HAS IN ITS TAIL AN ELECTRIC GENERATOR. SO IT MAKES ITS OWN ELECTRIC FIELD AS WELL. THIS IS THE MODIFIED MUSCLE THAT PUTS UP ABOUT A 10 FOLD MILLISECOND LONG PULSE, HERE'S WHAT THE PULSE LOOKS LIKE AND THAT IS USED BY AN ACTIVE SYSTEM, THE SECOND OF THESE CEASES AND THAT'S WHAT MAKES THIS A WEEKLY ELECTRIC FISH SO THIS IS NOT A FIELD BIG ENOUGH TO HARM THIS PREY, BUT BIG ENOUGH TO SENSE IT. AND SO, BY THAT SYSTEM, WHAT IT DOES THROUGH THESE ACTIVE SENSORS, THAT ARE TUNED TO MUCH STRONGER FIELDS, IS IT SENSES THE DISTORTION OF THE SELF-PRODUCED ELECTRIC FIELD IN THE WATER DUE TO A PREY. NOW THAT'S A NICE SYSTEM AND OBVIOUSLY THIS PULSE IS THE KEY TO MAKING THIS ACTIVE SYSTEM WORK BUT AS FAR AS THE PATH OF THE SYSTEM GOES, IT'S KIND OF A DISASTER. I KIND OF LIKEN THIS TO, YOU KNOW IF YOU'RE IN A DARK ROOM USING YOUR WADS TO DETECT A DIM OBJECT AND SOMEBODY COMES ALONG AND STARTS FLASHING THE FLASH OF THE CAMERA, THAT WILL CAUSE A PROBLEM. THAT'S WHAT LIKE THIS IS. THIS FIELD IS SOMETHING LIKE A THOUSAND TIMES BIGGER THAN THE FIELD THAT'S SUPPOSED TO SENSE, SO THE SYSTEM I WILL TALK ABOUT IS THE SYSTEM BY WHICH THIS SYSTEM KEEPS WORKING DESPITE THE FACT THAT IT'S BEING SLASHED, PLASTED SEVERAL TIMES A SECOND BY THIS ACTIVE PULSE. AND I'LL COME BACK TO WHY THAT'S GENERALLY INTERESTING, BUT LET ME FIRST JUST SHOW YOU THAT THE SYSTEM ACTUALLY WORKS. SO WHAT'S COMING INTO THE PATH OF SYSTEM THEN IS A BIG SHOCK, AS I SAY SEVERAL TIMES A SECOND AND WHATEVER ELECTRIC FIELDS ARE BEING PRODUCED BY OBJECTS OF INTEREST. THEY COME IN THROUGH AFRIP FIBERS FROM THE RECEPTORS INTO A ELECTROSENSORY LOAD AND STRIKE FIRST MEDIAN GANGLION CELLS AND THIS--THIS STRUCTURE AS YOU'LL SEE AS A GO ALONG IS VERY SIMILAR TO--IT IS A CEREBELLUM BASICALLY. IT'S A SIMILAR TO OUR CELL BELLUM AND THIS IS THE CELLS OF THE CEREBELLUM AND THERE ARE OTHER CELLS LIKE THE DEEP CEREBELLAR NUCLEUS CELLS OF THE CEREBELLUM, THE UT OUTPUT CELLS OF THIS SYSTEM THAT GET AFTER INPUT AND GET THIS MEDIUM GANGLION SIGNAL AND WHAT I WILL DO IS SHOW YOU EXPERIMENTAL DATA RECORDED FROM THESE AFFERENTS, AND THESE COME IN FROM THE PIN UT AND THE OUTPUT. AND IF THIS CIRCUIT DOES ANYTHING, YOU SHOULD SEE THAT THE SIGNAL'S BETTER HERE THAN HERE. THAT'S WHAT WE'RE TRYING TO STUDY. SO I WILL SHOW YOU THAT QUICKLY. THIS IS VERY RECENT DATA, MINIMALLY ANALYZED BUT IT'S WORTH SHOWING YOU TO SHOW THE NATURE OF THE SYSTEM. SO IN THIS CASE, WE DON'T ACTUALLY HAVE A BUG OR NATE, AND COLLABORATORS DON'T HAVE A BUG IN THE WATER, THEY MAKE AN ELECTRIC FIELD THAT'S SUPPOSED TO SIMULATE THE FIELD OF THE BUG, THAT'S WHAT YOU'RE SEEING HERE, ALTHOUGH THIS IS THE FIRING RATE, THIS IS NOT THE FIRING RATE HERE, IN THIS TRACE, THIS IS A FIELD BEING PRODUCED BY A COIL IN THE WATER. THESE MARKS ARE WHEN THE FISH DISCHARGE THE ELECTRIC ORGAN, OKAY? YOU CAN PICK THAT UP FROM A NERVE THAT THE FISH SENDS A PULSE FROM ITS BRAIN DOWN TO ITS TAIL AND DISCHARGES THE ORGAN AND THAT'S WHAT'S HAPPENING AT THESE RED PULSES AND HERE'S THE--NOW THIS IS A FIRING RATE, DETECT FRIDAY THESE AFFERENT FIBERS, THE INPUT TO THIS CIRCUIT IN RESPONSE TO ALL THE STUFF GOING ON. AND WHAT YOU CAN SEE, I THINK PRETTY CLEARLY, IS THAT EVERY TIME THE FISH DISCHARGES ITS ELECTRIC ORGAN, YOU SEE ONE OF THESE WIGGLES. AND SO THIS--THIS AFFERENT IS PRIMARILY RESPONDING TO THE FISH'S OWN FIELD AND THEN WHATEVER'S PRODUCED BY THIS, WEAKER, YOU SORT OF SEE IT THERE BUT IT'S PRETTY DISGUISED. YOU CAN QUANTIFY THAT--LET ME SHOW YOU THE OUTPUT CELL FIRST. IF YOU NOW LOOK AT THE OUTPUT OF THIS CIRCUIT, YOU SEE A MUCH CLEARER VIEW OF THE BUG, THE FAKE BUG. WHAT YOU CAN SEE IS THESE LARGE RESPONSES CORRESPONDING TO THESE. THERE'S A LITTLE BIT OF RESPONSE TO THE SELF-GENERATED FIELD THAT HAVE BEEN GREATLY SUPPRESSED SO I THINK YOU AGREE THAT THIS IS A MUCH BETTER BUG DETECTOR THAN THIS. THIS IS THE ESSENTIALLY A SELF-FUELED DETECTOR AND YOU CAN QUANTIFY THAT BY DOING A STANDARD R. O. C. ANALYSIS, SAYING HOW WELL GIVEN THESE RECORDINGS AND REPEATS OVER MULTIPLE TRIALS, HOW WELL CAN WE DETECT THE PRESENCE OF THIS FAKE BUG BY THESE SYSTEMS. IF YOU DO IT ON THE AFFERENTS YOU'RE A BIT OVER 50% DETECTION THRESHOLD. IN FACT, IN OTHER WORDS YOU'RE BARELY BETTER THAN CHANCE. AND WITH THE OUTPUT OF THIS CIRCUIT, YOU'RE ALMOST AT A HUNDRED%. SO CLEARLY THIS CIRCUIT IS DOING SOMETHING TO IMPROVE THE SITUATION. NOW, THE SOMETHING SUGGEST ITSELF QUITE READILY FROM THIS RECORDING. I POINTED OUT THAT THIS RECORDING AND CORRUPTED BY THE CELL FIELDS AND OF COURSE THE CELL FIELD DOESN'T TELL YOU ANYTHING ABOUT WHETHER THERE'S A BUG IN THE WATER. SO, ONE THING YOU CAN IMAGINE DOING IS SUPPOSE YOU COULD SUBTRACT THE BUG, YOU CAN NOTICE THAT THEY'RE QUITE STEREOTYPE, NOT EXACTLY BUT QUITE STEREOTYPED SO WE CAN MATHEMATICALLY DO THAT, WE CAN TAKE THE AVERAGE OF THESE SELF-PRODUCED FIELDS, AND JUST SUBTRACT IT OUT FROM THIS TRACE. JUST ON THE COMPUT, AND WHEN YOU DO THAT, YOU SEE, YOU DO GET A MUCH BETTER IMAGE OF THE BUG. IN FACT, THIS ONE IS INVERTED. IN THE AFFERENT, ITVERTED INVERTED PICTURE BUT YOU GET A MUCH BETTER DETECTION. NOW I WILL COME BACK AT THE END AS HOW MUCH BETTER, I WILL SAVE THAT AS A SURPRISE/PUNCH LINE. BUT THAT'S THE IDEA. SO WHAT WE WILL DO IS LOOK AT THE SYSTEM BY WHICH THIS--THIS CIRCUIT, THE E. L. L. SUBTRACTS OUT THE EFFECTS OF THIS SELF-INDUCED FIELD FROM THE PASSIVE SYSTEM AND ALLOWS THE REAL SIGNAL TO GET THROUGH. SO RIGHT NOW, WE KNOW THERE'S BUG IN THE WATER. SO WE'RE JUST LOOKING AT THE EFFECT OF THE SELF-INDUCED SHOCK AND WHEN YOU LOOK AT WHAT THAT DOES, TO THE ELECTRORECEPTORS, IF E. O. D. WILL AEAR ON MY SLIDE, THAT STANDS FOR ELECTROORGAN DISCHARGE, THAT'S THE FIRING FIELD THIS, IS THE FIRING RATE OF ONE OF THE RECEPTORS, IT REACTS AND GOES INTO THIS LONG OSCILLATION AND IT'S BEEN WHACKED VERY HARD AND SO HAVE YOU ABOUT A 200 MILLISECOND LONG OSCILLATION THAT YOU WANT TO GET RID OF BECAUSE A BUG IF IT APPEARED WOULD SORT OF LOOK LIKE THIS. IT WOULD BE A VERY TINY LITTLE PIECE EVER THE SIGNAL. SO, I WILL TALK ABOUT THE SYSTEM BY WHICH THIS GOT SUBTRACTED. NOW, TO DO JUSTICE, I HAVE TO TALK ABOUT THE WORK THAT'S PARTICULAR ABOUT CURTIS BELL, CURTIS BELL STUDIED THIS SYSTEM THROUGHOUT HIS CAREER