Machine Learning Street Talk (MLST)

Bob Coercke is a celebrated physicist, he's been a Physics and Quantum professor at Oxford University for the last 20 years. He is particularly interested in Structure which is to say, Logic, Order, and Category Theory. He is well known for work involving compositional distributional models of natural language meaning and he is also fascinated with understanding how our brains work. Bob was recently appointed as the Chief Scientist at Cambridge Quantum Computing.

Bob thinks that interactions between systems in Quantum Mechanics carries naturally over to how word meanings interact in natural language. Bob argues that this interaction embodies the phenomenon of quantum teleportation.

Bob invented ZX-calculus, a graphical calculus for revealing the compositional structure inside quantum circuits - to show entanglement states and protocols in a visually succinct but logically complete way. Von Neumann himself didn't even like his own original symbolic formalism of quantum theory, despite it being widely used!

We hope you enjoy this fascinating conversation which might give you a lot of insight into natural language processing. 

Tim Intro [00:00:00]

The topological brain (Post-record button skit) [00:13:22]

Show kick off [00:19:31]

Bob introduction [00:22:37]

Changing culture in universities [00:24:51]

Machine Learning is like electricity [00:31:50]

NLP -- what is Bob's Quantum conception? [00:34:50]

The missing text problem [00:52:59]

Can statistical induction be trusted? [00:59:49]

On pragmatism and hybrid systems [01:04:42]

Parlour tricks, parsing and information flows [01:07:43]

How much human input is required with Bob's method? [01:11:29]

Reality, meaning, structure and language [01:14:42]

Replacing complexity with quantum entanglement, emergent complexity [01:17:45]

Loading quantum data requires machine learning [01:19:49] 

QC is happy math coincidence for NLP [01:22:30]

The Theory of English (ToE) [01:28:23] 

... or can we learn the ToE? [01:29:56] 

How did diagrammatic quantum calculus come about? [01:31:04

The state of quantum computing today [01:37:49] 

NLP on QC might be doable even in the NISQ era [01:40:48] 

Hype and private investment are driving progress [01:48:34] 

Crypto discussion (moved to post-show) [01:50:38] 

Kilcher is in a startup (moved to post show) [01:53:40

Debrief [01:55:26] 

Performing reliably on unseen or shifting data distributions is a difficult challenge for modern vision systems, even slight corruptions or transformations of images are enough to slash the accuracy of state-of-the-art classifiers. When an adversary is allowed to modify an input image directly, models can be manipulated into predicting anything even when there is no perceptible change, this is known an adversarial example. The ideal definition of an adversarial example is when humans consistently say two pictures are the same but a machine disagrees. Hadi Salman, a Ph.D student at MIT (ex-Uber and Microsoft Research) started thinking about how adversarial robustness  could be leveraged beyond security.

He realised that the phenomenon of adversarial examples could actually be turned upside down to lead to more robust models instead of breaking them. Hadi actually utilized the brittleness of neural networks to design unadversarial examples or robust objects which_ are objects designed specifically to be robustly recognized by neural networks. 

Introduction [00:00:00]

DR KILCHER'S PHD HAT [00:11:18]

Main Introduction [00:11:38]

Hadi's Introduction [00:14:43]

More robust models == transfer better [00:46:41]

Features not bugs paper [00:49:13]

Manifolds [00:55:51]

Robustness and Transferability [00:58:00]

Do non-robust features generalize worse than robust? [00:59:52]

The unreasonable predicament of entangled features [01:01:57]

We can only find adversarial examples in the vicinity [01:09:30]

Certifiability of models for robustness [01:13:55]

Carlini is coming for you! And we are screwed [01:23:21]

Distribution shift and corruptions are a bigger problem than adversarial examples [01:25:34]

All roads lead to generalization [01:26:47]

Unadversarial examples [01:27:26]

In today's show we are joined by Francois Chollet, I have been inspired by Francois ever since I read his Deep Learning with Python book and started using the Keras library which he invented many, many years ago. Francois has a clarity of thought that I've never seen in any other human being! He has extremely interesting views on intelligence as generalisation, abstraction and an information conversation ratio. He wrote on the measure of intelligence at the end of 2019 and it had a huge impact on my thinking. He thinks that NNs can only model continuous problems, which have a smooth learnable manifold and that many "type 2" problems which involve reasoning and/or planning are not suitable for NNs. He thinks that many problems have type 1 and type 2 enmeshed together. He thinks that the future of AI must include program synthesis to allow us to generalise broadly from a few examples, but the search could be guided by neural networks because the search space is interpolative to some extent.

https://youtu.be/J0p_thJJnoo

Tim's Whimsical notes; https://whimsical.com/chollet-show-QQ2atZUoRR3yFDsxKVzCbj

Dr. Christian Szegedy from Google Research is a deep learning heavyweight. He invented adversarial examples, one of the first object detection algorithms, the inceptionnet architecture, and co-invented batchnorm. He thinks that if you bet on computers and software in 1990 you would have been as right as if you bet on AI now. But he thinks that we have been programming computers the same way since the 1950s and there has been a huge stagnation ever since. Mathematics is the process of taking a fuzzy thought and formalising it. But could we automate that? Could we create a system which will act like a super human mathematician but you can talk to it in natural language? This is what Christian calls autoformalisation. Christian thinks that automating many of the things we do in mathematics is the first step towards software synthesis and building human-level AGI. Mathematics ability is the litmus test for general reasoning ability. Christian has a fascinating take on transformers too. 

