Ziyang Li
@_ziyang_
Followers
114
Following
32
Media
4
Statuses
14
No Code No Life | UPenn CIS PhD Student
Joined April 2022
Joint work with my awesome colleagues Jiani Huang (@moqingyan233) and Neelay (@NeelayV), professors Mayur Naik (@AI4Code) and Rajeev Alur (@RajeevAlur), and amazing undergrads Jason, Liam, Felix, and Eric from @CIS_Penn. (9/9).
0
0
3
To learn more:.- Read our AAAI’24 paper: - Try Vieira online: - Join our Discord Channel: - Check out Vieira’s source code: (8/9).
github.com
Framework and Language for Neurosymbolic Programming. - scallop-lang/scallop
1
2
11
For instance, Vieira’s solution for the Date Reasoning task from BIG-Bench achieves an accuracy of 92.41% compared to GPT-4’s 71.00%, and GPT-4 (CoT)’s 87.26%. (7/9)
1
0
4
We evaluate Vieira on 9 challenging tasks that span language, vision, and structured and vector databases. Our evaluation shows that Vieira programs are concise, can incorporate modern foundation models, and achieve comparable or better accuracy than competitive baselines. (6/9)
1
0
3
Vieira introduces a uniform relational programming interface for invoking diverse foundation models. With an extensible plugin library, programmers have access to 12 pre-built foundation models (GPT, Llama, Clip, Segment Anything, etc.) and can even build their own plugins. (5/9).
1
0
3
Why relational?.1) vast amount of data exists in relational databases;.2) relations can also represent scene graphs in vision and syntax trees in natural and formal languages; and.3) relational programs can easily be augmented with probabilistic or differentiable semantics. (4/9).
1
0
3
Vieira builds atop Scallop, our neuro-symbolic language and compiler, to integrate probabilistic relational programming with foundation models. Our key insight is that foundation models are stateless functions with relational inputs and outputs. Cf. (3/9).
One thing I really like about language models is that they are stateless (they are functional programs of type text -> text). This allows us to share prompts (essentially currying the LM) and reproduce results.
1
0
4
The powerful but incomplete nature of foundation models has spurred many efforts to augment them with capabilities such as calling functions, searching the web, and using other tools. Examples include ChatGPT Plugins, LMQL, and VisProg. (2/9)
1
0
4
I am delighted to introduce Vieira, a declarative language and framework for neuro-symbolic programming with foundation models. (1/9)
2
22
62
RT @AI4Code: I am excited to share a preview of Scallop: a new programming language and toolchain for neurosymbolic AI. Website: https://t….
0
104
0