AND REALLY WHAT YOU'LL SEE THAT NATE DID AND WE DID WITH THE THEORY WAS TO PUT THE FINISHING TOUCHES ON A PROGRAM IS LARGELY CURTIS BELL'S AND IN ADDITION THERE'S REALLY BEAUTIFUL THEORY WORK BY PATRICK ROBERTS DONE THAT I WANT TO ACKNOWLEDGE AND I'LL TRY TO SAY WHERE WE'VE ADDED BUT YOU CAN SEE, WE PUT THE ICING ON THE CAKE BUT THESE GUYS BAKED THE CAKE. OKAY, I WANT TO SAY JUST ONE MORE THING ABOUT THE GENERALITY OF THIS. THIS SYSTEM IS A COROLLARY DISCHARGE SYSTEM. WHAT YOU HAVE IS SIGNAL GOING FROM THE BRAIN BACK TO THIS ORGAN, TELLING IT TO DISCHARGE, THIS ORGAN IS A MODIFIED MUSCLE SO THAT'S VERY SIMILAR TO THE KIND OF SIGNAL THAT WE GO BACK AND TELL THE FISH TO FLIP IT'S TAIL. NOW, HAVE YOU A COROLLARY DISCHARGE OF THAT SIGNAL THAT GOES TO THE E. L. L. CIRCUIT AND SAYS, BY THE WAY I JUST BLUE OFF THE ELECTRIC FIELD, YOU SHOULD EXPECT AN OSCILLATION LIKE THIS. SO THIS SYSTEM IS VERY SIMILAR TO THE KIND OF THINGS I DO IF I WIGGLE MY HEAD OR SOMETHING, I KNOW THAT THE--YOU KNOW THE ROOM IS NOT ROCKING BACK AND FORTH BECAUSE I KNOW I'M SENDING SIGNALS TO MY NECK MUSCLES, SO THE GENERAL ABILITY OF ALL ANIMALS TO REMOVE SELF-CREATED SENSORY EXPERIENCE AND REALIZE WHAT'S REAL IN THE EXTERNAL WORLD AS OPPOSE TO WHAT I'M DOING BY MOVING AROUND UP HERE. THIS IS AN EXAMPLE OF THAT, IN A PARTICULARLY CLEAR EXAMPLE TO UNEARTH THE CIRCUITRY THAT ALLOWS US ALL TO DO THINGS LIKE THAT. SO LET ME INTRODUCE YOU TO THE CIRCUITRY, AGAIN IN PARTICULAR THE HEAVY LIFTING ON THIS SUBTRACTION OF THE SELF-INDUCED FIELD IS DONE BY THESE MEDIAN GANGLION CELLS AND THE AN LOG OF THE CELLS IN THIS CEREBELLUM LIKE STRUCTURE. THEY RECEIVED AS I ALREADY TOLD YOU THIS INPUT COMING FROM THE SENSORY PERIPHERY AND EVERY TIME THERE'S AN ELECTRIC ORGAN, DISCHARGE THIS, IS THE IDEALIZED SIGNAL, YOU GET THIS OSCILLATION LASTING ABOUT 200 MILLISECONDS. NOW IF YOU DID NOTHING, THESE CELLS WOULD OUTPUT SOME VERSION OF THAT OSCILLATION, BUT IN THE CASE LIKE THIS, WHERE THERE'S NO BUG PRESENT, WHAT YOU WANT TO HAVE IS SOMETHING COMING OUT. THERE'S NOTHING INTERESTING OUT THERE, IT'S A SELF-INDUCED SIGNAL, NOW HOW DO YOU DO THAT? WELL, I MENTIOED THE COROLLARY DISCHARGE SIGNAL, THIS CIRCUIT HAS TO KNOW THAT THE MOTOR SYSTEM ACTIVATED AN ELECTRIC PULSE, AND IT GETS THAT SIGNAL THROUGH PHOSPHORYLATEDY FIBERS, I THINK YOU--ANYWAY, I'LL GO ON WITH THAT. SO YOU GET A COROLLARY DISCHARGE SIGNAL COMING INTO THIS CIRCUIT, IT SYNAPSES ON GRANULAR CELLS SO YOU START TO SEE THE TERMINOLOGY SOUNDING LIKE THE CEREBELLUM, THOSE GRANNUAL CELLS MAKE PARALLEL FIBERS AND THOSE PARALLEL FIBERS SYNAPSE ON THE CELLS AND THERE'S A FEW MORE PLAYERS IN THE GAME. THERE'S A CELL CALLED THE UNIPOLAR BUSH CELL, AND IT RECEIVES ADDITIONAL INPUT INTO THE GRANULE CELLS, THOSE GO DIRECTLY TO THE GRANULE CELLS AND INDIRECTLY THROUGH THE UNIPOLAR BRUSH CELLS AND THEN THERE ARE INHIBITORY, GOLGI CELLS, PLAYERS IN THE GAME. NOW THERE WAS ANOTHER PIECE PROVIDED BY CURTIS BELL AND HIS LAB, WHICH IS A FORM OF PLASTICITYOT SYNAPSIS BETWEEN THESE CELLS AND THE MG CELLS, BUT KIND OF AN ANTISPIKE TIMING DEPENDENT PLASTICITY. IT HAS TWO COMPONENTS. ONE HAVE YOU TROUBLE SEEING, NAMELY EVERY TIME THE GRANULE CELLS FIRE, THE SYNAPSE GETS A BIT STRONGER, SO IT'S JUST A TINY BIT AND THEN, IF THE PRESYNAPTIC SPIKEOT GRANULE CELL IS FOLLOWED BY WHAT'S CALLED A BROAD SPIKE IN THE MG CELL, YOU GET A STRONG DEPRESSION OF THE STRENGTH OF THE SYNAPSE, THAT'S THIS THING, ONLY WHEN IT'S PRE-BEFORE PULSE'D. SO THIS IS THE PLASTICITY WE'RE LEARNING, SO YOU SEE THERE'S QUITE A BEAUTIFUL CIRCUIT IN THE SENSE THAT YOU KNOW WHAT'S COMING INTO IT, AND YOU KNOW WHAT YOU WANT TO HAVE COME OUT OF IT, YOU KNOW THE PLASTICITY RULE, THE ONLY PIECE MISSING WAS THAT YOU DON'T KNOW WHAT'S HAPPENING ON THESE GRANULE CELLS. I WILL SHOW YOU THAT IN A SECOND. I GOT AHEAD OF MYSELF. SORRY. SO THE IDEA HERE IS THAT THE COROLLARY DISCHARGE SIGNAL, TELLS YOU THAT THE ELECTRIC ORGAN DISCHARGE AND THIS CIRCUITRY, GOING THROUGH THESE ADJUSTABLE SYNAPSES, HAS TO CONSTRUCT WHAT'S CALLED A NEGATIVE IMAGE, SO A NEGATIVE VERSION OF THESE WIGGLES SO THAT WHEN YOU ADD UP THIS AND THIS, YOU ADDA UP THIS PATHWAY AND THIS PATHWAY, YOU GET ZERO. SO THAT'S THE FUNCTION OF THIS APPARATUS, AND AS I WAS SAYING, YOU KNOW ALL THESE PIECES BUT UNFORTUNATELYOT THE TIME OF THE BEAUTIFUL WORK OF ROBERTS AND BELL, THERE WAS NO KNOWLEDGE OF WHAT WAS ACTUALLY GOING ON HERE. SO OF COURSE, THESE PARALLEL FIBERS FORMED THE BASIS FROM WHICH YOU WILL HAVE TO GENERATE THE NEGATIVE IMAGE. SO THAT'S WHAT NATE DID. HE RECORDED, IN FACT FROM THE WHOLE UPPER END OF THIS THING, THIS IS THE BUSINESS END AS FAR AS DEVELOPING THIS PREDICTION, BY THE WAY, THE REASON I'M SAYING PREDICTION IS BECAUSE THIS CIRCUIT IS SUPPOSED TO SAY WHEN THAT COREULARY DISCHARGE COME, I PREDICT THAT THE AFFERENTS ARE GOING TO REACT IN A CERTAIN WAY. SO HE RECORDED EXTRA CELLULARLY FROM THE MULTIFIBERS AND INTRACELLULARLY FROM THE UNIPOLAR BRUSH CELLS AND THE GRNULE CELLS QUITE A HEROIC AND DIFFICULT UNDERTAKING. OKAY. SO HERE'S WHAT HE GOT FOR THE GRANULE CELLS: THE GRANULE CELLS ARE INTERRACELLULAR RECORDING, THEY'RE NORMALIZED BUT VOLTAGE ON A NORMAL SCALE O YOU CAN SEE IT WELL AND THIS IS THE CRITICAL PERIOD FROM THE TIME OF THE ELECTRIC ORGAN, DISCHARGE OUT TO 200 MILLISECONDS, SO THIS IS THE BASIS FROM WHICH YOU'VE GOT TO BUILD A NEGATIVE IMAGE. NOW, SO THESE ARE VOLTAGES BUT I SHOULD ADD AND THIS IS ABOUT 170 CELLS I BELIEVE, OF COURSE WHAT YOU'RE LEGALLY INTERESTED IN THE SPIKES AND OF THESE CELLS ONLY ABOUT 10% SPIKE AND THEY ONLY SPIKE ONE OR TWO TIMES, SO WHEN I FIRST SAW THIS, I JUST THOUGHT, OH MY GOD, HOW ARE WE GOING TO EVER MAKE A NEGATIVE IMAGE FROM THIS SMALL QUANTITY OF ACTIVITY IN PARTICULARLY LATE, WHAT YOU SEE IS, THERE ARE A TON OF CELLS THAT REACT EARLY ON DIRECTLY TO DISCHARGE AND THAT'S FINE. HAVE YOU LOTS OF ACTIVITY FROM WHICH TO BUILD THE EARLY PART OF THE NEGATIVE IMAGE BUT YOU HAVE TO BUILD THESE OSCILLATIONS WAY OUT HERE AND YOU DON'T HAVE A LOT TO WORK WITH. I MEAN THESE CELLS ARE FIRING BUT THEY'RE SORT OF FIRING AT A CONSTANT RATE, THERE'S A LITTLE BIT HERE, BUT IT LOOKED PRETTY DISMAL. ANYWAY, WHAT--WHAT NATE HAS DONE IS ORGANIZE