Haiku Tech Center

 This is a conversation with Jeffery Shainline, who is a physicist at NIST with a specialty in optical computing and optoelectronics on different approaches to developing neuromorphic computing hardware for building artificial intelligence.  The wide range of topics discussed include the physics and practice of existing computer processor manufacturing, super conductivity for computation, optical computing, loop neurons, machine learning, evolution, and how the brain works. Lex has been hitting it out of the park with some of his recent interviews, and this is another great one.  Anyone with an interest in computer hardware and/or machine learning and artificial intelligence should find this discussion fascinating and informative.

This is a fascinating conversation between Jay McClelland and Lex Fridman on the history of the PDP group and connectionist models of cognitive neuroscience.   Jay was one of the authors of the famous PDP (Parallel Distributed Processing) book from the 1980s (that came with associated code so you could run the models themselves) that really helped jump-start the connectionist revolution in cognitive neuroscience as well as introduced the back propagation algorithm for training neural nets to the computer science community.

 The Qt 6.2 update release is almost upon us. Qt6.2 is the first long term stability release for Qt6.  It is also the first one to fully support M1 ARM chip native compiles on the mac side. Lar's direct involvement in the QuMultimedia framework is also very encouraging. This KDAB news update includes info on upcoming QT 6.2, a short interview with Maurice Kalinowski of the Qt Company, and info on C++20. It looks like Sept 30th is the current release date projection for Qt 6.2.

 Today's presentation is a talk by Edgar Riba on the Kornia open source library.  Kornia is a differentiable computer vision library.  We have covered it before here in previous HTC posts, but this recent presentation from 2 weeks ago will help you get up to date with the latest and greatest developments.   This is from the OpenCV weekly webinar series.  They also allude in the webinar to the new Oak-lite kickstarter campaign, which is active now. Observations: 1:  One thing Edgar points out is that the Kornia library works on tensors, so unlike OpenCV you don't have to worry about how many channels are needed for an input to a particular algorithm.  The library just deals with it. 2:  In theory one could use the TorchScript part of PyTorch to get C++ code output from a Kornia algorithm. Edgar mentioned 80% compatibility with Torchscript right now.  I do get nervous when he says 'in theory' for this or Onyx output.

This is a conversation between Lex Fridman and Wojciech Zaremba, who is a co-founder of OpenAI and the tech leader of the OpenAi Codex project. They talk about a wide variety of topics, but the section in the middle that focuses on the OpenAi Codex project is particually interesting. Observations: 1:  It did not occur to me that the OpenAi Codex software could be configured to teach itself.  Without necessarily requiring humans to build the data sets.  Which leads to some interesting conclusions.  It will be fun to see how this project develops over the next year of 2. We covered OpenAi Codex last week with The Tale of Two Demos post.

 The crew at Machine Learning Street Talk really did a super excellent of putting this presentation together. It is well worth watching the entire thing.  Take breaks as necessary. "Symmetry, as wide or narrow as you may define its meaning, is one idea by which man through the ages has tried to comprehend and create order, beauty, and perfection." and that was a quote from Hermann Weyl, a German mathematician who was born in the late 19th century.  The last decade has witnessed an experimental revolution in data science and machine learning, epitomised by deep learning methods. Many high-dimensional learning tasks previously thought to be beyond reach -- such as computer vision, playing Go, or protein folding -- are in fact tractable given enough computational horsepower. Remarkably, the essence of deep learning is built from two simple algorithmic principles: first, the notion of representation or feature learning and second, learning by local gradient-descent type methods,

 Today's presentation continues our sweep through some of the Full Stack Deep Learning lectures from their spring 2021 course.  This one is an overview and more detailed breakdown of the various components of machine learning tools and infrastructure.  So software, compute hardware options, resource management, frameworks, distributed training, experiment management, hyper-parameter tuning. I had to puzzle a lot of this stuff out with a ton of web searching and puzzlement, so seeing it all explained in one place is a great resource.  The full stack course really tries to cover the nuts and bolts of everything you might need to get up to speed on for working with deep learning projects. Observations 1:  They discuss Visual Studio Code early on.  If you are looking for a good Python coding environment i think it is a good choice. 2:  Streamlit was new to me, but sounds interesting.  Especially the part about easy ways to build interactive applets from your Python code. 3:  PyTorch L

 The Full Stack Deep Learning course is great, because it really dives into the nuts and bolts of the details associated with the actual practice of deep learning.  This lecture on data management infrastructure is a good example.   This lecture is worth the price of admission (which is basically just the time required to watch it) just for the section 41 minutes in that contrasts a Hadoop Cluster vs just using unix command line tools to implement a particular part of a deep learning data management stack.  Once again proving the time honored KISS (keep i simple stupid) principal that seems to be an alternate acronym for Occam's Razor   (the simplest solution is probably the best, and certainly the most desirable unless you are billing by the hour, then by all means use that complex solution ). Observations 1:  Being much more of a signal processing and algorithms guy, i was not really familiar with a lot of what was covered in this lecture. So i learned a lot of useful information

 Todays presentation is a lecture by Turing award winner Yoshua Bengio presented at the Heielberg Laureate Forum in 2019.  It provides a higher level overview of the current state of the art and future directions than the kind of info you get in something like yesterdays nuts and bolts full stack lecture. This lecture will look back at some of the principles behind the recent successes of deep learning as well as acknowledge current limitations, and finally propose research directions to build on top of this progress and towards human-level AI.  Notions of distributed representations, the curse of dimensionality, and compositionality with neural networks will be discussed, along with the fairly recent advances changing neural networks from pattern recognition devices to systems that can process any data structure thanks to attention mechanisms, and that can imagine novel but plausible configurations of random variables through deep generative networks.  At the same time, analyzing the

