Synopsis of our March 15, 2018 YHouse Luncheon talk by Kenji Doya
Presenter: Kenji Doya (Okinawa Institute of Science & Technology, Japan)
Title: How Does the Brain Wire up Itself on the Fly?
Abstract: “The standard paradigm in functional brain imaging is to ask subjects to perform tasks requiring a certain computation or not, to see which brain areas are more activated, and to conclude that those areas subserve the computation. However, we do not really know why and how those specific brain areas can be activated and connected when they are needed. As deep learning provides solutions to specific computations like image recognition and language processing, how to select and combine those networks as needed in different tasks and situations is a critical challenge in general and autonomous artificial intelligence. In this talk, we will explore possible anatomical and computational mechanisms that realize modularity and compositionality of the brain.”
Present: Piet Hut, Yuko Ishihara, Ayako Fuqui, Yentl Dudink, Thomas Doctor, Daniel Polani, Eran Agmon, Steve Lin, Randy Beer, Olaf Witkowski, Vincent Paladino, Michael Solomon
Kenji began his presentation with a slide saying: “To explain the universe, people called for god(s). To explain life, people called for soul. To explain human intelligence, people now call for consciousness.” Rather than working on the neuroscience of consciousness, he proposes “mental simulation” as the target of research. Mental simulation can be defined as brain processes using an action-dependent state transition model: s’ = f(s,a) or P(s’Is,a) and its use can be verified from the behaviors. Mental simulation allows us to estimate the present from past state/action in the face of sensory distortion by noise/delay/occlusion. It can be used for predicting the future using model–based decision or action planning. It also allows us to imagine in a virtual world, which is the basis for thinking, language, or building theory in science. The importance of having causal models of the world and running mental simulation is recognized by other researchers in AI and cognitive science (Lake et al., BBS, 2017; Hassabis et al., Neuron, 2017).
How is mental simulation realized in the brain? He presented his view that the cortex is specialized for unsupervised representation learning (capturing data distribution), the basal ganglia for reinforcement learning (predicting reward), and the cerebellum for supervised learning (input-output mapping from samples). He outlined three ways of action selection: 1) Model free, in which the action values learned in the basal ganglia selects an action representation in the cortex. 2) Model based, in which a candidate action held in the working memory in the cortex is sent to a transition model in the cerebellum to predict the next state, which is then evaluated by the basal ganglia to decide if it is worth execution, and 3) Memory based, in which the state-action pairs represented in the cortex is memorized as a state-action mapping in the cerebellum, allowing automatic execution.
He tested this theoretical construct using a computer game called a “grid sailing task” (Fermin et al., Sci. Rep., 2016). In the game one must move a cursor to a goal on a grid, where the cursor could move in only 3 out of 8 directions. The player was asked to begin moving immediately or given time to plan the steps. The improved performance with pre-start time proved the use of model-based planning. He then proceeded to use fMRI (functional magnetic Resonance Imaging) to see which areas of the brain are activated during the pre-start time. These studies confirmed his hypothesis that the global network linking the cortex, the cerebellum and the basal ganglia are involved during mental simulation of cursor movements.
However, fMRI research including this study depends on the empirical rule or assumption that brain areas that perform required computations for a given task are activated. The question still remains, “How can these modules in separate brain areas become activated and connected as needed?” He next outlined four possible mechanisms of module selection/connection: As the computational principles, 1) a learning switch board with selection gating, 2) Competition by prediction (e.g. MOSAIC architecture by Wolpert & Kawato, 1998), 3) weighting by certainty (e.g., Bayesian cue integration), and 4) subsumption by more elaborate mechanisms. Possible circuit mechanisms in the brain would include: 1) central executive in the prefrontal cortex, 2) gating by the basal ganglia, 3) selection/connection by the thalamus, 4) Brain waves, and 5) Spike coherence.
His future plans involve testing these ideas by developing modular/compositional architecture that really works; using behavioral, fMRI, and ECoG experiments focusing on module selection/connection mechanisms; and using optogenetic experiments to manipulate particular pathways.
His presentation ended here and the discussion began.