THESE CELLS BY HOW EARLY THEY FIRE AND THEN THERE ARE THESE CELLS THAT HAVE AN INTERCEPTION IN THE ACTIVITY, OTHERWISE THEY'RE TONIC, YOU KNOW A CERTAIN ORDERING OF THE CELLS. THAT'S WHAT THEY GOT. SO THIS IS WHAT YOU WILL HAVE TO BUILD THIS NEGATIVE IMAGE FROM. NOW FORTUNATELY FOR US, THEORORRISTS, WE HAVEN'T PLAYED A ROLE YET BUT YOU'LL SEE WHAT OUR ROLE IS, NATE ALSO RECORDED FROM THE INPUTS TO THESE GUYS. SO IF YOU LOOK ON THE MOSSY FIBERS, YOU SEE WHEY TOLD YOU WAS THERE. NAMELY A SIGNAL THAT SAYS, THE COROLLARY DISCHARGE, YOU KNOW THE DISCHARGE HAS BEEN MADE AND THAT'S THESE EARLY AND MEDIUM SIGNALS. YOU JUST DIVIDED THEM OUT BY HOW--HOW FAR THEY EXTENDED IN TIME BUT BASICALLY, THEY'RE AT THE TIME,OOSE THE COMMAND SIGNAL, ELECTRIC ORGAN DISCHARGE SIGNAL, QUITE EARLY ON. NOW WHEN YOU RECORDED HOWEVER FROM THE UNIPOLAR BRUSH CELLS YOU GOT MUCH MORE INTERESTING ACTIVITY PATTERN. SO THERE WERE CELLS WITH THE LATE RESPONSES AMONG THEM AND THERE WERE CELLS WITH PAUSE RESPONSES SO THEY HAD STEADY ACTIVITY THAT PAUSED AND THEN CAME BACK AND NOW, YOU'RE SEEING SOME TEMPORAL STRUCTURE, THINGS LOOK MORE INTERESTING IN THIS PERIOD WHEN WE NEED SOME ACTIVITY FROM WHICH TO BUILD A TIME VARYING NEGATIVE IMAGE. AND THAT WAS ALL IN THE UNIPOLAR BRUSH CELL. SO ONE SURPRISE IN THIS IS THAT THE HEROES OF THIS--OF THIS GAME ARE THE UNIPOLAR BRUSH CELLS, THEY'RE WHAT'S GOING TO MAKE THIS CIRCUIT WORK. THE INHIBITORY CELLS, THE GOLGI CELLS LOOK LIKE PERHAPS THEY MAKE THIS GAP IN THE UNIPOLAR BRUSH CELLS BUT OTHERWISE, THEY ARE NOT BIG PLAYERS IN THE LATE GAME. SO CAN YOU SEE RIGHT HERE, THE UNIPOLAR BRUSH CELLS WILL BE WHAT MAKES THIS SYSTEM WORK. OKAY, SO, NATE RATHER HEROICALLY RECORDED FROM 170 GRANULE CELLS, NOW, THE PROBLEM HERE IS THAT LIKE A CELL, THEY GET INPUTS, AND BECAUSE THEY'RE SPOA SPARSE IN SPIKING MEANS THAT YOU NEEDED ALL 20,000 FOR THIS GAME. SO NATE HAD THE DIFFICULT CHOICE OF RECORDING FROM, YOU KNOW, WHATTATE NUMBER, 19,830 MORE CELLS, OR CALLING IN THE THEORISTS, SO HE TOOK THE--HE SWALLOWED THE POISON AND CALLED US IN. ALL RIGHT, SO WHAT DID WE DO? BASICALLY WE HAVE TO TURN 170 CELLS INTO 20,000. THAT'S FORTUNATELY KIND OF EASY BECAUSE OF THE SIMPLE STRUCTURE OF THESE GRANULE CELLS, THEY'RE TINY CELLS AS YOU KNOW, THEY ARE IN THE CEREBELLUM AND THEY HAVE VERY FEW INPUT SO, CAN YOU UP AND SEE THEM, THEY CAN INPUT THROUGH THE CLAW LIKE STRUCTURES, YOU CAN SEE THREE HERE AND THEN AN AXON. THEN THREE OR FOUR SYNAPSES, SO YOU KNOW YOU CAN MODEL, THE ENTIRE INPUT TO THEM, SO WHAT WE DID WAS TO BUILD A SIMPLE RIGHT NOW, JUST A SIMPLE INTEGRATE AND FIRE A PASSIVE MODEL CELL THROUGH THE THRESHOLD THINGS, BUT BUILD MODELS OF SYNAPSES, GIVEN THE IPSPs THAT HAD BEEN RECORDED. AND WE TRIED TO BUILD THE RESPONSES AND IN OTHER WORDS LET'S SAYOT PARTICULAR TRIAL TRR A PARTICULAR CELL, HERE'S ONE OF THE RECORDINGS, THE MEMBRANE POETIC TECTIAL LOOKS LIKE THIS AGAIN IN THE RESPONSE TO THE ELECTRIC ORGAN DISCHARGE. THIS IS THE ROYAL WE, THIS IS REALLY THE THEORY STUDENT, THE PICTURES I SHOWED YOU IN A WHILE, PACT PATRICK GREG AND AN WAS TO SORT THROUGH, ALL OF THE PRESYNAPTIC RECORDINGS THAT NATE HAD AND THEY WERE--THEY WERE QUITE A FEW OF THESE PARTLY BECAUSE THEY WERE EXTRA CELLULAR, SO IT'S--IT WAS THESE SOME MORE EXTRA CELLULAR, SO THEY LOOKED--NOW THESE ARE NOT VOLTAGES, THESE ARE FIRING RATES SO THEY FOUND A SAMPLE FROM THE MANY, MANY RECORDINGS OF THREE FIRING RATES THAT WHEN THEY GENERATED SPIKING FROM THESE RATES RATES AND INTEGRATED THEM, THEY COULD GET THIS RESPONSE AND I'LL SHOW YOU IN FACT, WHAT YOU GET FROM THE MODEL. IT'S REALLY QUITE A REMARKABLY GOOD FIT GIVEN THAT THESE RECORDIGS AND THESE WERE NOT MADE THE SAME TIME OR IN THE SAME ANIMAL. BUT JUST FROM THE SAME CIRCUIT. SO, WE WERE ABLE TO LOOK LIKE LIKELY CANDIDATES FOR HAVING PROED THE RESPONSE AND WE COULD DO THAT FOR ALL THE CELLS SO HERE'S A REPRESENTATIVE SAMPLE. SO THESE ARE DIFFERENT MG-- GRANULE CELL RESPONSES--SORRY. HERE YOU CAN SEE THE EPSPs FROM SPIKES, WE FOUND THE SET OF INPUTS WERE APPROPRIATE, MOST APPROPRIATE USING THE SORTS OF A SPARSE FITTING PROGRAM AND THEN--WHOOPS WE FIT SO NOW YOU SEEM GREEN TO FIT. THE FITS ARE TRULY UNBELIEVABLE, LIKE HERE, YOU SEE WE'RE FITTING EVERY LITTLE BUMP HERE, AGAIN DESPITE THE FACT THAT THAT WERE RECORDED AT THE SAME TIME OR IN THE SAME ANIMAL. SO WE REALLY GOT QUITE GOOD FITS. ALL RIGHT, SO WHAT. SO WE FIT THE DATA. WHAT WE EXTRACTED FROM THESE FITS IS THE PROBABILITY THAT A GRANULE CELL WOULD GET THESE DIFFERENT CLASSES OF INPUTS. IF YOU REMEMBER I DIVIDED THE INPUT INTO'S EARLY AND MEDIUM. THOSE ARE WERE THE MULTIFIBERS, LATE AND PAUSE, THOSE WERE THE UBCs, UNIPOLAR BRUSH CELLS AND SO, WE WERE ABLE TO GET A HISTOGRAM THAT LOOKS LIKE THIS. THIS IS THE PROBABILITY OF A GRANULE CELL RECEIVING EARLY INPUT. CAN YOU SEE THAT'S BIG BECAUSE THERE WERE A LOT OF EARLY INPUTS, MEDIAN LATE PAUSE AND WE ADDED A TON OF COMPONENTS WHICH WOULD BE AN INTRACELLULAR PIECE BUT WE NEEDED THAT. SO THEN WE MADE A BOLD ASSUMPTION WHICH IS--THIS IS ALL YOU NEED TO KNOW TO MAKE GRANULE CELLS. IN OTHER WORDS, FOR US TO MAKE A GRANULE CELL, BUT WE TAKE THREE OR FOUR SYNAPSES AND WE JUST DRAW ONE FROM EACH GROUP, AND WE TAKE THEM FROM THE DATA, ONE FROM EACH GROUP, JUST COMPLETELY RANDOMLY, UNTIL WE FILLED THE INPUT, AND WE BUILT 20,000 GRANULE CELLS NAWAY. NOW HOW DO YOU KNOW THAT THIS IS GOING TO WORK. WELL WHAT WE DID WAS WE TOOK OUR 20,000 GRANULE CELLS AND SAMPLED 170 OF THEM RANDOMLY WHICH IS PRESUMABLY WHAT NATE WAS DOING WHEN HE STUCK HIS ELECTRODE INTO THIS TISSUE. AND THEN WE SORTED THEM EXACTLY THE SAME WAY THAT NATE HAS SORTED THEM AND WE GOT THERE. SO THIS IS OUR SAMPLE SET OF CELLS AND THIS IS NATE'S EXPERIMENTAL CELLS AND WHAT I WANT TO STRESS IS THESE CELLS WERE NOT FIT TO BE CELLS, THIS IS ONE RANDOM SAMPLING FROM THE MODEL. THIS IS ONE RANDOM SAMPLING FROM THE DATA, SO I WOULD SAY THIS LENDS CREDENCE TO THE IDEA THAT THESE GRANULE CELLS REALLY DO SAMPLE RANDOMLY WITH THOSE INTRUSION, SO THIS WORKED EXTREMELY WELL. OKAY, SO NOW WE FILLED IN THE MISSING PIECE. WE KNOW IT'S AN EXCEPTIONAL CIRCUIT HERE WHERE WE KNOW EVERYTHING WE NEED TO KNOW FOR THIS PARTICULAR EXERCISE BECAUSE NOW