Latest news, recently, the ICLR conference, known as the "King of Deep Learning Conferences," was exposed that nearly half of the reviewers for ICLR 2020 have not published papers in the field.
This news immediately caused a stir in the industry. According to reliable sources, the review results of ICLR 2020 were finally released yesterday. However, the announcement was delayed by one day due to the extremely high number of submissions this year, with a total of 2600 papers. The previous year's conference, ICLR 2019, had 1500 submissions, while ICLR 2018 had just over 900, nearly doubling in two years. Therefore, some people believe that the explosive growth in the number of paper submissions is the fundamental cause of this absurd event.
ICLR Conference
ICLR, the International Conference on Learning Representations, was established in 2013 and is widely recognized by academic researchers as a top conference in the field of deep learning.
ICLR's paper review process mainly includes single-blind, double-blind, and open reviews. Single-blind review means that the reviewers know the authors' names, schools, and other identity information, but the authors do not know who reviewed their papers. Double-blind review means that both parties do not know each other's identity information. These two methods are the main approaches for reviewing papers in many academic conferences and journals.
Open review is different. According to the rules, all submitted papers will have their names and other information made public, and they will be subject to evaluation and questioning by all peers. Any scholar can evaluate the papers anonymously or with their real names. After the open review process, the authors can also make adjustments and modifications to their papers. ICLR is not the first conference to use this review mechanism, but it has done an excellent job in terms of openness and influence, receiving high praise from industry professionals.
OpenReview.net, an open review system founded by Andrew McCAllum of the University of Massachusetts Amherst, is the official submission portal for ICLR. It fully preserves the content of all papers and review discussions from previous conferences. At the same time, it has attracted attention from the academic community and gained support from industry giants such as Facebook, Google, and NSF.
This questioning of the reviewers is undoubtedly a significant blow to ICLR, a top conference. However, this is not the first time ICLR has faced review issues.
Double-blind Review Incident
It all started with a paper that received scores of "4, 4, 3".
A paper that should have been rejected outright was unexpectedly accepted by the AC, despite the strong objections from the three reviewers. This became the highlight of the year.
At first, an onlooker, A, said: Since the AC has the power to accept papers, what is the point of having three reviewers? Just for show?
Onlooker B said: My own paper with scores of "7, 7, 5" was rejected...
Onlooker C said: I also found a paper with scores of "4, 5, 6" that was accepted.
A paper with scores of "4, 4, 3" was unexpectedly accepted by a top conference, causing a significant uproar. By the way, the controversial paper is titled "Encoding Tree Structures into Word Embeddings using Geometric Methods".
In this paper, the author proposes a new method to implicitly encode the hierarchical category information of tree structures into word embeddings, resulting in a so-called "hyper-spherical representation". The method has two criteria: 1) the hierarchical relationships between categories should be implicitly represented as inclusion relationships between corresponding n-balls; 2) the tree structure of category relationships should be accurately encoded. In other words, the energy loss should be zero.
The author of this paper believes that the research has three contributions: 1) proposing a new geometric method for constructing/training n-balls to encode hierarchical category relationships as inclusion relationships between n-balls with zero energy loss; 2) considering both the position and size of n-balls, which is more accurate than cosine similarity in experiments; 3) creating a new n-ball embedding benchmark dataset based on Glove, extracting the tree structure from WordNet 3.0, which is freely accessible.
The paper is well-structured and has no issues, but it was rejected by the three reviewers. Reviewer A said: It's okay, but not good enough, reject. Reviewer B said: It's okay, but not good enough, reject. Reviewer C said: Absolutely reject!
Normally, this paper would have been rejected, but the result is often unexpected. The AC said: The author has proposed an interesting method to inject hierarchical information into existing word vectors. This may be helpful for various tasks that require knowledge base information and co-occurrence counts of text. Although the reviewers pointed out some shortcomings of this paper, I believe that the problem may be due to the failure to connect symbolic information/sets/logic/KB with neural networks. Therefore, I recommend accepting this paper at the conference.
The review and acceptance of this paper sparked heated discussions on Reddit. It is not for non-professionals to judge who is right or wrong, but it is undoubtedly problematic. However, compared to this incident, the double-blind review incident was just an appetizer, a mere sprinkle of water.
Having "non-professional" reviewers evaluate professional papers is a disaster no matter how you look at it. It is no different from asking a physical education teacher to grade a calculus exam!
Letting outsiders judge the professional abilities of professionals must be the idea of a "genius". What will happen to ICLR 2020? Let us wait and see.
Reference sources: Google News, Lei Feng Network, AI Technology Headquarters, Baidu Zhidao.
Alphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Anthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Apple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Google launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
YouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Google grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Mistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.