With Yannic Lightspeed Kilcher and Dr. Mathew Salvaris

Whimsical Canvas with Tim's Notes:

https://whimsical.com/mar-26th-christian-szegedy-CpgGhnEYDBrDMFoATU6XYC

YouTube version (with detailed table of contents) https://youtu.be/ehNGGYFO6ms

The race is on, we are on a collective mission to understand and create artificial general intelligence. Dr. Tom Zahavy, a Research Scientist at DeepMind thinks that reinforcement learning is the most general learning framework that we have today, and in his opinion it could lead to artificial general intelligence. He thinks there are no tasks which could not be solved by simply maximising a reward. 

Back in 2012 when Tom was an undergraduate, before the deep learning revolution he attended an online lecture on how CNNs automatically discover representations. This was an epiphany for Tom. He decided in that very moment that he was going to become an ML researcher.  Tom's view is that the ability to recognise patterns and discover structure is the most important aspect of intelligence. This has been his quest ever since. He is particularly focused on using diversity preservation and metagradients to discover this structure. 

In this discussion we dive deep into meta gradients in reinforcement learning. 

Video version and TOC @  https://www.youtube.com/watch?v=hfaZwgk_iS0

First episode in a series we are doing on ML DevOps. Starting with the thing which nobody seems to be talking about enough, security! We chat with cyber security expert Andy Smith about threat modelling and trust boundaries for an ML DevOps system. 

Intro [00:00:00]

ML DevOps - a security perspective [00:00:50]

Threat Modelling [00:03:03]

Adversarial examples? [00:11:27]

Nobody understands the whole stack [00:13:53]

On the size of the state space, the element of unpredictability [00:18:32]

Threat modelling in more detail [00:21:17]

Trust boundaries for an ML DevOps system [00:25:45]

Andy has a YouTube channel on cyber security! Check it out @ 

https://www.youtube.com/channel/UCywP24ly6h6NTusX88TQKTQ

https://www.linkedin.com/in/andysmith-uk/

Video version:

https://youtu.be/7Tz-3S4lypI

Christoph Molnar is one of the main people to know in the space of interpretable ML. In 2018 he released the first version of his incredible online book, interpretable machine learning. Interpretability is often a deciding factor when a machine learning (ML) model is used in a product, a decision process, or in research. Interpretability methods can be used to discover knowledge, to debug or justify the model and its predictions, and to control and improve the model, reason about potential bias in models as well as increase the social acceptance of models. But Interpretability methods can also be quite esoteric, add an additional layer of complexity and potential pitfalls and requires expert knowledge to understand. Is it even possible to understand complex models or even humans for that matter in any  meaningful way? 

Introduction to IML [00:00:00]

Show Kickoff [00:13:28]

What makes a good explanation? [00:15:51]

Quantification of how good an explanation is [00:19:59]

Knowledge of the pitfalls of IML [00:22:14]

Are linear models even interpretable? [00:24:26]

Complex Math models to explain Complex Math models? [00:27:04]

Saliency maps are glorified edge detectors [00:28:35]

Challenge on IML -- feature dependence [00:36:46]

Don't leap to using a complex model! Surrogate models can be too dumb [00:40:52]

On airplane pilots. Seeking to understand vs testing [00:44:09]

IML Could help us make better models or lead a better life [00:51:53]

Lack of statistical rigor and quantification of uncertainty [00:55:35]

On Causality [01:01:09]

Broadening out the discussion to the process or institutional level [01:08:53]

No focus on fairness / ethics? [01:11:44]

Is it possible to condition ML model training on IML metrics ? [01:15:27]

Where is IML going? Some of the esoterica of the IML methods [01:18:35]

You can't compress information without common knowledge, the latter becomes the bottleneck [01:23:25]

IML methods used non-interactively? Making IML an engineering discipline [01:31:10]

Tim Postscript -- on the lack of effective corporate operating models for IML, security, engineering and ethics [01:36:34]

Explanation in Artificial Intelligence: Insights from the Social Sciences (Tim Miller 2018)

https://arxiv.org/pdf/1706.07269.pdf

Seven Myths in Machine Learning Research (Chang 19) 

Myth 7: Saliency maps are robust ways to interpret neural networks

https://arxiv.org/pdf/1902.06789.pdf

Sanity Checks for Saliency Maps (Adebayo 2020)

https://arxiv.org/pdf/1810.03292.pdf

Interpretable Machine Learning: A Guide for Making Black Box Models Explainable.

https://christophm.github.io/interpretable-ml-book/

Christoph Molnar:

https://www.linkedin.com/in/christoph-molnar-63777189/

https://machine-master.blogspot.com/

https://twitter.com/ChristophMolnar

Please show your appreciation and buy Christoph's book here;

https://www.lulu.com/shop/christoph-molnar/interpretable-machine-learning/paperback/product-24449081.html?page=1&pageSize=4

Panel: 

Connor Tann https://www.linkedin.com/in/connor-tann-a92906a1/

Dr. Tim Scarfe 

Dr. Keith Duggar

Video version:

https://youtu.be/0LIACHcxpHU

Academics think of themselves as trailblazers, explorers — seekers of the truth.

