Neurofeedback
  1. Humans can learn very abstract reward environments. We can play the saxophone, find that applying pressure just so leads to this sound. The saxophone is a function from the space of motor neuron activations to the space of audio neuron activations.
  2. The function is quite arbitrary; we have no reason to expect the Saxophone function to look like other functions we've encountered before, like the Singing function, or the Typing function, or the Nose-Scratching function.
  3. All functions can be learned. Provide a consistent map from motor neurons to sensory neurons and the brain will do the rest.

We should be able to program skills into humans. Create new functions, e.g. from vocal cords to music synthesizer or from finger joints to image, that feel arbitrary when you first try them, but become an extension of yourself as you learn the function. Could we interrogate the process of learning this way?

If I recorded neurons from rat vocal cords and mapped it to English words, shouldn't rats be able to speak English?

Beyond plays and musicals making life that much more enjoyable, they also act as fuel for better action. When you leave a good show you feel a deeper connection with emotions words can’t yet describe. This deeper sense of connection acts as fuel for action in the real world. A good example of this is Hamilton. After seeing it you can’t help but feel a sense of urgency. This caused me to 1) dream bigger and 2) pursue these dreams as quickly as possible. In times of crisis it’s still the show I run back to.

The qualia of Hamilton is just so good. Images of union bells, young Alexander in gold and black, 'people shouting in the square', rolled into a gleaming ball of meaning.

Books and stories create this high-dimensional blob. That is the meaning of the book. We forget all the lines and everything that happens, as they are the steps we took to get to the meaning. But much like the words "couch", "red", and "window", the meaning of the words "Hamilton", "A Hundred Years of Solitude", and "Your Name" are remembered forever, becoming a primary part of our cognition.

Afika said something funny today: "Steve Jobs-ing around campus".

I stayed at two dorms in MIT and Waterloo this winter. I felt silly and sheepish about being there as a non-student when everyone around me was a student, and I found it hard to explain why I was there.

Afika used the word 'Steve Jobs-ing', and it was like, Oh. I've been doing it, and many of my friends have done it on independent occasions. It's just following your interests!

What is most of the world made of? Some things are small, repetitive units — water, or salt crystals, or elemental gases. Some things are slightly larger units repeating in more interesting ways — microtubules, lipid vesicles, gluten. Some things are enormous, single units that go and do things — proteins!

Having a warm soup and feeling the warmth flow from the insides to the bloodstream to the surface of the skin, breaking out as sweat.

Is it really just molecules jiggling? All of them, small or big, water or bone, jiggling at once. If there is memento mori for all men must die, shall there be memento calor — all molecules must jiggle?

Monday was not a collie or a setter or a hound or a Newfoundland. He was just, as Jem said, "plain dog" — very plain dog, uncharitable people added. Certainly, Monday's looks were not his strong point. Black spots were scattered at random over his yellow carcass, one of them blotting out an eye. His ears were in tatters, for Monday was never successful in affairs of honour. But he possessed one talisman. He knew that not all dogs could be handsome or eloquent or victorious, but that every dog could love.

Inside his homely hide beat the most affectionate, loyal, faithful heart of any dog since dogs were; and something looked out of his brown eyes that was nearer akin to a soul than any theologian would allow.

from Rilla of Ingleside, the last book in the Anne of Green Gables series

I thought this was so expressive. Her dog is a complete lemon yet wonderful all the same.

There's a spectrum of ways to think about a person: how creative, how productive, how successful they are happen to be on one end, the clinical, high-level characteristics. The low-level characteristics are very qualitative: the things they say, their reactions to you, how they emote and relate. I wouldn't be able to describe the type of person I get along with, much as I can't draw the cloud of all things I find funny, but as soon as we interact I know what this person is at a low level.

In SF, a lot of people like each other at a high level. Of course, wouldn't you want to be around the most awesome, creative, and interesting people you can find? But for friendship, what does that matter? To love, laugh, and play has everything to do with who they are and nothing to do with what they've done.

My eyes have been bad recently, maybe 5% worse than normal vision. I can still read, but it eats at me. One time yesterday, I opened an essay, started reading, then dropped it completely. Never mind, I don't feel like it.

I wonder if these papercuts are present elsewhere. The iPad replaces clicks with taps and makes you more likely to play with it. What are we not doing that we would like to, but don't because it's unpleasant?

There's a sense in which websites are just places. Like college, a cafe, or Singapore. Web users can teleport anywhere, so places on the web tend to be formed around their internal characteristics, like interests, hobbies, and problems.

People make places. Tumblr was filled with teenagers. Pinterest started off with middle-aged moms. The more middle-aged moms there were, the more middle-aged moms it attracted, and now it has 400 million people, most of which are middle-aged moms.

Places are bespoke. A cafe might have reading lamps, jazz music, green velvet chairs. Our psychological space is enormous, so the value of a place is how it constrains it. The cafe has rich affordances to get you to read, to linger.

Dogs are the perfect interface.

I pet the dog because I want to. I play with the dog because I want to. Every interaction is the means and the end. No extra movements to navigate the interface.

Sleep is fascinating. It is when the brain is decoupled from inputs, and patterns of oscillation arise from the synapses and connections as they were right before you sleep.

Artists

Neural interface companies need artists.

  • Steve Jobs's craft was creating demos that make your heart and soul sing. While the other companies were creating massive feature sets, Steve focused on a single compelling demo
  • The original Pixar movie was good because of the storytelling
  • Microsoft hired a YouTuber to envision AR technology
  • Andy Matuschak was most excited about a feature on the Apple watch where he could tap the watch, and his partner's watch would glow.

