Interested in Bayesian optimization? Check out Google BayesOpt Speaker Series! We’ve had a fascinating series of talks given by speakers including
@arkrause
, Yoshua Bengio,
@peter_i_frazier
, Kirthevasan Kandasamy,
@PhilippHennig5
.. Stay tuned for more!
Still don’t believe reinforcement learning is an appropriate model for human-level intelligence. We are taught delayed gratitude all our lives while RL agents use discounted reward?! Like wut?! I’ll be so depressed if I’m an RL agent.
Alex Terenin's new Google TechTalk on Gaussian processes is a must-watch!
@avt_im
He offers some really cool perspectives on this statistical tool. I learned a lot from it, and I'm sure you will too!
Our team at
@GoogleDeepMind
is hiring a student researcher based in Cambridge MA in 2024 to work on uncertainty-aware decision making for LLMs and/or text-to-image models. Check out the full info and how to apply:
Very fancy lunch at MIT ML retreat. Grad students can’t believe they are having cup cakes with the cup made of actual chocolate. Formal sit down lunch rather than sandwich box. Mind blown.
Hyper Bayesian optimization (HyperBO) is a highly customizable interface that pre-trains a Gaussian process model and automatically defines model parameters, making Bayesian optimization easier to use while outperforming traditional methods. Learn more →
Using LLMs for a new domain-specific language?
We developed Grammar Prompting to teach them rules&structures, boosting their skills in semantic parsing, planning 🤖, and even molecule generation!
Excited to share this work
#NeurIPS2023
poster
#329
Tuesday at 5pm!
@Ziwphd
and I will present our Grammar Prompting work tomorrow at
#NeurIPS2023
Happy to chat more on
- LLMs for structured language generation (e.g., programs, molecules and robotic plans)
- Structured/formal chain-of-thought for reasoning
Ever struggled with transferring knowledge in Bayesian optimization due to varying dimensions? Our TMLR paper solves this with MPHD! New Gaussian process Model Pre-training method works seamlessly across Heterogeneous Domains.
#MachineLearning
Awesome talk by
@anmakaro
from ETH Zurich on Bayesian Optimization in the Wild: Risk-Averse Decisions and Budget Constraints, featuring best paper award at
#AutoML_Conf
. Check it out on Google BayesOpt @ YouTube!
#bayesopt
#automl
Do we really need a gigantic amount of data to train AI models? We've been exploring this question from several facets and here's our first step to ensure robustness during data downsampling!
How to downsample data to safeguard against worst-case generalization performance issues (stemming from class imbalance and distributional shifts)? Great work with
@bankes_william
, G. Hughes and
@ziwphd
Today on the blog learn about the OptFormer, one of the first Transformer-based frameworks for hyperparameter tuning, learned from large-scale optimization data using flexible text-based representations →
Thank you to those who expressed interest in the position. I've received hundreds of emails and unfortunately won't have the bandwidth to reply. Please make sure to apply through the official channel directly
Our team at
@GoogleDeepMind
is hiring a student researcher based in Cambridge MA in 2024 to work on uncertainty-aware decision making for LLMs and/or text-to-image models. Check out the full info and how to apply:
MPHD enhances Bayesian optimization with domain adaptability. Using a neural net pre-trained in existing domains, MPHD generates tailored Gaussian process models for new domains, leading to smarter optimization decisions.
Super cool talk by Frank Hutter at the latest Google BayesOpt Speaker Series: “Deep Learning 2.0: How Bayesian Optimization May Power the Next Generation of DL”. A new era of deep learning? Thank you
@FrankRHutter
for joining us!
@evensgn
@nahelocin
@Harvard
Our paper reveals a fundamental connection: multi-task learning and single-task learning are equivalent for stationary kernels. This theoretical insight means MPHD models are guaranteed to improve over time, ultimately reflecting the real-world process that produces the data.
@alex_y_ku
@jasonbaldridge
@cocosci_lab
@_beenkim
GPP judges uncertainty rationally, just like a rational human, e.g. no extreme judgements if no evidence, and getting close to a firm judgment only if episteme is high. We can even prove theoretically that episteme has no upper bound, i.e., there is no limit to knowledge.
With episteme, GPP knows when it does not know.
This means GPP not only probes, but also detects out-of-distribution data! On 3D Shapes and ImageNet OOD detection tasks, GPP performed consistently well and on par with the best benchmark method for each task.
@alex_y_ku
@jasonbaldridge
@cocosci_lab
@_beenkim
GPP probes model representation with epistemic and aleatory uncertainty evaluated as episteme (knowledge / confidence), alea (intrinsic randomness) and judged probability.
This helps GPP distinguish fuzzy concepts and unknown concepts, an ability linear probes do not possess.