Author Archives: pgfeldman

Phil 10.19.2021

Book

  • Do another review, and do something in the introduction that works with the idea that we are barely individuals. How that changes from childhood through adulthood to old age, and how technology has had a huge impact

SBIRs

  • 9:15 Standup
  • 3:00 Army
  • Timesheets!

GPT-Agents

  • Run the new LIWC data and generate two spreadsheets. One with the word numbers and one with the default settings. Done

Phil 10.17.2021

Neural circuit policies enabling auditable autonomy

  • A central goal of artificial intelligence in high-stakes decision-making applications is to design a single algorithm that simultaneously expresses generalizability by learning coherent representations of their world and interpretable explanations of its dynamics. Here, we combine brain-inspired neural computation principles and scalable deep learning architectures to design compact neural controllers for task-specific compartments of a full-stack autonomous vehicle control system. We discover that a single algorithm with 19 control neurons, connecting 32 encapsulated input features to outputs by 253 synapses, learns to map high-dimensional inputs into steering commands. This system shows superior generalizability, interpretability and robust-ness compared with orders-of-magnitude larger black-box learning systems. The obtained neural agents enable high-fidelity autonomy for task-specific parts of a complex autonomous system.

Phil 10.15.2010

Apparently I have to roll over my 401k? We no longer have a retirement plan?

Baselines for Uncertainty and Robustness in Deep Learning

  • In “Uncertainty Baselines: Benchmarks for Uncertainty & Robustness in Deep Learning”, we introduce Uncertainty Baselines, a collection of high-quality implementations of standard and state-of-the-art deep learning methods for a variety of tasks, with the goal of making research on uncertainty and robustness more reproducible. The collection spans 19 methods across nine tasks, each with at least five metrics. Each baseline is a self-contained experiment pipeline with easily reusable and extendable components and with minimal dependencies outside of the framework in which it is written. The included pipelines are implemented in TensorFlowPyTorch, and Jax. Additionally, the hyperparameters for each baseline have been extensively tuned over numerous iterations so as to provide even stronger results.

Book

  • Twitter and Tear Gas

SBIRs

  • Adding optional buttons to TopicCombo so it’s possible to add a topic and not set a seed.
  • Need to check for the case where I am adding a topic to the group that provided the seed. No need to link to yourself
  • Save graph to DB, hopefully
  • Woohoo!
  • Taking a break
  • Saving the full graph to the DB!
  • How it looks today

Phil 10.14.2021

Book

GPT Agents

  • Generated the spreadsheet for the baseline gpt model
  • Got the template set up

SBIRs

  • Change status on stories – done
  • 9:15 standup
  • 10:00 RFP discussion
  • 11:00 LAIC meeting
  • 2:00 Weekly adversarial RL – interesting chat. Create a story to look at replacing the Markov process with a small NN
  • Finish graph save, start on load? Nope, other tweaking instead.
The tool so far

Phil 10.13.2021

Book

SBIRs

  • Create stories for 1) DB buildout 2) Model save/load 3) GML generation
  • 9:30 sprint planning
  • Creating tables and getting Graph read/write to the DB

GPT Agents

  • Rebuilt the code that takes into account the LIWC2015 components and how they relate/rollup. And I found one real difference between the ground truth and the gpt:
LIWC 2015 “Informal” word counts out of 500k words
  • 4:15 Meeting. Going to make a spreadsheet of the untrained GPT yelp, and be done with data. When that gets back, re-run the spreadsheets, and also add a version of the z-test code that produces rollups with the original LIWC data.

JuryRoom

  • 6:00 Meeting

Phil 10.12.2021

Triggering Failures: Out-Of-Distribution detection by learning from local adversarial attacks in Semantic Segmentation

  • In this paper, we tackle the detection of out-of-distribution (OOD) objects in semantic segmentation. By analyzing the literature, we found that current methods are either accurate or fast but not both which limits their usability in real world applications. To get the best of both aspects, we propose to mitigate the common shortcomings by following four design principles: decoupling the OOD detection from the segmentation task, observing the entire segmentation network instead of just its output, generating training data for the OOD detector by leveraging blind spots in the segmentation network and focusing the generated data on localized regions in the image to simulate OOD objects. Our main contribution is a new OOD detection architecture called ObsNet associated with a dedicated training scheme based on Local Adversarial Attacks (LAA). We validate the soundness of our approach across numerous ablation studies. We also show it obtains top performances both in speed and accuracy when compared to ten recent methods of the literature on three different datasets.

SBIRs

  • 9:00 Sprint demos – done
  • More DB – working!
  • More GPT – got the seed response running. Now working on the topic response – done!
  • Added switching between raw content to support topics and details
  • Send John a gml file with more data – done
  • Stories for next sprint

Phil 10.11.2021

Call Jim Donnie – done

Call Outlaw – done. Working out a tire mound and a compressor to be added to the build

Book

  • Rewriting the positioning statement and probably a lot of the proposal to reframe the books as a problem/solution as opposed to “look at these interesting things”

SBIRs

  • Enable calls to GPT
  • Set up for database, add row_id and parent_id to objects and then wrtite to_db and from_db methods (base class?)
  • Set up basic password management
  • Got db access working, and am writing/updating project tables. Each project has its own password so they can be shared
  • Meeting with John on the UI. I’m thinking that the demo can run entirely off of gml files that are generated by Graphbuilder

Phil 10.8.2021

Looks like tomorrow will be damp. Wool!

