2014-05-09 What I wrote on Facebook - Thought about [speed reading - Writing Method We talked about support tools. - We talked about Electronic KJ method tool and the Narrow screen problem. - Bundling Functions - Consideration of the review tool “should have been promptly rewarded and reinforced for the behavior it sees. - joint editing may have been futile. - As of 2019: Even one person needs to co-edit.. - Export method should be able to be done offline.

  • What is the purpose of speed reading?
  • It’s not good to be in a state of “I was going to read it, but I didn’t.”
    • If you decide to read it, read it.

    • If you are storing it in the hope that it will be found in a search when you need it in the future, make sure it will be found in a search.

    • If it is neither, throw it away.

    • I want to make [Bots to assist in reading

    • → hippocampus.


  • As a result of my speed-reading training over the past while, the books I haven’t read have been eradicated from my home.

  • In a sense, the goal has been achieved, but the result is a lost cause.

  • What should we aim for next?

  • Journals you haven’t read

  • It tends to accumulate because it arrives on its own with content unrelated to my context.

  • I thought I should read this, but I have an idea that you should just throw it away.

  • Downloaded Papers

  • Because of the lack of substance, as soon as it becomes invisible, its existence is forgotten.

  • Should I read it?

  • Some things in the world you should read and some you shouldn’t.

  • You don’t even have to read the papers you downloaded with the intention of making them into a corpus.

  • How do I decide if I should read it or not?

  • Not that it matters, but I like the way the list appears at the top when you write.

  • My idea writing program should also look like this for mobile pages

  • On the other hand, it’s not good that when you write, the keyboard closes and you have to tap again.

  • Ah, I see, my one has a section for bibliographic information as well as text, so the screen is already half full and there’s no place to list it.

  • Without a list, there’s no way to recover when you lose your train of thought.

  • Trying to use the same UI when writing out ideas in a one-person brainstorming session and when making excerpts while reading a book is a mistake in the first place.

  • You don’t need bibliographic information when you’re brainstorming.

  • Then the blurbs are like this.

  • I guess you don’t need the history of the excerpt the other way around.

  • Both are too focused on writing it out to figure out how to review it.

  • The next thing to writing out is an immediate KJ method map, inappropriate?

  • Small screen

  • listing

  • There is not enough room in the space to make sense of the placement.

  • I’m limited to 100 sheets in a row.

  • Isn’t it better that the bundling feature is not yet implemented so it can’t be compressed?

  • I’m thinking that extracting and rereading with appropriate frequency is a much higher priority.

  • Reading over excerpts and leveraged notes is less motivating when the volume exceeds 8 printed pages.

  • Because it’s harder to estimate how long it will take to finish?

  • Is the leverage memo the first thing to read back?

  • I developed kabuki accessory representing a woman’s head on a white background when the amount of leverage memos and excerpts I read back became too much and I became less and less motivated


  • I thought, why not just read the excerpts instead of looking at Twitter, the density of interesting stuff is much higher here


  • Okay, I guess I’m going the wrong way in trying to do the read back “properly”.

  • It’s not a good idea to default to “oldest reviewed first” when the random display was implemented.

  • They should have promptly rewarded and reinforced the behavior of “looking.”

  • I’m fed up with creating services on GAE and shrinking the free quota, but there are people who have learned their lesson creating on GoogleDocs


  • I was attracted to the real-time collaborative editing, but in the end there’s no room for multiple people to edit on a small screen.

  • Only a few confident people can edit without hesitation in a tool that only allows collaborative editing.

  • The majority of people think about editing it alone and making it somewhat “unashamed to show to others” before releasing it to the public.

  • Confident people are not afraid to show others that the editing process is trial and error, that it’s unfinished and of low quality, because they’re confident.

  • Also, I’m not sure if such a person would publish edits to an unspecified number of people.

  • If you are confident in what you have created, you are uncomfortable with others editing it.

  • I wonder what the mindset of the people writing on Wikipedia is?

  • Ah, so that’s why the editing wars happen.

  • You simply can’t predict that “anyone can edit, so what you edit can also be altered or deleted, but you shouldn’t be angry”.

  • We were talking about creating a tool to review excerpts so that they could be used offline and synchronized when online.

  • The one I’m working on now, I was worried because it hasn’t been tested to cover the path of editing offline and then going online and so on.

  • Each item should not have a last modified date and time, if it does, you can overwrite the older one


  • No, no, you can’t, you need to merge when both are updated.

