final set of proofreading changes #494
Conversation
sections/06_discussion.md
Outdated
| over plausible values of an input patch to more accurately estimate its | ||
| contribution. | ||
| different input patches. More recently, marginalizing over the plausible values | ||
| of an input has been suggest as a way to more accurately estimate contributions |
sections/06_discussion.md
Outdated
| the predicted probability of a selected class. When tested on image data, their | ||
| method took about 300 iterations to converge, compared to the ~5000 iterations | ||
| used by LIME. One drawback of this approach is that the use of gradient descent | ||
| the predicted probability of a selected class. Their method converged in many |
| perturbation-based approaches is to propagate an important signal from a target | ||
| output neuron backwards through the layers to the input layer in a single | ||
| backpropagation-like pass. A classic example of this is calculating the gradients | ||
| Backpropagation-based methods, in which the signal from a target output neuron is propagated backwards to the input layer, are another way to interpret deep networks that sidestep inefficiencies of the perturbastion-basd methods. |
There was a problem hiding this comment.
It's hard for me to follow the details of the diff here. Did you keep the same references but shorten the narrative, or were references dropped? I'm in favor of shortening this section.
There was a problem hiding this comment.
I think the references (or nearly all of them) were preserved but the section was dramatically shortened.
sections/06_discussion.md
Outdated
| weaknesses [@tag:Mahendran2016_salient], and new methods are being developed | ||
| to address them [@tag:Selvaraju2016_grad @tag:Sundararajan2017_axiomatic @tag:Shrikumar2017_learning]. | ||
| Lundberg and Lee [@tag:Lundberg2016_an] noted that several importance | ||
| scoring methods, integrated gradients and LIME, could all be |
sections/06_discussion.md
Outdated
| @tag:Li2014_minibatch]. However, GPUs also have a limited quantity of memory, | ||
| making it difficult to implement networks of useful size and complexity on a | ||
| @tag:Li2014_minibatch]. However, GPUs also have limited memory, | ||
| making networks of useful size and complexity it difficult to implement on a |
sections/06_discussion.md
Outdated
| application-specific integrated circuits (ASICs) [@arxiv:1704.04760]. Specialized hardware promises | ||
| application-specific integrated circuits (ASICs) [@arxiv:1704.04760]. | ||
| Less software | ||
| available for such highly specialized hardware [@tag:Lacey2016_dl_fpga]. |
sections/06_discussion.md
Outdated
| learning researchers to problems in genomics and healthcare. We have even | ||
| quickly on a CPU, are important for training students and attracting machine | ||
| learning researchers to problems in genomics and healthcare. | ||
| `TODO: Cite syllabus or this last sentence should probably go. Unclear what it |
There was a problem hiding this comment.
Please cut the last sentence about DragoNN in the course
| dropping from more than 20% to less than 6% [@tag:Speech_recognition] and | ||
| finally approaching or exceeding human performance in the past year | ||
| [@arxiv:1610.05256 @arxiv:1703.02136]. The phenomenal improvements on benchmark | ||
| datasets are undeniable, but greatly reducing the error rate on these benchmarks did not | ||
| fundamentally transform the domain. Widespread adoption of conversational | ||
| speech technologies will require not only improvements over baseline methods but | ||
| truly solving the problem, in this case exceeding human-level performance, as |
There was a problem hiding this comment.
This phrasing was awkward, but I wrote it with such complexity to emphasize "solving the problem". In speech recognition, human performance may be the goal line for solving the problem. For many tasks we review, human performance isn't relevant though. Is there a way to convey that more concisely?
There was a problem hiding this comment.
made a little change that goes more to what you had
sections/07_conclusions.md
Outdated
| performance than either individually [@arxiv:1606.05718]. Especially for sample | ||
| and patient classification tasks, we expect deep learning methods to complement | ||
| and assist biomedical researchers rather than compete with or even replace them. | ||
| semantics of the objects presented. Work in this area is continuing |
There was a problem hiding this comment.
I think the phrase "work in this area" should make it clear that the ongoing work is to guard against attacks and adversarial examples. Maybe the second half of the sentence is clear enough though?
|
|
||
| Even if deep learning in biology and healthcare is not yet transformative today, | ||
| we are extremely optimistic about its future. Given how rapidly deep learning is | ||
| We are extremely optimistic about the future of deep learning in biology and medicine. Given how rapidly deep learning is |
There was a problem hiding this comment.
I'll save it for a later issue or pull request, but some of my extreme enthusiasm has dampened a little after re-reading and reflecting upon the entire review. Let's merge this as-is and discuss separately.
|
plane should board soon - if this looks good feel free to merge or i can merge later |
|
Actually - noticed your point. Will merge & create issue. |
No description provided.