 Todays presentation is the Research Directions lecture in the Full Stack Deep Learning course (spring 2021) , and is presented by Pieter Abbeel.  The lecture has a lot of focus on robotics, probably because of Pieter's background and research interests. Hot topics such as unsupervised learning, reinforcement learning, meta learning, one-shot learning, data augmentation techniques, and new uses in science and engineering are covered.  Other hot topics such as generative models, relationships to biological models, and reference frames (things like capsule networks) are not covered, so like i said the topics covered tend to match the lecture presenters interests . We will be following up today's presentation with another one tomorrow that takes a higher level approach to laying out current and potential future research directions by Turing award winner Yoshua Bengio. Observations 1:  The domain randomization section is particularly interesting.  Once again pointing out the high e

 This is a followup to yesterdays Full Stack Transformer lecture post that is a tutorial on Transfer Learning and the HuggingFace Transformers Library by Thomas Wolf of HuggingFace presented in spring 2020.  He starts by covering recent breakthroughs in NLP that resulted from combining transfer learning and various transformer architectures.  Then he dives into some specific information on open source tools created by HuggingFace that implement these ideas in various open source apis. You can learn more about HuggingFace and their Transformers Library here .

 Todays  presentation is the transformer lecture from the Berkeley full stack deep learning course (which is quite excellent).  They title it 'Transfer Learning and Transformers', but the transfer learning part is really just a few minutes at the very beginning that is really just a lead in to their view of the events leading up to transformer architectures. So we quickly mention transfer learning, then move on to vector embedding models of language, then on to the real meat of the lecture, the transformer architecture.  We recently covered the OpenAI Codex automated coding system, and they talk about how that fell out of this transformer research on natural language modeling.  They also briefly talk about ethic and ai systems like GPT-3 or Codex. This lecture is from the spring 2021 Full Stack Deep Learning online course. HTC has a number of blog posts on transformers you can check out here , including a full tutorial on how to code one in PyTorch. We also have covered transfe

This lecture by Leslie Valiant on 'What are the Computational Challenges for the Cortex' is a followup lecture to yesterdays talk by Edvard Moser.  It focuses on the theoretical limitations of cortex models based on known physics and biological constraints in neural assemblies. Over a lifetime the brain performs hundreds of thousands of individual cognitive acts of a variety of kinds, including the formation of new associations and other kinds of learning. Each such act depends on past experience, and, in turn, can have long lasting effects on future behavior. It is difficult to reconcile such large scale capabilities quantitatively with the known resource constraints on cortex, such as low connectivity. Here we shall describe an approach to this problem, in terms of concrete functions, representations, and algorithms, that seeks to explain these phenomena in terms that are faithful to the basic quantitative resources available. Until recently an algorithmic understanding of co

We're continuing our presentation of background material to help understand Jeff Hawkins thousand brains theory described in a HTC post earlier this week.   Today's lecture was given by Nobel Laureate Edvard Moser on 'Space and time: Internal dynamics of the brain's entorhinal cortex' in 2019. Observations 1:  Scale space, lateral inhibition, orientation maps.  Fascinating how the mechanisms described in this talk mirror mechanisms in the brain's visual cortex processing. 2:  The optical recording of brain activity via fluorescence from viral genetic manipulation of the brain cells is pretty slick.

 Today's lecture is by Nobel Prize winner May-Britt Moser on structures in the brain that enable spatial orientation and navigation.  She covers the history of the discovery of place and grid cells in the mammalian brain, as well a many details of her own work in this area.  The specific research she covers was focused on the hippocampus, but the previous HTC posts covers Jeff Hawkins and Numenta's theoretical work on showing how these same grid cell structures might be a significant part of how the neo-cortex functions.

 Jeff Hawkins of Numenta has a new book out that tries to explain the results of Numenta's research into modeling how the cortex of the brain functions, and how intelligent behavior build off of that underlying structure. The thousand brains theory of intelligence puts forward a novel theory of how the neocortex works.  It proposes that each cortical micro-column builds a model of an object, and they then collectively bubble out a consensus about what they are modeling actually means.  Other people have discussed micro-columns being a replicated neural computation element before, so other than de-emphasing hierarchy, i'm not sure there is really anything new here. The real meat of the theory is the part that delves into grid cells in the brain.  Specifically grid cells in the neo-cortex.  And the notion that all computation in the cortex is ultimately composed of replicated grid cell structure that is used for all modality of sensory input, and for the mechanics of thought its

Our summer break here at HTC so we could focus on releasing new software to the greater universe is almost over.  It was the best of times.  It was the worst of times.  It was and continues to be in fact a wild ride time for AI, since new exciting things are coming out so quickly it is difficult to keep track. Time to dive back into recent developments in deep learning, AI, generative models, computer graphics, etc.  And boy there have been a lot of those new developments over the last 2 months. OpenAI just released something really interesting. Now how useful the current incarnation really is, well that is worth understanding by watching these 2 somewhat different demos. Let's begin with the slick and enthusiastic openAi demo of their new OpenAI Codex automated coding system. Pretty slick, eh. But what about that other demo i promised.  The boys at Machine Learning Street Talk have a slightly different take on it.  Here is their live coding session. So take what you will from thes