04/30: Munro

[shevajia]

Comment below with questions or thoughts about the reading for this week's workshop.

Please make your comments by Wednesday 11:59 PM, and upvote at least five of your peers' comments on Thursday prior to the workshop. You need to use 'thumbs-up' for your reactions to count towards 'top comments,' but you can use other emojis on top of the thumbs up.

[tonofshell]

Thanks for sharing your work! While, everything you talked about in your book is really interesting, it is also pretty technical. Do you think these methodologies can be used to make deep learning techniques more accessible in any way?

[luxin-tian]

Thank you for this presentation. As you are also an active responder to the current COVID-19 pedantic, could you please share some insights into the applications of machine learning into the public health and crisis management fields?

[JuneZzj]

Thank you for presenting. The idea of active learning is very fascinating. It allows people to concentrate more on innovating because pytorch has already built the block innovation inside. Why do you think active learning is breaking the boundary of other ML fields and what impact it will have on those fields? Thanks.

[sanittawan]

Thank you in for your talk and I am looking forward to it. I am very interested in the idea of diversity sampling since it tackles "unknown unknowns." You mentioned that one way to approach diversity sampling is to use a clustering algorithm (or a bunch of them). Since we normally do not know how many clusters they are going to be in the data set (unless one has a specific domain knowledge), do you have any suggestions on how to make sure that the result from this approach is reliable? How do we know that we will sample from every possible group in the data set? Is this approach best supplemented by other approaches?

[yongfeilu]

Thank you for the excellent presentation! My question is how you guarantee that the model derived from machine learning can accurately model the behavior of human beings. Even though algorithms can simulate rational human, how can they model the behavior of real human who just switches between rational an irrational mode under different circumstances?

[jsgenan]

Hi Thank you for bringing your industry insights to our workshop! The book is fascinating. You mentioned your experience of working in disaster response and health at Harvard, so I was curious about what the story of "human in the loop" is like in that industry? Was it different from the data science world?

[skanthan95]

Thank you for presenting! I work with human annotation for transcription, so chapter 7 of your book was particularly interesting to me. I was wondering about how your findings relate to turn taking behavior in conversation analysis, and which machine learning approaches would work best for reliably diarizing (segmenting and allocating) speech. In my experience, it has been very difficult to mark TRPs (transition relevance places) and there's a lot of subjectivity about where they should be placed, so I'd be curious to see how a trained machine learning model would classify the boundaries.

[HaowenShang]

Thanks for the interesting presentation!I see you are currently helping with the response to COVID-19. Could you please elaborate how the machine learning techniques applied into the pandemic of COVID-19? Thanks!

[timqzhang]

Thank you for your presentation ! I also wonder what are the implications of machine learning in the consulting, which may be considered as a profession that need much human thinking due to its complex cases? Are there any specific ml tools on consulting?

[yalingtsui]

Thank you for your interesting presentation. I wonder how machine learning could help us deal with covid or other disaster?

[luyingjiang]

Thank you for your presentation! I think the idea of active learning and human involvement is very fascinating. Could you talk more about the unbiasedness of human labeling in these models? In particular, are there certain ethical concerns due to human involvement in these processes?

[KenChenCompEcon]

Thanks so much for presenting! It is so interesting and fundamentally meaningful to see searches on how human-in-active methods can help the machine learning algorithms to improve themselves? However, I am thinking about if there is a limit on models trained in such manners when humans even feel difficult fulfilling certain tasks? Thanks!

[TianxinZheng]

Thanks for presenting! I'm interested to learn more about how machine learning could help us responding to disasters such as covid19.

[adarshmathew]

Thank you for your presentation, and for writing this book! A lot of what you've written about auditing neural network architectures and designing systems that can annotate data makes me wish I had this book at my last job.

Your proposed methodologies to interrogate hidden layers are very interesting, and I loo forward to implement them soon in an ongoing project. So could you list out any notable case studies, and the kind of success they've had with any of these approaches?

