Better hyperparameters๐= bigger bang for the buck๐ธ! What if the next trillion-parameter model could be tuned by running a tiny one w millions of params instead? Our technique, ฮผTransfer, enables that by aligning the optimal HPs across model sizes. arxiv.org/abs/2203.03466
GPT-3 175B is powerful but too expensive to serve many finetuned copies. We use low-rank adaptation (LoRA) to learn task modules that are 10,000x smaller and can be swapped while the main model is frozen. No extra inference latency or quality drop! Paper: arxiv.org/abs/2106.09685
here is o1, a series of our most capable and aligned models yet:
openai.com/index/learningโฆ
o1 is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it.
ฮผTransfer allowed us to find better hyperparameters for GPT-3 on a single GPU. It has since been used by Cerebras and Google to train huge models.
But how does it actually work? What is the enabling insight?
Get your answers in 7 minutes๐
Sampling from energy functions is fundamental to machine learning, but mostly done by expensive MCMC. GFlowNet is a deep learning way to amortize that cost and can produce diverse samples for RL, NLP, etc. Our latest work builds its theoretical foundation arxiv.org/abs/2111.09266
In June we released LoRA which adapts NNs as big as GPT-3 with few parameters yet stays performant. Our new result beats finetuned RoBERTa on GLUE with 1/8 of total parameters! Try "pip install loralib" and add LoRA to your fav model in 10 lines of code! github.com/microsoft/LoRA
๐คจShould you care about GFlowNets?
What are they anyway?๐ง
Learn about how GFlowNets speed up drug discovery and help large language models reason better in my new video!๐ฌ๐
Low-rank adaptation (LoRA) is one of the most popular methods for customizing large AI models.๐ค
๐คWhat's the story behind its invention?
๐กHow did we come up with the idea?
๐งShould you use it?
Get your answers in my new video๐๐ฝ๏ธ
๐คShould you few-shot prompt or fine-tune an LLM when you have limited training data?
๐กHere is a better option!
๐Our paper uses amortized inference to extract hard-to-access knowledge from LLMs and boost data efficiency for reasoning.
More in ๐งต๐
arxiv.org/abs/2310.04363
How can we keep scaling if compute grows exponentially but data does not?๐ค
Find the answer in my guest lecture at Stanford CS25 this afternoon on our ICLR2024 Honorable Mention paper "Amortizing intractable inference in large language models" ๐
1/ Existing theories of neural networks (NN) like NTK don't learn features so can't explain success of pretraining (e.g. BERT, GPT3). We derive the *feature learning* โ-width limit of NNs & pretrained such an โ-width word2vec model: it learned semantics!
arxiv.org/abs/2011.14522
Can AI explain data with complex latent structures, like graphs? Classic EM algo fits only simple latent variable models like Gaussian mixture & HMM.
Our GFlowNet-EM uses a big NN to do the hard work & explains data with complex compositional latents.
๐arxiv.org/abs/2302.06576