Olaf: Combining a huge pool of computational elements but low connectivity, can you get complex behavior?
Kenji: The brain can be a big reservoir to produce complex output, but needs selective activation and connection for useful behaviors.
Dan: How much of the information stored do you need? Determining that depends on context, and the context can then change as you proceed. What in evolution induces new brain areas or ways to switch context within a brain area?
Kenji: That is still an open question.
Dan: For language, are there different areas in the language center for people who speak multiple languages?
Olaf: There are clearly different areas of the brain involved for those who are early bilingual than for those who learn a language later.
Eran: Early learners get a dedicated area while later learners get bulges from your existing language area.
Dan: If Broca’s area is absent at birth, children can still learn language.
Kenji: The cortex has plasticity. Blind people may use the visual cortex for tactile sense.
Eran: There are not specific brain areas but plasticity and process that result in areas.
Olaf: Can we identify dictionaries of corresponding brain areas?
A: The Human Connectome Project is working on that.
Dan: When in evolution did the brain develop these capacities?
Eran: A single e. Coli cell, which he is modeling now, can do context based switching. Genes for metabolizing glucose can be turned off and for using lactate turned on by sensing the change in the environment and then expressing those genes. The single e. Coli cell has no brain.
Dan: How would you find mechanisms for context switching in an octopus, which has 9 brains? He is fascinated by the octopus. They are not right or left handed, but the different arms do have different functions and strengths. He referred to a video of an octopus picking up a coconut, sitting on the coconut, and using its arms like legs and the coconut as a roller to “walk”.
Michael: Are you aware of the hetero-modal cortex? These are four areas located on the left and right frontal cortex and at the junction posteriorly of the temporal, parietal, and occipital lobes on the left and right. The hetero-modal cortex functions to interpret input from multiple uni-modal sensory areas in the brain to create an integrated model of the world. The Anton-Babinski syndrome is a fascinating situation where people who had sight but then lost both the right and left occipital (visual) cortex to stroke or other disease, and are now completely blind, actively deny their blindness presumably because the hetero-modal cortex continues to provide a visual interpretation of the world. Even when they bump into objects in the room, they deny that they cannot see.
Kenji: The parietal area does have a lot to do with integrating multi-modal inputs and maintaining perception when there is no input. Kenji showed slides of an experiment using an auditory virtual environment for the mouse. Even when the sounds are intermittent, the mouse could perform normally by updating the representation in the parietal cortex coding the goal location.
At Piet’s prompting, Yuko agreed that modules fit with the idea of Epoche – suspending judgment. Epoche sets aside your prior beliefs and that creates a space for other possibilities. You neither believe nor disbelieve.
Q: Do value judgements occur in the basal ganglia?
Kenji: Yes, probably in the striatum in the basal ganglia.
Piet: In lucid dreaming, what part of the brain tells you when you are dreaming?
Vincent: Much is contingent on training. Zen practice allows you to train your mind. Plasticity allows for change in distributive cognition.
Yuko: Suspension of judgment can have practical implications. In Zen practice, you let go of what has accumulated and get back to reality.
Michael: Attributing functions to geographic areas in the brain is likely an over simplification. Cognition is more likely a process involving the entire brain.
Eran: The areas are not present at birth, but develop as a consequence of sensory input as the brain matures.
Olaf: Focusing on Attention becomes important. Attention is not so much what you put your light on, but what you are able to forget or ignore temporarily. This is like Epoche. Under hypnosis, you can forget your own name by visualizing a series of steps: Put your name on a piece of paper, put the paper in a box, the box in a river, etc.
Piet: Epoche does not require that you no longer want a cookie, but only that you add a dimension – you are aware that you like cookies and can then consider the possibility that you don’t want a cookie.
Dan: The epoche example is like shedding the idea of the ether in physics. For a long time, the ether was postulated as a medium to explain how light is transmitted in space, until that model was no longer useful.
Piet: Husserl asked what could it be like to be a wave without a medium like water or air? Only a wave?
Vincent: Attention involves the idea of boundaries. “At Tension” is tension at a boundary.
We ended our discussion here.
Michael J. Solomon, MD