MIT research scientist working on AI, computer vision, and neuroscience.
Cambridge, MA
Joined March 2013
- Tweets24
- Following190
- Followers250
- Likes34
Want embeddings for your neural recordings? You'll have better decoders with less data and a much more interesting story around connectivity. Check out @GeelingC and @czlwang's awesome work at #ICLR2025. From BrainBERT to the Population Transformer arxiv.org/abs/2406.03044
🎉Excited to share: My first ML conference paper, Population Transformer 🧠, is an Oral at #ICLR2025! This work has truly evolved since its first appearance as a workshop paper last year. So thankful to have worked with the best advisors + collaborators! 🤗 More soon!
Interested in large-scale neuroscience of language and multimodal representations? We have the dataset for you, the Brain Treebank! Come to our oral at NeurIPS. Today at 3:50pm PST, East Meeting room braintreebank.dev/ Now with extra foundation models arxiv.org/abs/2302.14367
What if we solved every object detection benchmark? Would machines be human-level? At @NeurIPSConf we measure how hard object detection datasets are for humans for the first time. Over 90% of benchmark images are easy! news.mit.edu/2023/image-reco… 1/3
Our new metric, Minimum Viewing Time, MVT, can be used to objectively judge dataset difficulty, to guide new dataset development, and even to speed up object detectors by using simple models for simple images. 2/3
With @IBM and @USAF_AI, next time someone tries to sell their object detector, measure how hard the dataset they're making claims on is. We've also learned there's something special about language-supervised vision models, check out the paper! objectnet.dev/mvt/ 3/3
Is this 1821 or 2021? @FinanceMuseum's "Genealogy of American Finance" doesn't mention slavery and redlining anywhere in 324 pages. They do take time to praise Wells Fargo for lending to "ethnic minorities". A bank that had to apologize for using slaves as collateral for loans...
@FinanceMuseum Your website also says that the North started the civil war!? "the North ... force a confrontation with the South not over slavery ... over ... the truncation of national markets" moaf.org/exhibits/checks_bal… What kind of museum are you?
@smithsonian How is the @FinanceMuseum associated with the Smithsonian given its views on race and slavery? Does the Smithsonian support the view that the North stated the civil war as a form of economic aggression against the south?
You would this is from the Pocket Guide to Misogyny, but it's a popular book on negotiation. Reading the literature on negotiation strategies to teach machines how to bargain has been a shocking. How can business schools continue to teach this drivel in the 21st century?
Andrei Barbu retweeted
Our paper just won a Best Paper Award at @svrhm2020! Congrats to my wonderful collaborators 🎉 And thanks the organizers for the great workshop!
Andrei Barbu retweeted
@avivnet , @tianminshu , Boris Katz, @_abarbu_ & @MITCoCoSci (Josh Tenenbaum) will be giving an Invited Talk / Oral Presentation on "PHASE: Physically-grounded Abstract Social Events for Machine Social Perception" at 1.15-1.30 pm EST, this Saturday! Link: openreview.net/forum?id=_bok…
Not only do word embeddings have biases, our work with @candacerossio shows adding vision to ground embeddings makes biases worse using new data & tests. Vision datasets are very biased arxiv.org/abs/2002.08911 Doesn't matter, it was in the 40% of #IJCAI rejections without review!
Andrei Barbu retweeted
What happens when folks at @MIT_CSAIL get an adorable dog in their office? Now presenting the weekend project of @_abarbu_ #doggytraining #arduino @AdamYaari @yenling_kuo
Our dataset, ObjectNet, a new kind of object recognition dataset that includes controls for biases, is being released at #NeurIPS this week. Removing biases leads to a sharp drop in performance, opening the way to more robust computer vision. objectnet.dev
Our new #emnlp paper is out on grounded language learning with @candacerossio and @yevgeni_berzak. Learn semantic parsers the way kids do: by looking around and figuring out what what interpretations make sense. news.mit.edu/2018/machines-l… aclweb.org/anthology/D18-128…
Security questions are great! My password didn't work, answered the question about where I was born. Logged on. Not my account. Ah well ;) Time to remember my username.
Andrei Barbu retweeted
Valuable paper on anchoring bias in treebanking—needs analysis of errors/variation & how to improve accuracy arxiv.org/abs/1605.04481 #nlproc