  • If you start thinking about that, you end up having to keep the pre-edited version after editing, like version control.

  • Since the server is there, should the mobile data only be up-to-date when syncing with the server?

  • First of all, we should make a prototype that works with orthodox cases without going too deep into dealing with rare cases such as

  • Back to what I’m talking about, what to speed read.

  • The importance of rereading the excerpts

  • But we need clarity of purpose.

  • The goal of eliminating piles of reading was achieved.

  • Newly arrived books can now be looked through on the same day.

  • The following achievable objectives need to be identified

  • As an analogy, if you download a paper you intend to read, you should look over it that day,

  • When you receive a subscription to an academic journal or other publication, you should look through it that day.

  • What was not worth reading today will be worth reading tomorrow?

    • Can’t say I wouldn’t get it

  • leaving that aside

  • Not a whole lot to say about whether you should read what you’ve already accumulated.

  • Leaving it in a state of “I meant to read it but didn’t” is not a good idea.

  • If you intend to read it, read it. If you keep it in the hope that it will be found in a search when you need it in the future, you should make it clear that it is so and make sure it is found in a search. If it is neither, it should be discarded. It should be “processed” one way or the other.

  • In other words, you should “look over” everything at a level high enough to determine that

  • Some of the things that are being “looked over”, including books, should be read more carefully

  • What to do with it.

  • What does it mean to read carefully?

  • A state of mind where investing more time in reading that book could yield more new

  • Clarification of objectives

  • What’s the point of reading a book to gain new insights?

  • Trying to answer this question in a general way is too abstract and contradicts the fact that goals must be specific.

  • The general answer, by the way, is “improve the model in your brain and increase the probability that you can cope with possible unknowns in the future.”

  • To improve my model, I need to encounter things that cannot be explained by my current model or that contradict the explanation in my model.

  • It’s what you can’t easily swallow that’s worth chewing on.

  • Encountering the unknown should not be a goal, it reduces motivation because the effort will not be achieved with certainty.

    • Really?
  • As for the work that we are forced to do, it is demotivating when our efforts do not produce results, but from a flow-theoretic point of view, 1/3 success rate is the most exciting.

  • Goals for a quick sense of accomplishment are different from goals for a sense of accomplishment

  • Also, you can’t know the probabilities in advance, so it’s hard to adjust them yourself.

  • Is it easy to fall into the flow of things that can be adjusted by you?

  • A goal that is close and can be achieved for sure is that I will record the number of books I read and the amount of time I read them. Now that you are recording them individually, how many books have you read in total? How long did it take you to do that? Why not visualize it?

  • The goal, which is a distant and challenging one, is to find interesting things. N or more per week. With this way of setting up, you can adjust by reading more or daring to try different fields if you don’t find any. Record the time taken. You can measure the difficulty and adjust n to achieve a reasonable balance.

  • Blog about interesting things you find. Reinforce preferred behaviors by linking them to your need for approval.

  • First, a tally of what we’ve done so far, then measurements for the future.

  • Attending study groups, like reading, is also a mess where you invest time to get something interesting.

  • Let’s record it all. Week by week, tally up the time invested, the amount of money spent, and the interesting things gained!

  • If you conceive of a system that is too complex from the beginning, you’ll fail to implement it due to your poor administrative skills and you’ll hate the whole thing.

  • First, it should be implemented minimally to validate value.

  • I want to put a bot in here.

    • Recommendation from sentence to sentence
    • Suggestions with sources in blog corpus and library corpus
    • Enhanced with a like.
  • I’m thinking of the hippocampus model.

  • Game rules omitted.

  • Running or turning at zero knowledge is randomly determined.

  • Taking actions that seem good when knowledge is available, but also exploring, i.e., reinforcement learning.

  • Running and hitting the wall, recognizing the wall condition

  • Do we have the state of each step in our short term memory as a kotatsu, or is it a gradual culture of identification?

  • If you’re walking down the street and the unification happens at “oh, this is where I’ve been before,” then you’ve had it separately until then.

  • But what is that unify? If you equate states when given the same input, they don’t become separate in the first place.

  • If the input is not identical due to some noise or historical information, the UNIFY side will have to allow for some error, and we are back to square one.

  • Is the Blue Star Algorithm relevant?

  • For this issue to be resolved, it would mean that UNIFY is not a zero-one.

  • That is, the states are not a discrete set, but are distributed over a continuous space, with distances far and near.

