Key points:

1. Scale of data is especially good for large NNs
2. Having a combination of HPC and AI skills is important to have optimal impact (handle scale challenges and bigger/complex NN)
3. Most of the value right now comes from CNNS, FCs, RNNS. Unsupervised, GANs and others might be future but they are research topics right now.
4. E2E DL might be relevant for some cases in future like speech -> transcript, Image -> captioning, text -> image
5. Self driving cars might also move to E2E, but none of us have enough data image -> steer

Workflow:

1. Bias = Training error - Human error. Try Bigger model, run longer, New model architecture
2. Variance = Dev error - Train error. Try More data, Regularization, New model architecture.
3. Conflict between bias and variance is weaker in DL. We can have bigger model with more data.

More data:

1. Data synthesis/augmentation is becoming useful and popular: OCR (superpose alphabets on various images), Speech (Superpose various background noises), NLP(?) But does have drawbacks, if it is not representative
2. Unified data warehouse helps leverage data usage across company

Data set breakdown:

1. Dev and test should come from same distribution. As we spend a lot of time optimizing for Dev accuracy.

Progress plateaus above Human level performance:

• But there is theoretical optimal error rate (Bayes rate)

What to do when bias is high:

• Look at examples of the ones machine got it wrong
• Get labels from humans?
• Error analysis: Segment training - identify segments where training error is higher than human.
• Estimate bias/variance effect?

How do you define human level performance: Example: Error of a panel of experts

Size of data:

1. How do you define a NN as small vs medium vs large?
2. Is the reason large NN can leverage bigger data is because it would not cause overfitting unlike on smaller NNs?

The team on Google Translate has developed a neural network that can translate language pairs for which it has not been directly trained. "For example, if the neural network has been taught to translate between English and Japanese, and English and Korean, it can also translate between Japanese and Korean without first going through English."

A good blog post on why do you need to understand backpropagation links to videos from CS231n by karpathy.

In this Pearson sounds more like an edtech company than a content publisher. I wonder at what point will Pearson release a full LMS product that competes directly with BB, D2L, etc?

The tension in that last line on the cultural environments of technology vs academia is an important -and real-tension.

Hypothes.is nicely bridges the worlds of social media and formal education.

No representation of skill trees captures the concept completely, but what I hope is evident on this page is that any Badge, with its related Playlist, should be connected to other Badges and Playlists that come before (in this case, above) it, and it should be one of a few available choices (represented in this case by other Badges and Playlists on the same row), and that it leads to other Badges and Playlists (below it), and that what comes next has choices as well.

Wrong. If you just explained your skills to other people, like a textbook does, no one would understand, unless they have accumulated personal experience similar to yours -- which they would call, after the fact, "a circuitous route" to learning the textbook content.

Speech, writing, math notation, various kinds of graphs, and musical notation are all examples of cognitive technologies. They are tools that help us think, and they can become part of the way we think -- and change the way we think.

Computer interfaces can be cognitive technologies. To whatever degree an interface reflects a set of ideas or methods of working, mastering the interface provides mastery of those ideas or methods.

Experts often have ways of thinking that they rarely share with others, for various reasons. Sometimes they aren't fully aware of their thought processes. The thoughts may be difficult to convey in speech or print. The thoughts may seem sloppy compared to traditional formal explanations.

These thought processes often involve:

• minimal canonical examples - simple models
• heuristics for rapid reasoning about what might work

Nielsen considers turning such thought processes into (computer) interfaces. "Every theorem of mathematics, every significant result of science, is a challenge to our imagination as interface designers. Can we find ways of expressing these principles in an interface? What new objects and operations does a principle suggest?"

The technologies in learning the physics are:- To bring improvements in the students’ physics ability. To bring improvements in the negative reactions of students towards physics.

Every layer needs its own weights and bias. And in tensorflow, it is a good practice to put all weights inside a dictionary, which is easier for management.

I think any good course would do this. The deeper you look, the more there is to see.