Seagull prep

  • Pack wool gear
  • Pack riding food
  • Get burgers and rolls
  • Get beer
  • Load GPS
  • Load bike

Book

  • Write 2 short reviews. Did some more topical overview and wrote a paragraph on Meltdown

SBIRs

  • Verify that subsequent parent-node linking works – done. That took a while. There was more stuff to fix
  • Make sure that we don’t make redundant links, just update weights – done. No weight updating for now. I think I’d rather calculate them on the fly for now
  • Make sure that details get stored with topics – done. Not sure that overwriting the response is a good idea. I think a better idea is to store the current response and replace it in the raw text once the details are set.
  • Get to_string() to show when node is selected – done
  • Enable calls to GPT
  • Monday is database day?

GPT Agents

  • 3:00 Meeting

4:00 Interview

Phil 10.7.2021

Call outlaw about spare tire mount and compressor

Slowed canonical progress in large fields of science

  • The size of scientific fields may impede the rise of new ideas. Examining 1.8 billion citations among 90 million papers across 241 subjects, we find a deluge of papers does not lead to turnover of central ideas in a field, but rather to ossification of canon. Scholars in fields where many papers are published annually face difficulty getting published, read, and cited unless their work references already widely cited articles. New papers containing potentially important contributions cannot garner field-wide attention through gradual processes of diffusion. These findings suggest fundamental progress may be stymied if quantitative growth of scientific endeavors—in number of scientists, institutes, and papers—is not balanced by structures fostering disruptive scholarship and focusing attention on novel ideas.

SBIRs

  • 9:15 Standup
  • 11:00 LAIC Tagup. Send Aaron a screenshot for the paper – done and integrated into report
  • Adding ForceNodes to MapGroups -done!
Nodes!
  • Adding connections. Done! That took a while. some of my states did not make sense, so the parent node would change. Need to test with successive iterations, but I’m done for the day
Connections!

Phil 10.6.2021

Book

SBIRs

  • More bug hunting. Trying different types of callbacks. Nope, that didn’t work. In a way, I’m kind of relieved. Moving on to getting nodes to appear with links
  • Got seed and topic MapGroups adding members correctly, now adding connections between MapGroups. Done.
  • I think things are mostly working. Going to try to hook up ForceNodes tomorrow
So far, so good!

Phil 10.5.2021

Had a wild dream last night. It began with Prince and I talking motorcycles, though I have to say that Prince looked more like Little Richard. There was a nice red motorcycle too, with spinning parts on the motor that don’t really make much sense now when I think about it.

At some point later, I was walking down a rural road and into a church(?). There was some kind christening-type thing going on, but it involved surgically implanting a line between the pelvis and the shoulder so that the child would grow up visibly stunted. This was presented as a price paid by the child for the parents belonging to the wrong group. That way, the group could always be recognized, and the individuals would be in a constant level of pain. My sense was that the original intent was that this was originally done so that the child would grow to reject the group so that its children wouldn’t have the procedure. But instead it had become a ritualized tradition and insisted on by the parents. Maybe a bit like foot binding?

Book

GPT-Agents

  • Need to see if I can do Z-tests in Excel – you can (tutorial)!
  • Had to add the American data
  • Finished correlation, t-test, and z-test

SBIRs

  • still working on the logic for setting parent group and selected node/topic. There is some kind of problem with callbacks on the tk.Text widget. I can fix all my bugs by turning off all the Text callbacks. It may be because I create a callback for an in-class method that in turn calls an optional method that is passed into the class. Tomorrow I’ll try binding directly and see if that fixes things, otherwise proceed with only the select callback in the raw text

Phil 10.4.2021

Wheel!

Book

GPT-Agents

  • Start LIWC csv reader – got the reader and counts done. Need to see if I can do the rese in Excel
  • Ping Andreea – done!

SBIRs

  • Expense report – done again, fingers crossed!
  • current parent/child node logic
  • Got most of the logic together for setting parent group and selected node. I’m not really happy about this, there are too many states and hidden relationships. This will need a cleanup once it is working. I may be able to do a good deal with looking at the node info though

Phil 10.2.2021

Rubrix is a production-ready Python framework for exploring, annotating, and managing data in NLP projects.

Key features:

  • Open: Rubrix is free, open-source, and 100% compatible with major NLP libraries (Hugging Face transformers, spaCy, Stanford Stanza, Flair, etc.). In fact, you can use and combine your preferred libraries without implementing any specific interface.
  • End-to-end: Most annotation tools treat data collection as a one-off activity at the beginning of each project. In real-world projects, data collection is a key activity of the iterative process of ML model development. Once a model goes into production, you want to monitor and analyze its predictions, and collect more data to improve your model over time. Rubrix is designed to close this gap, enabling you to iterate as much as you need.
  • User and Developer Experience: The key to sustainable NLP solutions is to make it easier for everyone to contribute to projects. Domain experts should feel comfortable interpreting and annotating data. Data scientists should feel free to experiment and iterate. Engineers should feel in control of data pipelines. Rubrix optimizes the experience for these core users to make your teams more productive.
  • Beyond hand-labeling: Classical hand labeling workflows are costly and inefficient, but having humans-in-the-loop is essential. Easily combine hand-labeling with active learning, bulk-labeling, zero-shot models, and weak-supervision in novel data annotation workflows.