Showing posts from November, 2022

Locating and Editing Factual Associations in GPT

This is an overview of the paper 'ROME: Locating and Editing Factual Associations in GPT'.  It includes an interview with 2 of the papers authors as well as Yannic Kilcher's commentary explanations of the paper. The paper is an analysis of where the information in a large language deep learning transformer based model like GPT is stored, and whether one could reprogram said models to transform or modify specific facts learned by the model.  The results are kind of fascinating, and give new insight into the role of the MLP (multi-layer perception) part of the Transformer model. So maybe 'attention' is not all you need after all? Here's a link to the paper. Here's a link to a followup paper on Mass-Editing in a Transformer.  

Templates in C++ - back to basics

  This is a talk by Nicolai Josuttis at CppCon 2022 titled 'Back to Basics: Templates in C++'.  It is a good intro review of C++ Templates that also includes some new info on new features in the latest C++ version.   'Auto' is an example of one of the new features that is pretty cool.

Computational Photography, AI, and 3D

  This is a presentation by Fredo Durand of MIT CSAIL on some recent research work in his group, focusing on computational photography, AI, and 3D. We have HTC posts on Halide here . The paper 'Differentiable Monte Carlo Ray Tracing through Edge Sampling' is available here . The paper 'Light Field Networks, Neural Scene Representations with Single-Evaluation Rendering' is available here . The project page is here . The paper 'Learning to generate line drawings that convey geometry and semantics' can be found here . The project page is here .

Michael Levin on the Foundations of Cognition

  This is a podcast interview with Michael Levin, who is a developmental and synthetic biologist at Tufts University.  It focuses on the foundations of cognition. Michael's website with more info on his fascinating research can be found here .

Spatial Thinking and the Foundation of Thought

The video presentation is a talk by Barbara Tversky for SEMF's 2022 Spacious Spatiality event.  The talk abstract is as follows: All creatures must move and act in space to survive. The elementary act is to approach or to avoid, an inherently emotional act. In mammals, place cells in hippocampus cull multi-sensory information to represent places; they are spatially arrayed one synapse away in grid cells in entorhinal cortex. In people, place cells also represent events, people, and ideas and grid cells array them in temporal, social, and conceptual spaces. The spatial foundation of thought is evident in expression of thought, in gesture, diagrams, and language, with implications for design, creativity, and cooperation. There is a more detailed HTC post on Grid Cells and Cortical Maps that includes a lecture by Novel Prize winner May-Britt Moser that details the hippocampus cell and entorhinal cortex research that Barbara mentions in the talk. Jeff Hawkins of Numenta mentions this

Python and C++ Interoperability with Shiboken - Qt for Python project

  This is a Qt for Python project webinar presentation on 'Python and C++ Interoperability with Shiboken' presented in 2020.  Shiboken is a binding generator tool for binding C++ code for Python.  The Qt for Python project uses it to bind Qt (via pyside2) to Python, but you aren't restricted to using Qt and can use it for your own C++ code libraries that you want to access in Python code. You can find more information on the Qt for Python project here . The Qt company is working on building better development tools for working with Python and Qt, and this is an ongoing development associated with Qt 6.  You can learn more about the packaging and tooling here . Here's a blog post on the new Qt for Python features in Qt 6.4.  This includes pyside6-deploy for making a deployable executable, and improvements to Shiboken. This is a December 2021 Qt World Summit presentation on Hybrid Qt Development with Python.