Lifelong learning is essential for everyone, not just liberal arts grads. People in healthcare or engineering have to be able to adapt to changing times and technologies just like anyone else.

This work was in schools with teachers, right? When foundations and funders of after school programs ask about how to "spread and scale" the work, it baffles me that they don't begin to answer their questions by turning to how to effectively bring the innovative after school work to teachers and students in schools. Working on the connections between in school and out of school learning is important!

Verification signal

Ground truth in machine learning

parents wanting better education because of the developing system of teaching

improve learning

its true they need to preserve the languages so more ppl learn it and so it won't die

offering more instruction=more children to learn

In other words, Google does use everyone’s information (Data as New Oil) and can use such things to target ads in Higher Education.

Just highlighting and annotating the obvious, lest we forget. There is so much of design = pathway to objectives that I want to linger over, signal boost, Mia's fundamental point: design = the learning experience. A crucial distinction.

Why isn't this in the Horizon HE report? It's more applicable to HE students who have greater opportunities and resources for experiential/self-directed learning.

learn others culture from their stand point

Set aside belief in native realism to understand essential meaning. Be in others shoes to understand perspective.

have an inside view rather than watch from the outside

involves making inferences and knowing back round knowledge

Enthrography is when you learn from the people in the culture. You don't study how they are in their culture but they teach you something

Surprising mention of “open-source platforms”, here. Doesn’t sound like these issues are absent from proprietary platforms. Maybe they mean non-institutional platforms (say, social media), where these issues are really pressing. But the wording is quite strange if that is the case.

Really? These aren’t uncontroversial, to say the least. For instance, discussion board interactions often call for careful, mixed-method work with an eye to preventing instructor effect and confirmation bias. “Time on task” is almost a codeword for distinctions between models of learning. Research in cognitive science gives very nuanced value to “time spent on task” while the Malcolm Gladwells of the world usurp some research results. A major insight behind Competency-Based Education is that it can allow for some variance in terms of “time on task”. So it’s kind of surprising that this summary puts those two things to the fore.

Ha! The Chronicle’s summary framed these categories somewhat differently. Interesting. To me, the “application” part is really about student retention. But maybe that’s a bit of a cynical reading, based on an over-emphasis in the Learning Analytics sphere towards teleological, linear, and insular models of learning. Then, the “representation” part sounds closer to UDL than to learner-driven microcredentials. Both approaches are really interesting and chances are that the report brings them together. Finally, the Chronicle made it sound as though the research implied here were less directed. The mention that it has “the potential to influence institutional practices and interventions” may be strategic, as applied research meant to influence “decision-makers” is more likely to sway them than the type of exploratory research we so badly need.

Useful breakdown. Research, predictive models, and recognition are quite distinct from one another and the approaches to data that they imply are quite different. In a way, the “personalized learning” model at the core of the second topic is close to the Big Data attitude (collect all the things and sense will come through eventually) with corresponding ethical problems. Through projects vary greatly, research has a much more solid base in both ethics and epistemology than the kind of Big Data approach used by technocentric outlets. The part about recognition, though, opens the most interesting door. Microcredentials and badges are a part of a broader picture. The data shared in those cases need not be so comprehensive and learners have a lot of agency in the matter. In fact, when then-Ashoka Charles Tsai interviewed Mozilla executive director Mark Surman about badges, the message was quite clear: badges are a way to rethink education as a learner-driven “create your own path” adventure. The contrast between the three models reveals a lot. From the abstract world of research, to the top-down models of Minority Report-style predictive educating, all the way to a form of heutagogy. Lots to chew on.

Interesting scenario. Sounds quite a bit like the role of this one person in grad school who gives you the boost you need. Usually not your director, who’s more of a name than a resource. Possibly someone with a relatively low status. It becomes something of an “informal advisor” role. “Trust” is indeed key, here. My first reaction reading this was to balk at the “trust” part, because critical thinking skills warrant other methods to gather resources. But this is a situation where trust does matter quite a bit. Not that the resources are necessarily better. But there’s much less overhead involved if rapport has been established. In fact, it’s often easy to get through a text or to start a conversation with someone using knowledge of the angle through which they’ve been recommended. “If she told me to talk to so-and-so, chances are that this person won’t take it the wrong way if we start discussing this issue.”

