Praise

Explore resources related to praise to help implement AI solutions for your business.

[D] Should the praise of "Data-centric" AI be taken with a grain of salt?

I hear co-workers, professors, entrepreneurs talk about just how important it is to switch from model-centric AI to a data-centric approach. Personally I don't see a problem focusing on attaining better data but I have been seeing this simplification that AI performance = model + data which seems massively misleading. A better generalization would be AI performance = model data or even AI performance = model log(len(data)). Looking at the rapidly changing problem of image synthesis, it is clear that data alone could not produce the same results seen even two years ago. On the other hand, it can seem unproductive for the field as a whole to try to create another classification model that achieves a .02% increased performance on ImageNet... again and again... And this is maybe why we hear more talk in industry about focusing on the data instead of models: Industry generally faces simpler tasks with negligible improvement as of recently (regression, classification). Therefore it's more easy and valuable to increase and improve the training data. And yes, most machine learning models NEED a large training set to approximate a complex distribution. We have already known this since forever. I think the term has just recently become a pet peeve for me as a novel "catch all" solution to any problem. What are your opinions on the matter?

#reddit #should #praise

[D] The machine learning community has a toxicity problem

yusuf-bengio•This week

[D] The machine learning community has a toxicity problem

It is omnipresent! First of all, the peer-review process is broken. Every fourth NeurIPS submission is put on arXiv. There are DeepMind researchers publicly going after reviewers who are criticizing their ICLR submission. On top of that, papers by well-known institutes that were put on arXiv are accepted at top conferences, despite the reviewers agreeing on rejection. In contrast, vice versa, some papers with a majority of accepts are overruled by the AC. (I don't want to call any names, just have a look the openreview page of this year's ICRL). Secondly, there is a reproducibility crisis. Tuning hyperparameters on the test set seem to be the standard practice nowadays. Papers that do not beat the current state-of-the-art method have a zero chance of getting accepted at a good conference. As a result, hyperparameters get tuned and subtle tricks implemented to observe a gain in performance where there isn't any. Thirdly, there is a worshiping problem. Every paper with a Stanford or DeepMind affiliation gets praised like a breakthrough. For instance, BERT has seven times more citations than ULMfit. The Google affiliation gives so much credibility and visibility to a paper. At every ICML conference, there is a crowd of people in front of every DeepMind poster, regardless of the content of the work. The same story happened with the Zoom meetings at the virtual ICLR 2020. Moreover, NeurIPS 2020 had twice as many submissions as ICML, even though both are top-tier ML conferences. Why? Why is the name "neural" praised so much? Next, Bengio, Hinton, and LeCun are truly deep learning pioneers but calling them the "godfathers" of AI is insane. It has reached the level of a cult. Fourthly, the way Yann LeCun talked about biases and fairness topics was insensitive. However, the toxicity and backlash that he received are beyond any reasonable quantity. Getting rid of LeCun and silencing people won't solve any issue. Fifthly, machine learning, and computer science in general, have a huge diversity problem. At our CS faculty, only 30% of undergrads and 15% of the professors are women. Going on parental leave during a PhD or post-doc usually means the end of an academic career. However, this lack of diversity is often abused as an excuse to shield certain people from any form of criticism. Reducing every negative comment in a scientific discussion to race and gender creates a toxic environment. People are becoming afraid to engage in fear of being called a racist or sexist, which in turn reinforces the diversity problem. Sixthly, moral and ethics are set arbitrarily. The U.S. domestic politics dominate every discussion. At this very moment, thousands of Uyghurs are put into concentration camps based on computer vision algorithms invented by this community, and nobody seems even remotely to care. Adding a "broader impact" section at the end of every people will not make this stop. There are huge shitstorms because a researcher wasn't mentioned in an article. Meanwhile, the 1-billion+ people continent of Africa is virtually excluded from any meaningful ML discussion (besides a few Indaba workshops). Seventhly, there is a cut-throat publish-or-perish mentality. If you don't publish 5+ NeurIPS/ICML papers per year, you are a looser. Research groups have become so large that the PI does not even know the name of every PhD student anymore. Certain people submit 50+ papers per year to NeurIPS. The sole purpose of writing a paper has become to having one more NeurIPS paper in your CV. Quality is secondary; passing the peer-preview stage has become the primary objective. Finally, discussions have become disrespectful. Schmidhuber calls Hinton a thief, Gebru calls LeCun a white supremacist, Anandkumar calls Marcus a sexist, everybody is under attack, but nothing is improved. Albert Einstein was opposing the theory of quantum mechanics. Can we please stop demonizing those who do not share our exact views. We are allowed to disagree without going for the jugular. The moment we start silencing people because of their opinion is the moment scientific and societal progress dies. Best intentions, Yusuf

#reddit #machine #learning

[D] Why is the AI Hype Absolutely Bonkers

good_rice•This week

[D] Why is the AI Hype Absolutely Bonkers

Edit 2: Both the repo and the post were deleted. Redacting identifying information as the author has appeared to make rectifications, and it’d be pretty damaging if this is what came up when googling their name / GitHub (hopefully they’ve learned a career lesson and can move on). TL;DR: A PhD candidate claimed to have achieved 97% accuracy for coronavirus from chest x-rays. Their post gathered thousands of reactions, and the candidate was quick to recruit branding, marketing, frontend, and backend developers for the project. Heaps of praise all around. He listed himself as a Director of XXXX (redacted), the new name for his project. The accuracy was based on a training dataset of ~30 images of lesion / healthy lungs, sharing of data between test / train / validation, and code to train ResNet50 from a PyTorch tutorial. Nonetheless, thousands of reactions and praise from the “AI | Data Science | Entrepreneur” community. Original Post: I saw this post circulating on LinkedIn: https://www.linkedin.com/posts/activity-6645711949554425856-9Dhm Here, a PhD candidate claims to achieve great performance with “ARTIFICIAL INTELLIGENCE” to predict coronavirus, asks for more help, and garners tens of thousands of views. The repo housing this ARTIFICIAL INTELLIGENCE solution already has a backend, front end, branding, a README translated in 6 languages, and a call to spread the word for this wonderful technology. Surely, I thought, this researcher has some great and novel tech for all of this hype? I mean dear god, we have branding, and the author has listed himself as the founder of an organization based on this project. Anything with this much attention, with dozens of “AI | Data Scientist | Entrepreneur” members of LinkedIn praising it, must have some great merit, right? Lo and behold, we have ResNet50, from torchvision.models import resnet50, with its linear layer replaced. We have a training dataset of 30 images. This should’ve taken at MAX 3 hours to put together - 1 hour for following a tutorial, and 2 for obfuscating the training with unnecessary code. I genuinely don’t know what to think other than this is bonkers. I hope I’m wrong, and there’s some secret model this author is hiding? If so, I’ll delete this post, but I looked through the repo and (REPO link redacted) that’s all I could find. I’m at a loss for thoughts. Can someone explain why this stuff trends on LinkedIn, gets thousands of views and reactions, and gets loads of praise from “expert data scientists”? It’s almost offensive to people who are like ... actually working to treat coronavirus and develop real solutions. It also seriously turns me off from pursuing an MS in CV as opposed to CS. Edit: It turns out there were duplicate images between test / val / training, as if ResNet50 on 30 images wasn’t enough already. He’s also posted an update signed as “Director of XXXX (redacted)”. This seems like a straight up sleazy way to capitalize on the pandemic by advertising himself to be the head of a made up organization, pulling resources away from real biomedical researchers.

#reddit #hype #absolutely