WE KNOW WHAT'S HAPPENING HERE, AND SO WE CAN DO A SIMPLE EXPERIMENT WHICH IS TO START WITH RANDOM SYNAPSES HERE, IN WHICH CASE WE WON'T HAVE A NEGATIVE IMAGE AND WE SHOULD SEE THE RINGING ON THE RESPONSE, SO THIS GREEN LINE WHICH IS WHAT I'M GOING TO SHOW YOU WILL BE WIGGLING LIKE THAT AND OVERTIME WHAT WE'LL DO IN THE MODEL IS TURN ON THAT LEARNING ROLE, LET THE SYNAPSES ADJUST AND IF EVERYTHING WORKS, WE SHOULD-OF WE SHOULD BUILD THE NEGATIVE IMAGE AND CONSLE OUT THE ORIGINAL SIGNAL SO WHAT WE HAVE OUR MODEL 20,000 CELLS FOR INPUT AND WE HAVE A MODEL MG CELL, WE KNOW THIS INPUT SO WE JUST FEED THIS IN, FROM RECORDINGS AND WE KNOW THE LEARNING RULE. AND OF COURSE, THAT WAS A QUESTION MARK, OF COURSE YOU KNOW IT WORKS. OKAY, SO HERE IS THE INITIAL POSITIVE IMAGE, THE OUTPUT WHEN IT'S NOT PROPERLY ADJUSTED AND TRIALS HERE MEAN PULSES OF THE ELECTRIC ORGAN SO THAT THERE'S NO ANIMALS HERE BUT THE MODCELL PULSING THE ORGAN OVER THE PERIOD OF ABOUT A THOUSAND PULSES WHICH IS ABOUT FIVE MINUTES OF TIME, YOU SEE THIS BEAUTIFUL CANCELLATION EVEN NOT VERY NOISY OF THE SYSTEM. AND YOU KNOW ABOUT FIVE MINUTE SYSTEM THE RIGHT TIME. I WILL SHOW YOU THIS IN A SECOND BUT CAN YOU MARALIES THE ELECTRIC ORGAN OF THE FISH SO IT CAN'T PRODUCE ANY ELECTRIC FIELD BUT IT STILL KEEPS PULSING, AND KEEPING TRYING TO PRODUCE AN ELECTRIC FIELD BUT YOU PARALYZED IT AND YOU CAN PRODUCE YOUR OWN ELECTRIC FIELD. AND ASK HOW LONG DOES IT TAKE THE FISH TO MAKE AN ADJUSTMENT TO SOME WEIRD ELECTRIC FIELD THAT YOU PUT IN WHICH DOES NOT HAVE TO BE THE SAME AS THE NATURAL FIELD AND IT'S OVER THIS AMOUNT OF TIME. SO WE TOOK ADVANTAGE OF THAT STRICK TO DO THE FOLLOWING EXPERIMENT. SO THIS IS THE WHOLE SYSTEM AS I SHOWED YOU MPLET SO AGAIN IF YOU PARALYZE THE ORGAN, YOU PARALYZE THE FISH, SO THERE IS NO FIELD, NO REACTION HERE AND YOU CAN BUILD WITH ELECTRODES IN THE WATER, CAN YOU BUILD A FIELD. BUT IN THIS CASE, WHAT NATE DID WAS INSTEAD DIRECTLY EXCITE THIS CELL WITH THE PULSE. SO THE SENSORY INPUT IS READING OR REPLACED BY A DIRECT CURRENT INJECTION. SO AGAIN IF YOU PULSE LIKE THIS OF COURSE YOU CAN FORGET A PULSE IN THE OUTPUT AND YOU KNOW THAT WILL COME OUT IN THE OUTPUT. YOU WILL GET A PULSE IN THE OUTPUT. NOW IF YOU INSTEAD THOUGH, LOCK THAT PULSE TO THE ELECTRIC ORGAN DISCHARGE, SO, THE FISH IS TRYING TO DRIVE ITS ELECTRIC ORGAN, CAN YOU PICK UP THE NERVE SIGNAL OF THAT DISCHARGE SO IF YOU LOCK THIS TO IT WITH A CERTAIN DELAY, THAT'S WHAT THESE EXPERIMENTS ARE SO MAYBE 50 MILLI SECONDS AFTER THE ELECTRIC DSCHARGE, YOU KEEP PULSING THIS THING. WHAT WILL HAPPEN IS THIS SYSTEM WILL MAKE A NEGATIVE IMAGE BECAUSE THAT'S WHAT IT'S THERE FORE, ANYTHING PREDICTED BY THE ELECTRIC ORGAN DISCHARGE BECOMES PREDICTED BY THIS SYSTEM. AND SO OVER TIME THE SYSTEM WILL MAKE A NEGATIVE IMAGE. IT CAN MAKE A PERFECT SQUARE PULSE BUT IT'LL MAKE SOME APPROXIMATION TO IT AND THAT WILL, YOU KNOW ALMOST PERFECTLY CANCEL OUT THE OUTPUT AND THEN CAN YOU SEE THE NEGATIVE IMAGE BY ESSENTIALLY TRICKING THIS SYSTEM, LEAVING OUT THE PULSE AND THEN, YOU'LL ONLY SEE THE NEGATIVE AIMAGE. SO THAT'S HOW THESE EXPERIMENTS ARE DONE. SO FIRST LET ME SHOW YOU THIS EXPERIMENTS DONE IN THE MODEL. SO HERE WE SIMULATED A PULSE OF CURRENT AT THIS PARTICULAR TIME. SO IT LOOKS LIKE ABOUT 15-14 MILLISECONDS AFTER THE DISCHARGE AND SURE ENOUGH OUR SYSTEM BUILT THIS NICE NEGATIVE IMAGE ANDADS THEY SAY IF YOU DO THAT IN THE FISH, THE SAME THING HAPPENS THE INTERESTING THING OCCURRED WHEN WE DID IT LATER SO HERE'S THE PULSE AT 150 MILLISECONDS IT MADE THE IMAGE HERE BUT IT MADE THIS EXTRA BUMP HERE, THAT IS THE PROPERTY OF THE NATURE OF THE BASIS THAT WE FOUNDOT GRANULE CELLS, A FUNNY BASIS. IT'S HEAVYOT FRONT END AND YOU CAN SORT OF THINK ABOUT THE PRIOR SYSTEM, NAMELY IF THE SYSTEM SEES AN ELECTRIC FIELD HERE THAT NEEDS CANCELING, THIS SORT OF AN ASSUMPTION THAT THAT MUST HAVE BEEN PART OF AN OSCILLATION, BECAUSE THE ELECTRIC ORGAN DISCHARGE, YOU KNOW HAPPENED HERE, IT'S UNLIKELY THAT NOTHING HAPPENED AND THEN SUDDENLY YOU GET AN ELECTRIC FIELD AND IN FACT, THAT'S WHAT THE FISH DOES, TOO. THERE IS THIS ANATIONAL LIBRARY OF MEDICINE LOWS THING, THIS HAS BEEN OBSERVED BY CURTIS BELL WHO ALWAYS WONDERED WHAT IT WAS AND THOUGHT THERE MAY BE A PLASTICITY THAT MISSED BUT IN FACT IT'S THE CONSEQUENCE OF THE NATURE OF THE BASIS THAT THE FISH PRODUCED. IN OTHER WORDS IT'S A FUNCTION OF THE FACT THAT THERE'S SO MUCHOT FRONT END AND NOT A WHOLE LOT AT THE BACK END. OKAY. SO THAT'S OUR UNDERSTANDING THEN ABOUT THE CANCELLATION OF THE BUILDING OF THE NEGATIVE IMAGE CANCELLATION. SO LET ME COME BACK TO WHERE I STARTED AND THEN R THEN I'LL MOVE ON TO THE SLIDE. I SHOWED YOU THIS PICTURE IN WHICH, OF COURSE THIS IS NOT DONE BY THE FISHES BUT COMPUTER, BY HAND SUBTRACTING OUT THE EFFECTS OF THE DISCHARGE. SO IN OTHER WORDS IT'S A COMPUTER CONSTRUCTIVE NEGATIVE IMAGE AND I FOREVER SHOWED YOU THE R. O. C. RESULTS FOR THIS. SO I SHOWED YOU AT BET BEGINNING IF YOU DON'T DO ANY SUBTRACTION, YOU'RE A LITTLE ABOVE CHANCE, MAYBE 53-54%. THIS GETS YOU UP TO MAYBE 75%. SUBTRACTING THE NEGATIVE IMAGE BUT I ALSO SHOWED YOU THE OUTPUT UP THERE. SO OUR CURRENT BELIEF, IF THIS DOESN'T GET YOU ALL THE WAY AS BEING AS GOOD AS THE FISH AND WHAT WE'RE CURRENTLY WORKING ON, AS FATTER PART OF THE COLLABORATION IS REALLY TRYING TO FIGURE OUT WHAT GETS YOU THE REST OF THE WAY. ONE THING IS SUMMATION, THIS IS BASED ON AFFERENT BUT AN OUTPULL CELL RECEIVES INPUT FROM THE AFFERENTS, SO THERE'S AVERAGING THAT SHOULD IMPROVE THE SITUATION, AND THERE'S ALSO PROBABLY SOME SORT OF CENTER SURROUND SPACIAL SUBTRACTION GOING ON. SO EVERYTHING I TALKED ABOUT, IS A TEMPORAL SUBTRACTION, BUT, YOU COULD ALSO USE THE FACT THAT THE SELF-INDUCED ELECTRIC FIELD IS MORE UNIFORM THAN A TYPICAL ELECTRIC FIELD FROM A BUG, SO WE WILL LOOK AT THAT AND THERE ARE A COUPLE OF OTHER BEAUTIFUL THINGS I WILL MENTION ALTHOUGH I'M NOT INVOLVED THIS THEM AND ONE IS THAT WE CONSIDER A STATIONARY FISH, BUT OF COURSE THEY SWIM AND THE TAIL GOES BACK AND FORTH, AS THE TAIL SWINGS HERE'S THE ELECTRIC ORGAN, THE FIELD WILL GET STRONGER AND WEAKER DEPENDING ON WHICH SIDE THE TAIL'S ON AND ANN KENNED SCHETIM HAVE DONE EXPERIMENTS HAVE DONE