How can they say student-directed. It's a Facebook algorithm.

The "business" of univeristy?

xnor-net: a very efficient network

Interestingly enough, a lot of teachers either don’t know that such data might be available or perceive very little value in monitoring learners in such a way. But a lot of this can be negotiated with learners themselves.

Only partially built up.

Easier to do with a strict model for success.

Heard of SoLT? Or of the “Centre of Learning and Teaching”? Been using that order for a while, but nobody has commented upon that, to this day. There surely are some places where learning precedes learning in name and/or in practice. But the “field” is teaching-focused.

Going back to the “projects” in the Maker Movement. Not “project-based learning” with projects set through the curriculum. But the kind of “quest” that allows for learning along the way and which may switch at a moment’s notice.

Great ideas here.

Valuing the teaching profession.

传统算法的局限，在图像和语音领域，需要对不相干的钝感和对几个很小地方差异的敏感

四大金刚中的三个

深度学习的最重要的一方面就是多层特征自动学习

随机梯度下降算法，讲的很好

我擦！原来如此！！！

反向传播过程来计算一个具有多层模块权重的目标函数的梯度其实不过是求导链式规则实际应用。

数据无损压缩编码

ReLU的好

The 3D representation of a 2D image is ambiguous and multi-modal. We achieve such reasoning by learning a generative model of 3D structures, and recover this structure from 2D images via probabilistic inference.

Learning w/o Forgetting: distilled transfer learning

This is a neat idea, but do you think that this inhibits some of the students from annotating to their full potential? If I had a great idea, I might save it for myself instead of having someone else "steal" it.

This would be the optimal learning experience, for me.

How can we modernize our current education to fit the unknown needs of the future?

Or even a better question: How do we use these tools to help children learn, and acquire skills that they need?

Interesting dichotomy.

If it’s not counted, does it count?

Development as a field continues to evolve. Ideas that turn into experience generate new ideas and lessons. New ideas inform new experiences, and these are typically debated by those involved in development work.

You can read and watch the debates and discussions that took place over the past two years complementing the IDRC publication International Development: Ideas, Experience, and Prospects.

A lot hinges on those mentors. Apparently, they’re not exclusively technical, which might be an important point. The mentoring is a big part of the learning experience, surely. But there’s a certain level of complexity involved when we start discussing mentorships.

DIY, grassroots, bottom-up… but not learner-driven.

Why so coy?

R.G. Wilmot Lampros, chief product officer for Aleks, says the underlying ideas, referred to as Knowledge Space Theory, were developed by professors at the University of California at Irvine and are in the public domain. It's "there for anybody to vet," he says. But McGraw-Hill has no more plans to make its analytics algorithms public than Google would for its latest search algorithm.

"I know that there are a few results that our customers have found counterintuitive," Mr. Lampros says, but the company's own analyses of its algebra products have found they are 97 percent accurate in predicting when a student is ready to learn the next topic.

As for Ms. Thille's broader critique, he is unpersuaded. "It's a complaint about capitalism," he says. The original theoretical work behind Aleks was financed by the National Science Foundation, but after that, he says, "it would have been dead without business revenues."

MS. THILLE stops short of decrying capitalism. But she does say that letting the market alone shape the future of learning analytics would be a mistake.

Who Decides?

Though much of it may sound trite and the writeup is somewhat awkward (diverse opinions strung together haphazardly), there’s something which can help us focus on somewhat distinct attitudes towards Learning Analytics. Much of it hinges on what may or may not be measured. One might argue that learning happens outside the measurement parameters.