Any fundamental discovery involves a significant degree of risk. If an idea is guaranteed to work then it moves from the realm of research to engineering. Unfortunately, this also means that most research careers will invariably be failures at least if failures are measured via “objective” metrics like citations. Today we discuss the recent article from Mark Saroufim called Machine Learning: the great stagnation. We discuss the rise of gentleman scientists, fake rigor, incentives in ML, SOTA-chasing, "graduate student descent", distribution of talent in ML and how to learn effectively.  

With special guest interviewer Mat Salvaris. 

Machine learning: the great stagnation [00:00:00]

Main show kick off [00:16:30]

Great stagnation article / Bad incentive systems in academia [00:18:24]

OpenAI is a media business [00:19:48]

Incentive structures in academia [00:22:13]

SOTA chasing [00:24:47]

F You Money [00:28:53]

Research grants and gentlemen scientists [00:29:13]

Following your own gradient of interest and making a contribution [00:33:27]

Marketing yourself to be successful [00:37:07]

Tech companies create the bad incentives [00:42:20]

GPT3 was sota chasing but it seemed really... "good"? Scaling laws? [00:51:09]

Dota / game AI [00:58:39]

Hard to go it alone? [01:02:08]

Reaching out to people [01:09:21]

Willingness to be wrong [01:13:14]

Distribution of talent / tech interviews [01:18:30]

What should you read online and how to learn? Sharing your stuff online and finding your niece [01:25:52]

Mark Saroufim:

https://marksaroufim.substack.com/

http://robotoverlordmanual.com/

https://twitter.com/marksaroufim

https://www.youtube.com/marksaroufim

Dr. Mathew Salvaris:

https://www.linkedin.com/in/drmathewsalvaris/

https://twitter.com/MSalvaris

Microsoft has an interesting strategy with their new “autonomous systems” technology also known as Project Bonsai. They want to create an interface to abstract away the complexity and esoterica of deep reinforcement learning. They want to fuse together expert knowledge and artificial intelligence all on one platform, so that complex problems can be decomposed into simpler ones. They want to take machine learning Ph.Ds out of the equation and make autonomous systems engineering look more like a traditional software engineering process. It is an ambitious undertaking, but interesting. Reinforcement learning is extremely difficult (as I cover in the video), and if you don’t have a team of RL Ph.Ds with tech industry experience, you shouldn’t even consider doing it yourself. This is our take on it!

There are 3 chapters in this video;

Chapter 1: Tim's intro and take on RL being hard, intro to Bonsai and machine teaching 

Chapter 2: Interview with Scott Stanfield [recorded Jan 2020] 00:56:41

Chapter 3: Traditional street talk episode [recorded Dec 2020] 01:38:13

This is *not* an official communication from Microsoft, all personal opinions. There is no MS-confidential information in this video. 

With:

Scott Stanfield

https://twitter.com/seesharp

Megan Bloemsma

https://twitter.com/BloemsmaMegan

Gurdeep Pall (he has not validated anything we have said in this video or been involved in the creation of it)

https://www.linkedin.com/in/gurdeep-pall-0aa639bb/

Panel: 

Dr. Keith Duggar

Dr. Tim Scarfe

Yannic Kilcher

Today we are going to talk about the *Data-efficient image Transformers paper or (DeiT) which Hugo is the primary author of. One of the recipes of success for vision models since the DL revolution began has been the availability of large training sets. CNNs have been optimized for almost a decade now, including through extensive architecture search which is prone to overfitting. Motivated by the success of transformers-based models

in Natural Language Processing there has been increasing attention in applying these approaches to vision models. Hugo and his collaborators used a different training strategy and a new distillation token to get a massive increase in sample efficiency with image transformers. 

00:00:00 Introduction

00:06:33 Data augmentation is all you need

00:09:53 Now the image patches are the convolutions though?

00:12:16 Where are those inductive biases hiding?

00:15:46 Distillation token

00:21:01 Why different resolutions on training

00:24:14 How data efficient can we get?

00:26:47 Out of domain generalisation

00:28:22 Why are transformers data efficient at all? Learning invariances

00:32:04 Is data augmentation cheating?

00:33:25 Distillation strategies - matching the intermediatae teacher representation as well as output

00:35:49 Do ML models learn the same thing for a problem?

00:39:01 How is it like at Facebook AI?

00:41:17 How long is the PhD programme?

00:42:03 Other interests outside of transformers?

00:43:18 Transformers for Vision and Language

00:47:40 Could we improve transformers models? (Hybrid models)

00:49:03 Biggest challenges in AI?

00:50:52 How far can we go with data driven approach?