Artificial Intelligence @Google. Former Stanford PhD. Cultural, interdisciplinary enthusiast. I tweet about things that interest me.
San Francisco, CA
Joined April 2009
- Tweets767
- Following493
- Followers705
- Likes1.1K
Sercan Arık retweeted
Multimodal LLMs (MLLMs) excel at many tasks, but object hallucination (generated descriptions of non-existent objects) hinders widespread use. Learn about a method to reduce hallucinations in existing MLLMs and maintain their vision-language capabilities → goo.gle/4fBQSN3
Selective prediction can improve reliability of LLMs by allowing them to abstain from making predictions when they are unsure. Our novel framework, ASPIRE, is based on adaptation with self evaluation to push state-of-the-art in selective prediction. For details:
Introducing ASPIRE, a framework that enhances the selective prediction capabilities of large language models, enabling them to output an answer paired with a confidence score. Learn how it outperforms state-of-the-art methods on a variety of QA datasets. → goo.gle/3RWpChi
Sercan Arık retweeted
Google's new ASPIRE system teaches AI the value of saying 'I don't know' venturebeat.com/ai/googles-n…
Prompting can be very important to get accurate outputs with LLMs but challenging for humans - handcrafting even a small number of demos can be difficult or, for unseen tasks, impossible. Our recent work on automatic prompting addresses this...
Introducing a new approach for adaptive prompting of #LLMs that train with unlabeled samples + pseudo-demonstrations generated by the model itself to close the gap between few-shot and 0-shot performance on reasoning, NLU and language generation tasks. → goo.gle/3tY00IB
Sercan Arık retweeted
Although we have witnessed several breakthroughs by transformers (especially LLMs), simple MLPs (or mixers in a fancier name 🙂) can be competitive or even better on time-series forecasting problems 🚀 Feel free to give a try on your task!
Code: github.com/google-research/g…
Introducing TSMixer, an advanced multivariate model for long-term forecasting that leverages linear model characteristics for high benchmark performance. Learn how it outperforms state-of-the-art models when evaluated on real-world applications → goo.gle/3PulkOj
ALT Transformer block and TSMixer block architectures. TSMixer replaces the multi-head attention layer with time-mixing, a linear model applied on the time dimension.
Our recent work on using LLMs to query databases using natural language:
State-of-the-art Text-to-SQL with SQL-PaLM
- proposes SQL-PaLM, an LLM-based Text-to-SQL adopted from PaLM-2
- achieves SoTA in both in-context learning and fine-tuning settings
- incredible that the few-shot model outperforms the previous fine-tuned SoTA by 3.8% on the Spider benchmark
- few-shot SQL-PaLM also outperforms few-shot GPT-4 by 9.9%, using a simple prompting approach
arxiv.org/abs/2306.00739
Sercan Arık retweeted
At #GoogleIO, we unveiled four important ways we’re supporting customers and partners to leverage generative AI capabilities. cloud.google.com/blog/produc…
Privacy concerns can arise as a key bottleneck for data sharing, especially for sensitive domains like healthcare.
We propose a novel generative modeling framework, that yields privacy-preserving synthetic EHR data with high fidelity.
Analyzing electronic health records (EHRs) has great potential, e.g. to enhance patient care, but common anonymization methods can decrease the data’s utility. To that end, read how EHR-Safe generates high-fidelity & privacy-preserving synthetic EHR data→ goo.gle/3HZfpxn
ALT Block diagram of EHR-Safe framework.
We’re excited to announce that TabNet is now available in Vertex AI Tabular Workflows: lnkd.in/g7jB5EMe! Tabular Workflows provides fully managed, optimized, and scalable pipelines, making it easier to use TabNet without worrying about implementation details. (1/3)
TabNet on Vertex AI Tabular Workflows is optimized for efficient scaling to massive (billion-scale) tabular datasets. Moreover, TabNet on Vertex AI Tabular Workflows come with ML improvements on top of the original TabNet, yielding better accuracy for real-world challenges. (2/3)
Testimonial by Uber on adopting TabNet for one of their core business use cases: lnkd.in/gzx6bZwd (3/3)
Sercan Arık retweeted
Our paper (arxiv.org/abs/2209.07999) just got accepted to @NeurIPSConf 2022! Congratulations to @Alper_T_E for coming up with the idea and @ozsoyserdar, @ShadiSHamdan, @sercanarik for a great collaboration.
Sercan Arık retweeted
Self supervised learning is revolutionizing AI using large unlabeled datasets. We show that maximizing mutual information between alternative representations of the same input is a practical method for self supervised learning that is immune to the dreaded collapse problem.
"Algorithmic fairness in pandemic forecasting: lessons from COVID-19" - we present our perspectives for equitable ML innovations based on our experience building pandemic forecasting models, with brilliant collaborators from Harvard led by @Thomasctsai
nature.com/articles/s41746-0…
Sercan Arık retweeted
Grateful for the opportunity to have collaborated with @GoogleAI @tomaspfister @sercanarik and others on the Google-Harvard Covid-19 Public Forecast model that was part of this broader ensemble modeling effort.
Excited to this @PNASNews paper out.
pnas.org/doi/10.1073/pnas.21…
Our recent work introduces a novel way of designing attention based visual understanding architectures, demonstrating high accuracy, useful interpretability capabilities and learning benefits. See our paper for more details: arxiv.org/pdf/2105.12723.pdf
The Visual Transformer has helped advance many core computer vision applications, e.g., image classification, but training can be inefficient and models lack interpretable designs. Learn how the Nested Hierarchical Transformer addresses these challenges → goo.gle/3sAy7Ca
Our new blog post on integrating rules into deep learning towards interpretable, robust and reliable deep neural networks…
Introducing a novel approach for interpretable, robust, and reliable deep neural networks (DNNs) that employs controllable rule representations, which do not require retraining to adjust the rule strength at inference. Learn more below ↓ goo.gle/3IN8TGU
Our recent work: Temporal Fusion Transformer (TFT), for interpretable time series forecasting. 📈 TFT has been used to help retail and logistics companies for accurate and interpretable demand forecasting, and for applications related to climate change.
Announcing the Temporal Fusion Transformer, designed specifically to handle the heterogeneity of data in multi-horizon forecasting, which achieves more accurate forecasts with increased interpretability. Read more, including real-world applications ↓ goo.gle/3oPgE84
Sercan Arık retweeted
Beyond covid-19, how could mRNA-based vaccines deal with diseases such as malaria, TB and HIV? The founders of @BioNTech_Group explain econ.st/3oer1kw
Our latest paper on AI-augmented epidemiology from @GoogleCloud published in Nature Digital Medicine & Google AI Blog! Our models have been used in US & Japan for creating COVID-19 testing targets, allocating resources and simulating policies.
Learn more about a new ML-based framework for epidemiology that we applied to COVID-19, including forecasts that are released to the public daily. Read all about how it was developed and has been used by large organizations ↓ goo.gle/3Azo2Hj