A BEAUTIFUL JOB EXPLAINING THAT. SO THEY HAVE A MODEL LIKE THE MODEL I'VE DONE THAT CAN PREDICT THE ELECTRIC FIELD EVEN WHEN THE TAIL IS MOVING BACK AND FORTH. OF COURSE THAT MEANS THAT THE SYSTEM NOT ONLY GETS COROLLARY DISCHARGES OF THE ELECTRIC ORGAN BUT COROLLARY DISCHARGE OF THE MUSCLES THAT BEND THE BODY. AND FINALLY THERE'S ANOTHER STUDENT, CONNOR DESMSY IS WORKING ON THE FACT THAT AS THE FISH MOVE IN ON A PREY, THEY ACTUALLY INCREASE THE E. O. D. FREQUENCY SO THIS SYSTEM HAS TO WORK AT A LOT OF HIGHER FREQUENCIES THAN WHAT I'VE SHOWN HERE. SO ANYWAY, THAT'S JUST THE OING RESEARCH THAT HOPEFULLY WILL MAKE PROGRESS, AND GET THE R. O. C. CURVE UP ALL THE WAY OR AT LEAST CLOSE TO THE TOP HERE. BY THE TIME WE'RE DONE. OKAY. I WILL--I WILL SWITCH CREATURES AND TALK ABOUT THE FLY, A VERY SIMILAR SYSTEM IN THE FLY. DIFFERENT SET OF COLLABS, THIS IS WORK WITH RICHARD AND MEMBERS OF HIS LAB AND ALSO WITH A CREW AT GENELIA AND SOVEY AND VANESSA ARE ALL POST DOCS AND THIS IS A TYPICAL DAY AT THE LAB, WE TRY TO LOOK GOOD WHEN WE DO THIS WORK. OKAY. SO THIS IS THE OLFACTORY SYSTEM OF THE FLY, IT'S REMARKABLY LIKE YOUR OLFACTORY SYSTEM, MIND BLOWINGLY LIKE YOUR OLFACTORY SYSTEM, CONSIDERING THEY'RE NOT EVOLUTIONARILY CONNECTED, IT'S APPARENTLY THE RIGHT WAY TO BUILD AN OLFACTORY SYSTEM. SO INSTEAD OF A NOTICES, HAVE YOU FACT RESILIENCE SEPTORSORS ON THE AN FENNA IN A FLY, THOSE LIKE IN MAMMALS EXPRESS A SINGLE RECEPTOR TYPE. ALL OF THE ODOR RECEPTORS, WITH THE SAME TIME CONVERGE ON A GLUMARIUS, SO THESE ARE THE GLAMULARY LYE AND THESE ARE RESPONDING TO RECEPTOR TYPES WHERE WE HAVE MORE LIKE 400 RECEPTOR TYPES. FROM THERE THE SIGNAL GOES UP TO A STRUCTURE I'LL TALK ABOUT A LOT WHICH IS THE MUSHROOM BODY. IF YOU WANT THAT'S KIND OF LIKE PERIFORM CORTEX BUT IT'S NOT. IT'S STRUCTURED LIKE THE CEREBELLUM, SO IT'S--IN KEEPING WITH THE THEME OF THE DAY, IT'S A HIGHER LEVEL RECEPTOR, AND IT ALSO GOES TO A STRUCTSURE CALLED THE LATERAL HORN THAT I'LL MENTION, SO THAT'S THE GENERAL PATHWAY. HERE'S NICE PICTURES OF TSO HERE ARE THE RECEPTORS TO THE ANTENNA. YOU SHOULD REMEMBER O. R. N. BECAUSE THAT WILL APPEAR IN MY SLIDES, OLFACTORY RECEPTOR NEURON. THAT'S THE FRONT END OF THIS SYSTEM. HERE THE RECEPTORS EXPRESSING A CERTAIN TYPE OF RECEPTOR MOLECULE, HAVE BEEN LABELED OBVIOUSLY, SO THERE THEY ARE. I'VE GIVEN YOU THE MOUSE OR THE HUMAN ANALOGUE HERE. HERE CAN YOU SEE THIS BEAUTIFUL CONVERGENCE, SO THE AXONS FROM ALL OF THESE GUYS HERE ARE CONVERGING ON A SINGLE GLUE MARIA, THAT'S THE BEAUTIFUL ONE THAT TAKES PLACE IN MAMMALS. HERE'S THE AN LOG OF A MIGHTERAL CELL, A PROJECTION NEURON. I THINK I HAVE THE NAME--NO, THERE'S THE ANTENNA LOBE, SO P. N., YOU WILL SEE THAT ON MY SLIDE, THAT'S THIS PROJECTION STEP. AND IT GOES UP TO THE TWO REGIONS, THE MUSH BODY, THE LATERAL HORN, I WILL TALK ABOUT IT, I WILL TALK ABOUT LEARNED VOTER RESPONSES SO, THE--AND OF COURSE, HERE'S SORT OF THE ANALOGUE, BUT THE PATH I WILL TALK ABOUT IS FROM HERE, OLFACTORY RECEPTOR TO THE GLAMULARIS, UP TO THE MUSHROOM BODY AND UP TO THE MUSHROOM BODY UP TO THE REST OF THE BRAIN. OKAY, HERE'S JUST ANOTHER PICTURE OF THIS STRUCTURE, YOU SHOULD KEEP IN MIND THIS SHORT OF L-SHAPED. THIS PURPLE AND THEN THIS L-SHAPED THING, THAT'S THE MUSHROOM BODY, SO IT'S AN L-SHAPED STRUCTURE THAT'S ALL YOU HAVE TO REMEMBER FROM THIS AND THESE ARE THE END FRONTAL LOBE OUTS FRONT END IS NOT SHOWN HERE AND THIS IS ONE OF THE PROJECTION NODES. AND THERE THEY ARE. YOU KNOW JUST BEAUTIFUL PICTURE SO HERE IS A PROJECTION NEURON COMING FROM WHO KNOWS WHERE, UP CONNECTING IN THE CALYX AND THESE ARE THE CELLS OF THE MUSHROOM BODY, BIFURCATING INTO THE L-SHAPED STRUCTURE AND THE POINT OF THE CELLS DON'T LEAVE THIS STRUCTURE AND WE WILL TALK ABOUT OUTPUT NEURONS THAT LEAD TO THIS STRUCTURE AND HERE'S MORE PICTURES, I GET CARRIEDED AWAY FOR A THEORIST BUT THEY'RE PRETTY. HERE'S THE KENYON CELLS, AND THERE'S AN INTERESTING POINT THAT'S WORTH MAKING EVEN THOI WILL RUN OUT OF TIME. MOST OF THE CELLS IN THE FLY BRAIN ARE LIKE THESE GUYS IN THAT THERE ARE CELL TYPES WITH VERY FEW MEMBERS SO OF 150 CELLS MAYBE IN 70 TYPES THERE ARE A FEW TYPES PER MEMBER OF A CELL TYPE, FEW CELLS PER MORE AND THEY'RE VERY STEREOTYPED I'M NOT SHOWING THAT YOU HERE, THEY'RE MUCH A LIKE THATTA TRUE IN THE LEVEL TWO, BUT THESE ARE CELLS THERE ARE HUNDREDS OF CELLS PER TYPE. SO HERE'S THE THEORIST DIAGRAM, GET RID OF ALL THE BIOLOGICAL BEAUTY AND GET DOWN TO BASICS. SO THIS IS A REPRESTATION OF THE NEURONS, THE COLOR WILL REPRESENT WHICH TYPE OF RECEPTOR THEY'RE EXPRESSING. THIS IS THE IDEA OF THE CONVERGENCE WHERE ALL THE RED GUYS CONVERGE ON TO A RED GLUE MAR LIAISONS, ET CETERA LIKE THAT SO THIS IS STRUCTURED WIRE. BEAUTIFULLY STRUCTURED, I THINK YOU ALL KNOW THIS STORY IT'S ALSO TRUE IN THE MAMMAL AND IT'S PRETTY CLEAR THAT THAT STRUCTURE WHICH EXISTS IN PARALLEL OF EVOLUTION BETWEEN INSECTS AND MAMMALS IS DESIGNED FOR NOISE REDUCTION. YOU WANT TO TAKE THE SIGNALS FROM THE RED GUYS, BRING THEM TOGETHER AND AVERAGE THEM TO GET RID OF NOISE. IN THE ANTENNA LOBE, THOSE SYSTEMS ARE A NORMALIZATION SYSTEM, IT'S TRYING TO STAY A POWERFUL LOADER AND A WEAK ODOR SHOULD HAVE KIND OF EQUAL RIGHT TO CONTROL THE BEHAVIOR OF THE FLY. AND SO, THERE'S SOME NORMALIZATION THAT GOES ON HERE TO TRY TO GIVE THEM A MORE EQUAL FOOTING. I WON'T TALK ABOUT THAT IT'S JUST TO INTRODUCE. SO THEN WHAT HAPPENED AS I MENTIONED THERE ARE 50 TYPES OF THAT GETS SEABED UP BY THE PROJECTION NEURONS TO THESE CELLS SO YOU SHOULD GET THE THAT THIS IS SOME SORT OF EXPANSION OF THE REPRESENTATION IN TERMS OF CELL NUMBERS INTO A HIGH DIMENSIONAL REPRESENTATION FROM A RELATIVELY LOW DIMENSIONAL REPRESENTATION AND THE FIRST THING THAT WE SET OUT TO DO WITH MEMBERS DOWN HERE, WITH RICHARD'S LAB IS TO TRY TO SEE WHAT'S THE NATURE OF THIS CONNECTIVITY. DOES IT SUPPORT THE IDEA THAT YOU REALLY ARE TRYING TO EXPAND THE DIMENSIONALITY OF THE OLYMPIC IN FACTORY REPRESENTITION, SO THAT WAS DONE IN THE FOLLOWING WAY, EXPERIMENTALLY I CAN TELL WHAT YOU I DID THESE ARE A WHOLE BUNCH OF DENNIAN CELLS YOU SEE DIFFERENTLY THEY ARE EXPRESSING A GFP, AND IN THIS CELL THIS IS HAS BEEN ACTIVATED AND THAT'S WHY IT STABBEDS OUT LIKER THIS. YOU