Additionally, similar to what @bakerwho mentioned, I'm curious about how you think Diversity Sampling addresses unknown unknowns, because that seems to be a very strong claim.

We know the success of deep neural nets can be associated with its ability to learn latent features that are nonetheless present in the data. For a set of unknown features to be classified as unknown unknowns, they cannot be reasonably feature engineered. Short of a truly generative model that is able to make/generate meaning using language or constructs and labels we haven't used yet, how does your approach actually capture these unknowns? I realize that this may be splitting epistemological hairs, but any insight on how to identify what kind of additional data needs to be collected to address gaps in the model's intelligence would be massively helpful in the quest for models that can self-correct.

[Yiqing-Zh]

Thank you for your presentation in advance. I am wondering how to combine machine learning with traditional economic models and how ML modifies econometric techniques.

[ziwnchen]

Thanks for the presentation! As many other responses, I too feel like the idea of "Active learning", especially the knowledge quadrant really fascinating! Like its motivation has stated, I also believe active learning could be a potential approach in better identifying the boundary(and hence to some degree interoperability) of machine learning models. Could you comment on the current ability of active learning on finding meaningful items and exploring machine learning boundaries? What do you think would be the state of art/most exciting progress in the area of active learning?

[weijiexu-charlie]

Thanks for your presentation. Really looking forward to seeing you soon tomorrow!

[mingtao-gao]

Thank you in advance for your talk! Your research on Active Learning definitely brought many insights into how the sampling of raw data may affect significantly the machine learning models. The medium posts you made demonstrates how we as social science researchers can use the packages to achieve active learning. The visualizations tell us clearly how Active Learning and Active Transfer Learning work, but can you provide us with some examples and actual applications in machine learning models? That would be very helpful.

[Anqi-Zhou]

Thanks for sharing your inspiring work! Could you give more examples of active learning applying in social science research, especially in economics areas? I think your research on active learning is really meaningful!

[keertanavc]

Thanks for the talk. You've outlined some neat tricks to work better with data. Do you have any theoretical/empirical evidence for how much performance boost we can expect using these methods?

[Panyw97]

Thank you for your presentation! The topic is so interesting. I am wondering that what does the lack of information means when we are training a model to learn something from data? Does it mean we do not have so much data or the data provided is not sufficient for making a clear prediction? Thank you so much!

[bazirou]

Thank you for your presentation! The topic is so interesting. I am wondering that what does the lack of information means when we are training a model to learn something from data? Does it mean we do not have so much data or the data provided is not sufficient for making a clear prediction? Thank you so much!

[hesongrun]

Thank you so much for your presentation! The topic is really fascinating. We need to deploy our machine learning systems smartly to fully harness the advantage of 'human learning' versus 'machine learning'. However, from some applications, for example, the Alphago zero, which has no human input compared to the previous versions, turns out to outperform the previous versions that tap into human knowledge. I am wondering in the future, do we need better human assistance in the system or better and more intelligent algorithms that can go without human efforts? Which direction should we pursue? Thanks a lot!

[cytwill]

Thank you for the presentation! It is interesting to see another topic about analyzing human's role in machine-learning. As you mention, most machine learning methods require human input at the very beginning, this could a coin of two sides. I would like to know more advantages and disadvantages of human inputs, like the effects on the accuracy or bias. And still, how we would like to handle these unwanted effects in complex machine learning algorithms.

[AllisonXiong]

Thanks for bringing some industry insights to us! The idea of active learning and how to optimize the human-intervened part of ML algorithms is fascinating. My question is how do you select the sampling methods? Thanks!

[caibengbu]

Thank you for providing such an interesting paper to read! My question is that depending on the data and research domain, will the performance of active learning processes change?

[wu-yt]

Thank you for presenting such an interesting topic! What is your opinion on how active learning will change tech start-up business model?