  • Distributed representation of states

  • Should I think of a grid cell as a distributed representation that just happens to be a grid because the state is two-dimensional?

  • Let’s see, distributed representation of state

  • You will hit a wall, you will feel discomfort, you will anticipate that you will feel discomfort again as you proceed and you will avoid it, you need a mechanism that allows you to do this.

  • Needs a distributed representation of the state and a neural net that takes the next action as an argument and predicts the reward to be obtained

  • This can be learned every step of the way.

  • I need to relearn reinforcement learning.

  • I’m guessing they’ll predict the next state, too.

  • Q function returns a reward with the state and next action as arguments

  • Let’s also predict the next state of this

  • I’m sure my brain simulation will learn to make a U-turn when I hit a wall, for now.

  • If they can learn that if they make a U-turn, they’ll be left with nothing where there was nothing before, they’ll turn and explore as they see fit.

  • How is a distributed representation of the state created?

  • If the CA3 recurrent is randomly initialized, it would be possible to randomly initialize the variance representation, since we’d get an appropriate value at each step anyway.

  • Is there a grid-cell-like cycle in the time direction as well?

  • To make the state of nearness near.

  • If a neural net has been trained to return a state from a previously experienced state-input pair, where the input returns the state for the argument, then it seems possible for the recurrent net to relearn, but in that case the firing pattern is just an ID, not a distributed representation

  • Still, that’s okay.

  • Then the incoming states of similar inputs come to fire together, which is redundant, so compressing them allows unification.

  • I think I can do it if I put sparse constraints on it.

  • Then it’s going to be a place cell thing.

  • If this is a distributed representation with sparse constraints, would it be a grid?

  • Not so easy.

  • It is a good thing if a place cell is created.

  • Once we have a model that can produce place cells, let’s feed them natural language as input.

  • I hope the rat has a neural net that receives the state and returns the next action, and that guy is initialized randomly at first.

  • Then you’re repeating the same behavior.

  • How to do something like UCB1 with neural nets

  • A mechanism that narrows when nothing happens.

  • Thompson sampling.

  • No, I guess not.

  • Need a neural net that can express confidence

  • Is it enough to have lots of outputs?

  • There is a randomly initialized net that takes a state and returns an action, with several outputs for each action and a majority vote.

  • The condition looks better coming in before the sparse.

  • Actually, maybe the behavior is a distributed expression, too.

  • If we used rewards directly to learn this state-action mapping, we would not be able to deal with delayed rewards, so life began to predict and estimate the value of the next state sometime in the future.

  • We have to do something here, because if the rats don’t run around, the map won’t be made, and if they don’t understand the delayed rewards, they won’t run around.

  • All you need is a lot of output from a neural net that takes a state and an action and returns a value.

  • Initially initialized randomly, the distribution of the output is considered to be a uniform distribution and becomes sharper as the learning is repeated

  • If we choose the largest randomly selected one of these outputs, it is roughly Thompson sampling

  • Oh, you don’t have to work hard to realize Thompson sampling or UCB1, or even epsilon greedy.

  • If only we could have introduced neurons that randomly select

  • After sparsing the state, it’s normal reinforcement learning, so we should implement it normally and first test the expectation that nets from input to state and recurrent nets and sparsing will produce place cells.

  • I couldn’t sleep and I have an hour and a half before I leave for my trip


  • Sleeping now is dangerous, so we’ll just have to think about it further.

  • So far, we have created cells that respond to specific locations, but their positioning is still in question.

  • I wonder if it learns state transitions even after sparsing


  • There is no recurrent over here, though.

  • Hourly average.

  • If a stimulated neuron remains active longer than the temporal resolution of other neurons as a way to make transitions closer or farther away, then the neuron acts as a compressor of the temporal axis.

  • The orientation cells are born here because “running straight” is a frequent, natural behavior of rodents.

  • I hope the duration of the firing is determined by some concentration, and that there is a concentration gradient along the hippocampus.

  • I don’t know why it’s on the grid


  • Surprisingly, you might be able to get a grid just by compressing the two dimensions with time averaging, though that’s a negative.

  • It seems to be born out of the behavior patterns of rats.

  • My interest is in natural language anyway, so I don’t care about the grid.

  • It would be amazing if we could inadvertently grid

  • What happens when you create and raise an intelligent body that can only read and write?


This page is auto-translated from /nishio/速èȘ­ă«ă€ă„ăŠè€ƒăˆă‚‹ using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.