CAN SEE THE CELL BODY, REACHING INTO THE KACCT LICK TO GET THE INPUT AND SENDING THEAXON DOWN HERE DOWN BELOW THE FLOOR WOULD BE THAT L-SHAPED WHAT ARE CALLED THE LOBES OF THE MUSHROOM BODY. THESE STRUCTURES HERE ARE CLAWS, THEY SHOULD BE FAMILIES FROM THE GRANULE CELLS, THEY'RE CLAW LIKE STRUCTURES THAT REACH OUT FROM THE KENYAN CELLS, GRAB ON TO PROJECTION NEURON TERMINAL AND THAT'S HOW THEY GET THEIR PIN UT. --THAT ALLOWED THE TRICK THAT WAS USE INDEED THESE INSTRECTED INTO THE CONNECTION AND THAT DIE IS TAKEN UP BY ONE AND ONLY ONE PROJECTION NEURON AND HERE YOU CAN SEE WHERE IT GOT TAKEN UP HERE'S THE AXON OF THE PROJECTOR NEURON YOU CAN SEE OTHER SITES WHERE IT'S MAKING ADDITIONAL SITES WITH OTHER CELLS BUT THE TRICK HERE IS THAT YOU CAN FOLLOW THIS AXON BACK AND CAN YOU FIGURE OUT WHICH GLUE MARRUOUS IS IT IS, SO THESE ARE MAPPED OUT SO IF YOU KNOW THE AN TELA LOBE, YOU KNOW WHICH GLAMARRIAL THIS IS, WHICH RECEPTOR TYPE THIS IS, SO BY DO THANKSGIVING REPEATEDLY YOU CAN FIND THE CONNECTIONS TO THIS GUY; YOU CAN FIND WHICH GLAMAR LIEU LIE ARE FEEDING IT AND RESULT OF A HUGE AMOUNT OF CORRELATED WORK IS A MATE RICK LIKE THIS SO THIS IS THE CONNECTION, THE RED ARE SINGLE CONNECTIONS, THE ORANGE ARE DOUBLE CONNECTIONS FOR 100 DENNIAN CELLS GOING DOWN THIS WAY AND ALL THE DIFFERENT GLAMAR LIE THAT ARE PROVIDED INPUT. SO YOU GOT A NICE CONNECTION MATRIX. SO MY JOB, WHICH I WILL TELL YOU THE RESULTS MY JOB IS TO FIGURE OUT WHAT'S THE STRUCTURALLY ARE OF THIS MATRIX. NOW YOU MAY LIKE AT IT AND SAY IT DOESN'T LIKE IT HAS ANY STRUCTURE, BUT THE PROBLEM IS THAT THIS MATRIX S&P IN ARBITRARY ORDER THIS, IS ALPHABETICAL WHICH IS NOT THE KEY TO THE STRUCTURALLY ARE OF THE MUSHROOM BODY AND THIS IS JUST CHRONOLOGICAL, SO, THE KEY IS, YOU KNOW YOU HAVE TO ASK YOURSELF, IS THERE ANY WAY OF REARRANGING THE ROWS AND COLUMNS OF THIS MATRIX SO THAT STRUCTURE APPEARS. SO THE FIRST THING CAN YOU DO IS THUMB DOWN THE COLUMNS TO SEE HOW OFTEN DIFFERENT GLAMUE LIE PROVIDE INPUT TO A KENYON CELL AND YOU DON'T GET A UNIFORM DISTRIBUTION, YOU GOT A DISTRIBUTION LIKE THAT BUT THE REAL QUESTION THAT I WAS RESPONSIBLE FOR IS THERE ANYTHING ELSE? IS THERE--IS THIS THE WHOLE STORY. AND THE ANSWER, I WON'T GIVE YOU THE DETAILS, THE ANSWER IS NO. WE COULD NOT FIND ANY EXTRA STRUCTURE, IN OTHER WORDS THAT MATRIX IS COMPLETELY CONSISTENT WITH JUST RANDOM DRAWS FROM THAT PROBABILITY DISTRIBUTION AND THERE ARE OTHER PAPERS, AND REGS THAT PROVIDE SUPPORT FOR THIS AS WELL. SO, I HOPE YOU SEE THE ANALOGY HERE. THIS CASE IN ANATOMICAL ARGUMENT THAT THE FIRST ORDER OF STATISTIC IS ALL YOU NEED TO WIRE UP KENYON CELLS AND IN THE OTHER CASE IT'S A ELECTROPHYSIOLOGICAL CELLS WHEN WE MADE CELLS OF THE DISTRIBUTION OF THIS INTRUSION IS ALL YOU HAVE TO NOTE, WE SEEM SOPHISTICATED SUCCEED, SO I WOULD ARGUE IN BOTH OF THESE CASES THERE, 'S EVIDENCE FOR LACK OF STRUCTURE, YOU'RE JUST TRYING TO GET THE WIRES THERE AND THESE CLAWS GRAB ON TO THE FIRST THING THAT THEY FIND OUT THERE IN THE POOL. I GIVE YOU ANOTHER ARGUMENT SO WE ARGUE THEN THAT IN BOTH OF THESE CASES THE KEY IS TO BUILD AS RICH A REPRESENTATION AS POSSIBLE IN ORDER FOR THE SYSTEM TO DRAW THE IMPORTANT INFORMATION FROM IT, SO IN THIS CASE, YOU WANTED KIND OF A RICH TEMPORAL STRUCTURE AND YOU DID IT BY MIXING UP THESE INPUTS. IN THIS CASE, YOU WANT A HIGH DIMENSIONAL REPRESENTATION OF VOTERS AND DO YOU IT BY MIXING THESE MAXIMALLY CHRKS IS RANDOM AS GOOD AS ANYTHING ELSE. THERE'S ANOTHER NICE ARGUMENT YOU CAN DO TO SUGGEST THAT THE KENYON CELLS ARE BEING AS PROJECTILE AS POSSIBLE. SO SUPPOSE THEY RECEIVED ONE TYPE OF PROJECTION NEURON, SO THERE ARE 2000 GRANULE CELLS SO IF YOU DOES THAT YOU GET TONS OF MATCHING PAIRS, YOU GET MANY NEURONS THAT RECEIVED, YOU KNOW ONLY INPUT FROM THE FIRST GLAMULARLY AND THAT'S WHAT YOU'RE SEEING HERE, SO THIS IS JUST COMPUTER CALCULATION, SO IF THE--IF THE KENYON CELLS WERE MADE FROM A SINGLE INPUT, YOU WOULD GET, YOU KNOW MANY THOUSANDS OF MATCHING CELLS AND THAT NUMBER GOES DOWN IF YOU INCREASE THE NUMBER OF KENYON CELL INPUT BECAUSE THERE'S MORE COMBINATIONS AND THAT NUMBER GOES THROUGH ONE, JUST BETWEEN SIX AND SEVEN, AND THAT'S EXACTLY HOW MANY INPUTS KENYON CELLS GET, RIGHT BETWEEN SIX AND SEVEN SO, IT LOOKS LIKE THIS SYSTEM IS SET UP TO MAKE THUR THAT ALL THE KENYON CELLS ARE DIFFERENT TO MAKE IF AS HETEROGENEOUS AS POSSIBLE BUT NOT WASTE ADDITIONAL SYNAPSES WHEN YOU DON'T NEED THEM FOR THAT PURPOSE. SO THERE'S THIS VERY DRAMATIC CHANGE, YOU KNOW THE FIRST SYNAPTIC CONNECTION, IN THE OLYMPIC FACTUALLY SYSTEM IS HOW IT'S STRUCTURED. AND THE NEXT ONE SEEMS COMPLETELY UNSTRUCTURED, AS IF WE THROUGH EVERYTHING AWAY. SO WHAT UPON HAPPENS AT THE NEXT STAGE. THE KENYON CELLS DON'T LEAVE THE MUSHROOM BODY SO YOU HAVE TO GET THE SIGNAL OUT SOMEHO, AND THAT COMES THROUGH WHAT WE CALL MUSH AUTOMATIC BODY OUTPUT NEURONS. RIGHT SO THE KENYONS MAKE CONNECTIONS TO THESE GUYS AND THIS CARRIES THE SIGNAL HOW THE AND THEY FORM A REMARKABLY LOW DIMENSIONAL REPRESENTATION BUT THANKFULLY IF YOU CAN'T CHANNEL SOMETHING LIKE 16 DIFFERENT CHANNELS. SO YOU START WITH 15 CHANNELS, YOU EXPAND IT WAY UP AND YOU COLLAPSE IT DOWN AND THIS TO ME WAS EXCITING BECAUSE IT TELLS ME THAT THE OUTPUT NEURONS ARE NOT ANOTHER OLFACTORY REPRESENTATION AND YOU KNOW IN SENSORY SYSTEMS WE OFTEN SEE REREPRESENTATION, SAY THE VISUAL SYSTEM, REREPRESENTATION OF THE DATA IN MORE AND MORE COMPLICATED FORMS AND IT'S INTERESTING AND BEAUTIFUL BUT YOU'VE GONE WHERE THE BUCK STOPS, WHERE DOES IT MAKE THAT BEHAVIORIAL TO MAKE A BEHAVIORIAL SYSTEM, AND IT'S VERY HARD TO FIND THAT POINT, HERE, THAT OCCURS RIGHT HERE, RIGHT AT THIS SYNAPSE, BECAUSE THIS IS TOO DIMENSIONAL TO PROVIDE A COMPLETE REPRESENTATION OF THE OLFACTORY SPACE IBT--INTEGRATE STED WE PROVIDE THIS AS MANY OTHERS AS PROVIDING THE BEGINNINGS OF A BEHAVIORIAL BIAS. YOU TAKE THE LESSON LEARNED FROM THE ODOR AND START TO BIAS YOUR BEHAVIOR IN ONE WAY OR THE OTHER. SO SORT OF AN EXCITING MOMENT AND ELUCIDATING THIS AS PART OF A PAPER THAT WE COLLABORATED ON BUT THE BULK OF THE WORK--I'LL DESCRIBE IT BRIEFLY IN NOW IN ORDER TO GET THIS OUT OF THIS SYSTEM, YOU BETTER LEARN, SO THIS IS RANDOM, IT SHOULD BE DIFFERENT IN