Time-sensitive, mission-critical, just-in-time, realtime, 24/7…

The autopsy/biopsy distinction can indeed be useful, here. Leading to insight. Especially if it’s not about which one is better. A biopsy can help prevent something in an individual patient, but it’s also a dangerous, potentially life-threatening operation. An autopsy can famously identify a “cause of death” but, more broadly, it’s been the way we’ve learnt a lot about health, not just about individual patients. So, while Teamann frames it as a severe limitation, the “autopsy” part of Learning Analytics could do a lot to bring us beyond the individual focus.

This Systems or "factory farming" approach to education seems antithetical to (and virtually guaranteed to flummox) a community-based, engaged, serendipitous and spontaneous learning explosion in traditional Higher Ed. Where are some cracks and crevices where the System has failed to snuff out the accidental life of learning?

Guest users in Moodle - can we make it easier to get them into the space to engage with students? No more boring forums when the community members or guest speakers in a f-2-f class can contribute. What about a Google form for requests? Is there a way to limit guests to only one forum?

You may consult input from amazing people like @GoogleGuacamole and @actualham who have very intentionally integrated (not just mentioning or requiring) Twitter use in their courses and implicated its value in students' connections with their professional network.

Very unique network structure. Prediction results look promising.

Unless we integrate such a platform in something else.

Though the advice makes a lot of sense, leaving it aside makes for a very empowering experience.

This might be a useful reference for further study into active learning

2013 article about Douglas Hofstadter, who has continued to pursue an understanding of the human mind through experiments with AI.

In machine learning it's also important to evaluate not just against random, but against how well other methods (e.g. parole boards) do. That kind of analysis would be nice to see.

Sometimes this takes weeks or even months. I remember taking a walk with an art teacher several years ago, and I asked him how a particular student was doing in his class, and specifically what he was working on because it was hard for me to figure out how to get him connected to my work in English. It was November, just before Thanksgiving, and my colleague said, "I haven't figured out what his project will be yet," he said, before going on to explain a couple of things he had tried without success. I was struck with how patient he was being in letting the project come to the student, and not forcing him into a prescribed curriculum. Waiting is so hard, yet the work produced once there is a "flow" for a student makes it worth the wait. This has strong implications for school structures however! We need to be with students for longer periods of time. It also has implications for how groups work together. Perhaps a student who hasn't found his/her project yet can help others?

Funny thing is, one can imagine that students -- at least my students in the Bronx -- would come up with a similar list. They have! But you can't bring it to them. There are shades and subtleties that are important in any group's list of topics. Like my students wanted to explore why people from the Bronx are not treated the same as people from elsewhere.

connections over content. yes!

Active learning, engaged learning, experiential learning, and owning the learning best happens with transformational teaching.

It is easy to allow technology to replace memorization and other skills. We should be mindful of what we allow it to replace. Martin Luther King had a large store of writings memorized -- and it served him well when he wrote the Letter from a Birmingham Jail.

We need more tools that will aid skill development instead of replacing useful skills. Spaced repetition software to assist memorization is one example. Phrase-by-phrase music training programs are another. The same ideas can be applied to memorization of text.

We should have control of the algorithms and data that guide our experiences online, and increasingly offline. Under our guidance, they can be powerful personal assistants.

Big business has been very militant about protecting their "intellectual property". Yet they regard every detail of our personal lives as theirs to collect and sell at whim. What a bunch of little darlings they are.

Not only do schools not do that, traditionally they have "taught" creativity and curiosity out of students.

Would love to hear Campbell or Kuh elaborate on this. Identifying "connections" as more important than identifying content/information? A new way for searching the Internet? Mining connections among content/people? Mining the connections I've made among content/people on the Internet?

My understanding of what happens in Koranic schools is very limited but friends in Mali have described a process by which they had learnt the Koran by heart before they knew the meaning of any of the words. Something similar has been discussed in terms of the Bhagavad Gita.

An opportunity for library outreach? Public libraries offer access to lifelong learning resources already. Could they go further with offering assistance & promoting advantages?Problem is that their resources are very limited.