There must be a version of neurotech that feels compelling, and that version should have nothing to do with technology and everything to do with people

(BCI is interesting to me when we can share new mental spaces — what we love is still people, rather than things)

Visual stimulation

Recently I went to demo Varjo, a VR headset with human-eye resolution.

There were enough issues with motion artifacts that you couldn't fully believe the scenes were real. Even so, I felt like an infant, interested in the appearance of the world. I stared at desks and carpets, knelt to look at the floor, crawled to see things from below.

Two other things stuck out. Teleporting in the space felt real, the closest you could get to teleporting in real life. Also, I felt delight whenever I saw an object suspended in the air.

it's also a "built" language — 1800s people might not enjoy the pop songs of today? Could neurotech be a similarly "built" technology?

start with clear (primary motor, visual) inputs/outputs, use THOSE to move 1 layer deeper, use THOSE to move to cognition

MVP model - apply to neurotech? specifically, pursue the "complex", red-hot brain functions from the get-go

We should just publish videos instead of papers!

That sounds really fun! I'm most likely in Boston, though I'm super down if you want to explore upstate new york (Finger Lakes looks lovely!)

On another note, Winnie was asking if we wanted to do something tomorrow or sun?

Follow the science. A scientific problem has no reason to stay inside a particular field. If you start needing to reason about chemistry, then learn chemistry. Fields are fake; self-identity is a hindrance; nature will not be cognitively pleasant.

Know that you can learn We are all general intelligences. Differences in the current state of anyone's brain are all noise in light of our astounding biological ability to learn.

Learning is learning I often get caught feeling bad or feeling good about learning, but neither of those really make sense. Even if all I've done is struggle, all the same I've been learning.

I'm not saying this for spirit or for morale, but because it seems like the truth of what learning is. If you spend five hours on a problem and can't figure out the answer, it doesn't mean you haven't learned, it means you spent five hours learning. Your brain has rewired and formed new attractor states, you've done necessary work, the activity you've done is called learning.

Neurons operate at all levels of scale.

Stimulating a single neuron increases probability of a rat responding.

Milisecond precision is necessary to capture network dynamics. Action potentials last 2 ms and their closely associated calcium fluctuations last between 1 ms (rise) and 10-100 ms (drops). Optical dyes have a lot of nice properties for recording (e.g. we can record from 1 million neurons!) but they rely on calcium as an indicator: we can't draw causal relationships from the data at that time resolution.

possibilities for timing codes, spike-timing dependent plasticity mechanisms, and other effects relying on temporally-precise spiking patterns

What is the qualitative experience of a seizure? What happens when the oscillations of the brain, usually so finely mechanized for computation, become global in character, roll the entire ship to a single rhythm?

How would it "feel" to not have a fovea?

Depending on where an object is located in your visual field, its clarity may differ by over an order of magnitude. The difference is so great that trying to observe things outside your fovea is akin to blindness.

You can experience what a uniform visual field feels like by wandering around in the dark. A different set of receptors are active for low light conditions, and unlike the cones these receptors are evenly distributed throughout the retina.

So in low lighting, moving your eyes to an object doesn't increase clarity like it would in daytime lighting. Everything in your visual field is uniformly clear; or at least, uniformly unclear.

Cone activation is approximately inversely proportional to background illumination

Neuroscience has the same vibe as

In the human brain, no less than 1/5th of the visual cortex is dedicated to facial recognition. So long as it has improved survival, we've become more and more sensitive to the curves of the human face.

Compare that to how the face of, say, an elephant appears to us, and you and I can simulate how largely our perceptions are distortions of the objective world. In our world, human faces are fundamentally higher-resolution than anything else.

The interface theory of perception: the perceptions of an organism are a user interface between that organism and the objective world

Our perceptions care a little bit about the statistical structure of the universe, but mostly care about our survival. So if you add some noise to a scene which then renders it a physical impossibility, we don't actually care. We don't see it.

An ML model is trained under different conditions.

Could it be that an adversarial image — a picture with added noise, or a single black pixel — is actually a physical impossibility, a perversion of the way EM waves travel in the universe? Could it be that ML models actually learn the underlying statistics of images, and become confused when presented with what seems to them to be an obvious impossibility?

Rather than saying the ML failed because it didn't see the airplane after the noise was added, perhaps the questions is, how come we don't see all the blatant noise that has been added to the airplane?

Marley 11/4/2021, 4:38:00 PM

Adversarial examples for the OpenAI CLIP in its zero-shot classification regime and their semantic generalization | Stanislav Fort

  • it genuinely believes that it's seen a frog or a toad.

The 'Blind Watchmaker' is great, evolution is a general idea that you can see everywhere, it tickles your imagination. I've been trying to grasp at an explanation of our cognition through evolution.

Why, for instance, are trees in forests so tall? ... There has been an arms race in which forest trees became larger as the generations went by. At every stage of the arms race there was no intrinsic benefit in being tall for its own sake. At every stage of the arms race the only point in being tall was to be relatively taller than neighbouring trees.

The imagery is of an ecosystem stuck in bizarre local optimums.

Why are humans anxious, mimetic, self-conscious, ego-defending, status-seeking? We have all this cognitive infrastructure to be similar to each other. He who was most similar to others was the most likely to survive.

These things were developed at a time when our lives were local--

[our brains] are equipped to assess risks of things happening to us personally, or to a narrow circle of people that we know. This is because our brains didn't evolve under conditions dominated by mass media.

People are happy when they're in small towns and large families, of course they are, because their brains are tuned for it. Instead of thinking of modern media as the harbinger of mental illness, think of mental illness as the set of cognitive punishments that were meticulously designed in humans to keep them happy in small groups. They were always there, just properly tuned.

Marley 8/30/2021, 4:21:15 PM