DIFFERENT FLIES, YOU GET WHAT YOU GET WHEN YOU'RE BORN AND SO THERE'S NO WAY THESE WILL FIRE APPROPRIATELY RIGHT OHE BAT SO YOU HAVE TO LEARN ON THESE SYNAPSES AND THERE'S A TAUPA MEAN SYSTEM THAT GUIDES THAT LEARN TAG I'LL TALK ABOUT. POLITICAL ME START OFF BY CONSIDERING A SINGLE OUTPUT NEURON AND ASKING WHETHER IT WAS WORTH IT TO DO THIS RANDOM EXPANSION. SOY WHY DIDN'T YOU STICK WITH THIS DIMENSION AND YOU HAD BEAUTIFULLY CONSTRUCTED IT AND GO STRAIGHT TO THE OUTPUT NEURONS. SO THIS A MODEL THAT ANN KENNEDY BUILT AND IT'S AN EXAMPLE OF TRYING TO DO A LITTLE LEARNING EXPERIENCE. THERE ARE 110 ODORS AND THE REASON FOR THAT IS THEY HAD THEM IN CARLSON RECORDED THE OLFACTORY RECEPTOR NEURON RESPONSES TO A FAMILY OF 110 ODORS, SO THE BEGINNING OF THIS MODEL IS REAL DATA WHICH WE THEN SEND THROUGH THE MODEL AND THIS IS AN EXPERIMENT IN WHICH EACH OF THESE ODORS HAS BEEN LEARNE. SO FOR EXAMPLE, YOU MIGHT IMAGINE THAT WE'RE ASSOCIATING THESE ODORS WITH SHOCK OR SOMKE THAT SO THE IDEA IS THAT WE TRAIN THE SYSTEM, THE OUTPUT NEURON OF THIS SYSTEM TO REACT TO EACH OF THESE ODORS, SO WHEN YOU LOOK AT THE COLUMN 100, IT'S REACTING TO ODOR 100. NOW WHAT YOU'RE SEE NOTHING THE NUMBER OF ERRORS IS HOW MANY ADDITIONAL ODORS DOES IT REACT TO. SO IT REACTS TO THE ONE WE TRAINED AT BECAUSE THAT'S THE WAY WE BUILT IT BUT IN ADDITION, FOR EXAMPLE, THIS GUY WAS TRAINED TO RESPOND TO ODOR 20 BUT IT RESPONDS TO 10 ADDITIONAL ODORS OUT OF THIS SET. SO IT GIVES 10 FALSE ALARMS AND ONE REAL ALARM. THIS IS THE CASE WHERE YOU POTENTIATE, I SHOULD MENTION THAT THERE'S A DEBATE ALTHOUGH MOST PEOPLE THINK THIS SYNAPSE IS DEPRESSED BY DOPAMINE, NOT POTENTIATED BUT IT'S NOT TOTALLY KNOWN SO THIS IS THE MODEL WHEN YOU POSITIONAL CLONING POE TENTIATE, WHEN YOU DEPRESS TDOES WELL, VERY FEW FALSE ALARM WHEN IS YOU HEAR IT. NOW HERE THE ANALOGOUS CURVE, DIRECTLY FROM P. N. N. INPUT AND YOU SEE HUGE NUMBERS OF ERRORS IN IN THING. SO IT DOES BENEFIT EVEN THOUGH IT'S JUST A RANDOM RE-REPRESENTATION IN THE MODEL YOU DO BENEFIT FROM IT. THERE'S AN OVERGENERALIZATION. ALL RIGHT, SO NOW LET ME GO BACK TO 16 CHANNELS, WHAT DO THEY CORRESPOND TO? IF YOU TAKE THE MUSHROOM IF YOU TAKE THIS STRICTURE AND IT WAS OBSERVED EARLIER, THEY DIVIDE INTO KIND OF FIVE ZONES EACH, SO THERE ARE ARE THREE KINDS OF LOBES AND EACH ONE DIVIDES INTO FIVE. SO THERE'S THE ALPHABETTA LOAD, ALPHA PRIME BETA LOAD THAT'S 15 OF THE 16 AND THEN PEDUNCAL IN THE BACK ARE THE 16, AND THESE ARE 16 CHANNELS AND YOU SEE HERE ARE THE L-SHAPED STRUCTURE SO HERE ARE THE KENYONS AND HERE THE SYNAPSES I TALKED ABOUT, HERE'S THE L-SHAPED STRUCTURE. THESE OUTPUT NEURONS, I'M SHOWING YOU REGION WHERE IS THE OUTPUT NEURONS GET THEIR PIN PUT ARE BAFFLELY LAID OUT GEOMETRICALLY IN THESE FIVE ZONES. SO THESE ARE THE 15 OUTPUT CHANNELS OF THIS SYSTEM. THESE GUYS ACTUALLY HERE'S A BEAUTIFUL PICTURE, OF--FROM MULTICULTOR FLIP OUT, OF THESE REGIONS, LOOKS PRETTY. OKAY. SO THESE ARE PROVIDING THE 16 CHANNELS OF OUTPUT FROM THIS SYSTEM. THEY'RE LAID OUT IN QUITE AN ORGANIZED WAY DEPENDING ON THE TRANSMITTER SO THESE ARE THE COLONERGIC GUYS, GRABBAERGIC GUYS AND GLUTEA METERGIC GUYS AND THEY'RE BEAUTIFULLY LAID OUT AS FAR AS THE DOPAMINE INPUT. SO LET ME GO TO THE DOPAMINE INPUT. I WILL FIND TIME TO FINISH UP HERE. SO WHAT ABOUT THE DOPAMINE MODULATION? THE BEAUTIFUL PART OF THE DOPAMINE MODULATION IS THAT MAYBE I'LLUOUSLY JUMP, AHEAD FOR A SECOND. IF EACH OF THESE GETS THEIR OWN DOPAMINE MODULATION, COMPLETELY ISOLATED, THERE'S NO OVERLAP, DEEPA MINE THAT GUY ANOTHER ONE THAT GUY, ANOTHER ONE THAT GUY AND LET ME JUST SHOW YOU THAT. HERE'S THIS ALIGNMENT, SO HERE IS THE ALIGNMENT OF THE DIFFERENT REGIONS AS DEFINED BY THE DENDRITES OF THE OUTPUT NEURONS AND HERE'S THE ALIGNMENT OF THE DOPAMINE NEURON MODULATION, SO CAN YOU SEE THAT THIS SIGNAL, WHICH IS--PROBABLY A LEARNING SIGNAL IS BEAUTIFULLY TARGETED TO EACH COMPARTMENT SO CAN YOU DIAL UP OR DOWN IN OUTPUT INDEPENDENTLY BASED ON EXPERIENCE FROM ALL THESE CHANNELS. NOW IN ADDITION, THERE'S A BEAUTIFUL LAY OUT IN TERMS OF THE KIND OF INPUT THAT DOPAMINE INPUT THAT THESE RECEIVE, SO THE DEEPA MINE INPUTS ARE CRUDELY DIVIDED INTO THOSE ASSOCIATED WITH PUNISHMENT AND THOSE ASSOCIATED WITH REWARD, AND YOU CAN SEE THEY'RE LAID OUT, YOU KNOW TOPPOGRAPHICALLY HERE ON THIS OUTPUT HERE, SO WHAT YOU SEE IS THAT FROM HIGHLY ORDERED OVER HERE, COMPLETELY DISORDERED OVER HERE AND THEN BACK TO VERY HIGHLY ORDERED AND THE THING THAT WILL LINK THESE OF COURSE, IS THE DOPAMINE INPUT TRAINING THESE SO THAT THE APPROPRIATE INFERENCE IS MADE. IN OTHER WORDS THIS IS A SYSTEM BY WHICH A REACTION IS TURNED INTO PREDICTION OF THE FUTURE. SO WHAT YOU CAN THINK OF IS THAT AN ODOR COMES ON, ACTIVATES, THESE PARALLEL FIBERS, AND THING YOU SHOULD NOTICE IS THIS IS ANOTHER PARALLEL FIBER SYSTEM, THAT INFORMATION GOES TO EVERY SINGLE OF THESE CHANNELS, IDENTICALLY, BUT DEPENDING ON HOW THE WORLD TREATS YOU, YOU SMELL THAT ODOR, YOU MIGHT MODULATE SYNAPSES IN THAT COMPARTMENT OR THAT COMPARTMENT. AND THESE ARE PRESUMABLY DOWN STREAM LINKED TO DIFFERENT BEHAVIORS SO CAN YOU ASSOCIATE THE FACT THAT YOU KNOW YOU SMELLED AN ODOR AND GOT A SHOCK BY ACTIVATING LET'S SAY THAT GUY. AND THEN THAT CAN BE LEADING TO A FLIGHT RESPONSE OR SOMETHING LIKE THAT, SO THAT THE NEXT TIME YOU SENSE THAT ODOR, YOU GET OUT OF THERE. NOW I'LL SHOW YOU A MODEL OF THAT BUT MAYBE I SHOULD--YEAH, I SHOULD MENTION ONE MORE THING, THOUGH, BEFORE I DO THAT. SO THESE OUTPUTS GO OFF TO OTHER STRUCTURES IN THE BRAIN THAT ARE PRESUMABLY GOING TO GENERATE THE BEHAVIORS BUT THEY ALSO FEEDBACK ON TO THE DOPAMINE SYSTEM. SO THIS SYSTEM, THIS SEEMS LIKE IT'S SET UP SO THAT LEARNING COULD SHUT DOWN THE PROMOTER OF LEARNING. SO IN OTHER WORDS IF THIS GUY'S GETTING A DOPAMINE SIGNAL SAYING, YOU SHOULD HAVE PREDICT THAD ODOR WAS GOING TO GIVE ME A SHOCK, ONCE IT'S LEARNED IT CAN SHUT DOWN THE DOPAMINE SYSTEM, TURN OFF ITS OWN LEARNING AND I MENTION THAT BECAUSE A MODEL I'M GOING TO BUILD, I WILL SHOW YOU, IS BUILT ON THAT ASSUMPTION. SO, LET ME RUN THE MODEL. IT'S A SIMPLE MODEL, IT GIVES YOU AN IDEA. SO THESE ARE JUST SUPPOSED TO REPRESENT THREE DIFFERENT ODORS, BEING PULSED TO THE ANIMAL. SO THE DIFFERENT COLORS REPRESENT DIFFERENT ODORS AND NOW, I'M JUST GOING TO IN THE MODEL BY ACTIVATING THIS, GENERATE DOPAMINE RESPONSES TO THESE ODORS AND I JUST MADE THIS UP. BUT WHAT CAN YOU SEE IS THAT THE IDEA IS THAT YOU GET THE ODOR FIRST AND THEN DURING THIS BLACK BAR, MAYBE YOU GET A SHOCK, OR YOU GET SOMETHING ASSOCIATED WITH THE ODORS AND IF YOU NOTICE WHATEVER IT IS, THAT I'VE BEEN DOING IN THIS MODEL HAS DIFFERENT VALENTS SO CAN YOU THINK OF THIS AS A REALLY BAD OUTCOME, THIS AS A MEDIUM OUTCOME AND THIS IS A NOT TOO BAD. LITTLE BAD BUT NOT TOO BAD. SO WE HAVE A GRADED RESPONSE, AND THIS IS WHAT WE WOULD HOPEFULLY WANT TO PREDICT, RIGHT? RATHER THAN WAIT AROUND UNTIL THE BAD THING HAPPENS. SO WHAT I'M GOING TO DO IS TURN ON LEARNING. WHICH IS A SYSTEM, VERY SIMILAR TO WHAT--WHAT WE STUDY INDEED THE ELECTRIC FISH THIS, IS WHERE YOU GET THE BENEFIT OF THE TRADE OFF, WE ALREADY KNOW HOW TO DO THIS, BECAUSE WE IT IN THE REACTIVE FISH MOLDLE AND WHAT YOU SAW WHILE I WASIACKING IS THAT IT DRIVES DOWN THE DOPAMINE RESPONSE AND GENERATES OUTPUT RESPONSE THAT COMPLETELY MIRRORED, IF YOU REMEMBER THE VALENTS. SO THE VALENTS IS NOW CARRIED BY THE OUTPUT NEURON AND NOW IT'S STARTS IMMEDIATELY UPON ODOR RESPONSE WHERE THESE LITTLE RESIDUAL DOPAMINES START AT THE BLACK BAR. SO YOU TURNED THIS REACTIVE REPRESENTATION OF VALENT INTO A PREDICTIVE REPRESENTATION OF VALENTS. THAT'S THE IDEA. OKAY, LET ME GO BACK TO MY TALK. SO LET ME WIND UP. THESE SYSTEMS ARE, AS I MENTIONED BOTH CELLBELAR LIKE SYSTEMS AND THEY BOTH HAVE THE PROPERTY THAT YOU'RE TRYING TO PREDICT WHATEVER'S COMING IN ON THESE KIND OF BROWNISH COLORED LINES SO IN THIS CASE, SENSORY INPUT WAS COME NOTHING ON ON THOSE LINES AND WE DIDN'T WANT TO PREDICT THE WHOLE SENSORY INPUT BUT WE WANT TO PREDICT THE PREDICTABLE PART OF THAT AND SO, THIS SIGNAL COMES IN ALONG HERE, WHICH WAS THE ELECTRIC ORGAN DISCHARGE ALLOWS YOU TO CONSTRUCT A PREDICTION OF WHAT'S GOING ON HERE AND WHEN THAT HAPPENS, THE LEARNING SHOULD BE TURNED OFF. IN THIS SYSTEM, IT'S NOT EXACTLY THE SAME BUT AGAIN, YOU HAVE A DOPAMINE INPUT THAT THERE WERE SIGNALS SOME SORT OF REACTION TO AN ODOR. AND YOU'RE TRYING TO BUILD A PREDICTION OF IT ON THE BASIS OF THE OLFACTORY INPUT THAT ACCOMPANIED IT. SO HERE THE CORRELATION BETWEEN THESE TWO ALLOWED A PREDICTION, AND HERE A CORRELATION BETWEEN THESE TWO ALLOWED TO BUILD A PREDICTION AND ULTIMATELY YOU SHOULD BE ABLE TO ELIMINATE THE DOPAMINE INPUT AND ELIMINATE THE SHOCK AND RUNNING AWAY AS YOU SMELL THAT OFFICE OF DIVERSITIOR, THESE ARE LIKE CELLBELAR CIRCUITS. THEY ARE THREE QUARTERS OF THE CELLS IN YOUR BRAIN AND YOU KNOW THE STANDARD THEORY OF THE CEREBELLUM IS THAT THEY'RE TRYING TO PICK UP CORRELATIONS WITH CLIMBING FIBER ACTIVITY SO THE ANALOGUE OF THE DOPAMINE INPUT IN THAT CASE IS CLIMBING THE FIBER ACTIVITY WHICH IN THE STANDARD MODELS OF CEREBELLUM INDUCE PLASTICITY HERE SO THESE SYSTEMS ARE QUITE SIMILAR, I WOULD CALL THIS A VARIANT BUT THESE ARE ALL SYSTEMS IN WHICH YOU KNOW YOU TRY TO PREDICT AN ERROR SIGNAL BEFORE IT HAPPENS, ON THE BASIS OF VERY NUMEROUS AND APPARENTLY RANDOMLY REPRESENTED SIGNALS AND A CLEVER PLASTICITY MECHANISM AT THOSE SYNAPSES. I WILL STOP THERE. THANK YOU. [ APPLAUSE ] >> IF I'M FOR A FEW QUESTIONS? --TIME FOR A FEW QUESTIONS? >> SO IF THE FISH MODEL, YOU'RE REALLY ONLY HALFWAY THERE, TO LOOK AT THE NEURONS. >> RIGHT. --[INDISCERNIBLE]. WHAT'S THE STRUCTURE OF THE SIGNAL. BETWEEN THE MODEL AND THE DATA THAT'S WHERE YOU GET THE [INDISCERNIBLE]. >> YEAH, EXACTLY. >> [INDISCERNIBLE]--STRUCTURE--[ INDISCERNIBLE] >> YEAH, MY GUESS IS IT'S MOSTLY THE SPACIAL STUFF. AFTER --I HAVEN'T TALKED ABOUT THE SPACIAL STUFF AT ALL BUT YOU WANT TO LOOK AT WHERE IT IS. SO THIS STUFF IS SET UP SPACIALLY. ON THE SKIN THERE'S A CENTER SURROUND STRUCTURE AND THAT'S PERFECT BECAUSE OVER THE PATCH FOR ANY ONE OUTPUT CELL, THE SELF-INDUCED FIELD IS FAIRLY UNIFORM, BUT THE BUG PROBABLY IS NOT, AND SO I THINK THERE'S A CENTER SURROUND MECHANISM THAT WE HAVE NOT TALKED ABOUT AND MY BET IS THAT THAT GETS YOU MOST OF THE REST OF THE WAY BUT WE DON'T KNOW THAT YET. >> [INDISCERNIBLE]. >> YEAH, SO DIFFERENT TYPES OF DOPAMINE NEURONS CONVERGE ON TESE APARTMENTS, SO AS FARsA THOSE GO IT'S BECAUSE AXONS ARE CONFINED TO A PARTICULAR COMPARTMENT. AS FAR AS WHAT SETS THEM OFF. PRESUMABLY WE DON'T KNOW THIS BUT PRESUMABLY WE GETS DIFFERENT TYPES OF INPUTS, AS I MENTIONED IN THIS ONE, THESE PAM NEURONS TEND TO GET REWARD ASSOCIATED INPUT, THESE GUYS TEND TO GET PUNISHMENT, ONE THING CAN YOU DO IS BECAUSE--YOU KNOW, IN THIS PROGRAM THEY BUILT THESE LINES OF ALL OF THESE NEURONS, YOU CAN ACTIVATE THESE WITH LIGHT AND YOU CAN REPLACE A REWARD OR A PUNISHMENT BY DOPAMINE NEURON ACTIVATION. SO, NOW WHAT REALLY ACTIVATES THAT CIRCUIT REMAINS TO BE KNOWN, WHERE THEY GET THEIR ACTIVATION BUT IT'S CLEAR, THAT THESE GET ACTIVATED BY REWARD TYPE SITUATIONS. >> [INDISCERNIBLE] >> YEAH BUT THAT'S VERY CRUDE. RIGHT. YOU KNOW THESE ARE THE CLASSIC TWO GROUPS ABOUT IN FACT, I THINK YOU SHOULD THINK OF EACH OF THE DOPAMINE NEURONS OF THE DIFFERENT COMPARTMENT IS SOME SLIGHTLY DIFFERENT TAKE ON IT, WHICH REFLECTS THE SLIGHTLY DIFFERENT BEHAVIORIAL BIAS THAT'S PRODUCED BY THAT COMPARTMENT. SO TELL BE COMPLICATED. IT'S SORT OF A 16-DIMENSIONAL REPRESENTATION OF EMOTION OF VALENTS. >> IF YOU WERE DESIGNING THE SYSTEM WOULD HAVE YOU PUT [INDISCERNIBLE] IN THERE ALSO? >> YEAH, YEAH. A MACHINE LEARNING PERSON THEY WOULD DO THIS. >> WHY? >> BECAUSE YOU CAN THINK OF THESE OUTPUT NEURONS AS SUBNEURONS THEY'RE SIMPLE NEURONS, THEY HAVE A BINARY ANSWER. WHEN THE SUBNEURON WAS DEVELOPED THERE WAS A PROBLEM WITH THAT CAN'T DO AN EXCLUSIVE [INDISCERNIBLE] PROBLEM. THERE'S CERTAIN LOGICAL THINGS IT CAN'T DO. THE EASIEST WAY TO SOLVE THAT IS WITH A PARAMETER INJECTION AND TRANSLINEAR TRANSFORMATION INTO A LARGER FAMILY. SO A MACHINE LEARNING PERSON WOULD HAVE EXACTLY [INDISCERNIBLE] THE SYSTEM, SUPPORT VECTOR MACHINE PERSON WOULD HAVE TOLD YOU TO DO THIS WAY. I JUST TOOK OUT THE SLIDES, I GAVE THE TALK TO THE PHYSICS AUDIENCE BUT I AND I HAD THE SLIDE IN THE FRONT. I FIGURED IT WOULDN'T FLY, BUT I WAS WRONG. >> OKAY, LET'S THANK LARRY AGAIN. >> [ APPLAUSE ]