This short episode is a narration of Richard Sutton's The Bitter Lesson.

Richard Sutton is a distinguished research scientist at DeepMind and a professor of computing science at the University of Alberta. He is considered one of the founders of modern computational reinforcement learning, having several significant contributions to the field, including temporal difference learning and policy gradient methods.

If you haven't yet read his piece, it is definitely worth reading. This episode is a narration of it.

I would like to invite more discussion about these thoughts, so if you know of anybody who would like to discuss them, please send me to me at [email protected].

You can read the article here: http://www.incompleteideas.net/IncIdeas/BitterLesson.html

More drama in the contest between traditional machine learning models and deep learning models when it comes to tabular data.

We have on the show Vadim Borisov, a research fellow at the University of Tubingen as well as Kathrin Sessler, a PhD student from the same university.

This episode will be led by Ankush Garg, exploring Vadim and Kathrin's recent paper “Deep Neural Networks and Tabular Data: A Survey”.

Paper: https://arxiv.org/pdf/2110.01889.pdf
SAINT paper: https://arxiv.org/abs/2106.01342
Deep learning architectures and the hippocampus paper: https://arxiv.org/abs/2112.04035

We will also do a brief form of house keeping about the show and some of the upcoming episodes.

As always, if you have ideas, questions, feedback, or would like to be interviewed on the show: please reach out at [email protected].

Today we’re speaking with Jeff Clune about a new path towards general artificial intelligence, that he calls AI Generating Algorithms (AI-GAs).

Jeff is a Professor of Computer Science at the University of British Columbia. He was previously a Senior Research Scientist at Uber AI, and more recently a Research Team Leader at OpenAI. He's currently a Faculty Member at the Vector Institute.

We’re going to be talking about whether it is even possible that humans never create AGI, the enormous potential upside of AGI, the risks of AGI (including existential risks), Jeff’s fascinating transition into the field of AI, whether academia remains receptive to similar academic transitions, how long it might take to get to AGI and whether we have the right playbook for it, the prevailing status quo track (he mentions a couple): The first being a large language model approach, that he calls “standing on the shoulders of giant human datasets”, some of the limitations of this approach, the main track, that he calls “the manual path to AGI”, his new contribution, the AI Generating algorithms, which is essentially how we get out of the way and let the AI do much more, whether we already have an existence proof for how this can work, the three pillars of the AI Generating approach, AI Generating environments and capability oriented evaluation, the production of interesting or useful behaviors, the hidden Turing award to the person who would figure out AI Generating environment, learning pathologies, the circuitous paths of innovation, the seeming miracle of Darwinian evolution and the challenges of replicating it, the increasingly complex environments Darwinian evolution produces, the moral dimension to toying around with increasingly intelligent agents, AI and vegetarianism, and the extent to which we should be dismissive of the subjective claims of increasingly intelligent models.

His paper can be found here: https://arxiv.org/abs/1905.10985
Follow him on Twitter here: https://twitter.com/jeffclune
Learn more on his website here: http://jeffclune.com/

Today I'm speaking with Jan N. van Rijn about metalearning.

Jan is an assistant professor at Leiden University, where he also did his PhD. He is one of the founders of the OpenML Foundation, he previously did a post-doc at Freiburg in the Frank Hutter lab, and he is one of the authors of the metalearning book, which we'll be discussing.

We’ll be primarily examining the contents of their chapter titled “Evaluating Recommendations of Metalearning/AutoML Systems”.

We'll be covering such topics as OpenML and metalearning, how the space of AutoML has changed in the last few years,  benchmarks in the space of AutoML, how benchmarks measure progress of a field, some of the challenges benchmarks present, the need to build better tooling around benchmarks, the rise of NAS within the AutoML umbrella, the scope of AutoML (algorithm selection, hyperparameter optimization, the CASH problem, pipeline optimization, etc.),  how meta learning helps traverse through the various scopes of AutoML problem types, metalearning in the context of hyperparameter optimizations, the importance of properly designing meta-datasets, approaches to the inputs and outputs of meta-models and their advantages and disadvantages, surrogate models, how AutoML systems interact with meta-models, how to think about metalearning across dataset difficulties, diagnosing meta-models and meta-datasets, how to compare different metadata systems, loss-time curves, meta-features and the various approaches to creating them (including dataset 2 vec), contemporary meta-learning in a deep-learning context, and other topics.

Link to book - https://link.springer.com/book/10.1007/978-3-030-67024-5
Link to OpenML - https://www.openml.org/

Today we’re speaking with three researchers: Karan Grewal, Abhi Iyer and Akash Velu, about multi-task learning and how their new brain-inspired approach can help tackle it.

We’ll be discussing what a task is, what exactly we mean by multi-task systems, distances between tasks, the difference between continual learning and multi-task learning, catastrophic forgetting, catastrophic interference and their causes, various approaches out there like context-dependent gating and synaptic intelligence, the role of scale, and sparsity, we’ll cover some basics of the brain like dendrites, proximal and distal dendrites, apical and basal dendrites, how active dendrites can help us solve the challenges of gradient interference in multi-task learning, how they attach dendrites to each of their neurons in their deep learning models, their various approaches to representing context vectors, the challenges of brain-inspired approaches to machine learning, and some speculation about the future of this line of research.

Some references:
Their paper can be found here: https://arxiv.org/abs/2201.00042

All three of them had a great appearance on the Yannic Kilcher YouTube channel here: https://youtu.be/smxwT82o40Y

They mentioned Hierarchical Temporal Memory as the foundation stone for a lot of their research: https://numenta.com/blog/2019/10/24/machine-learning-guide-to-htm

Today I’m speaking with Mayukh Das about using neural architecture search for resource-constrained devices and about a new multi-objective reinforcement-learning based framework that he recently published called AUTOCOMET.

We’ll be covering such topics as how NAS research is done both at Samsung and at Microsoft,  the relationship between NAS and product teams, devices and the various types of constraints they expose, how to featurize hardware contexts, layer-wise latency calculations, surrogate models and the kinds of hardware-aware data they require, the current limitations of NAS, reinforcement learning and NAS, multi-objective optimization in the context of reinforcement learning, reward sparsity, reward shaping and shaping functions, primary and secondary rewards, their concept of co-regulated shaping, Q functions and the effects of potentials, AUTOCOMET, the future of NAS and other topics.

Thank you for tuning in!

To learn more about AUTOCOMET, find the paper here: https://arxiv.org/pdf/2203.15408.pdf

Today we’re speaking with José Hernández-Orallo. José is a Professor at the Polytechnic University of València in Spain and a Senior Research Fellow at the Leverhulme Centre for the Future of Intelligence, at Cambridge.

We'll be covering an enormous amount of ground surrounding intelligence and its evaluation.

We’ll touch on topics such as operating conditions in ML, agent characteristic curves,, the challenge with average performance scores, task-oriented evaluation, capability-oriented evaluation - in humans and other animals, animal-AI Olympics, meta-data annotation, performance robustness as a function of latent attributes, the limitations of aggregating behavioral performance into a handful of metrics, whether intelligence is meaningless or all-explanatory, the concept of generality in intelligence, AGI and the problem of distributions of tasks, calibration of task difficulties, whether capability metrics are incommensurable, complexity and compressibility, compression-based algorithmic information theory, the limits of intelligence, the intelligence of hybrid systems, whether we’re ready to talk about consciousness given our limited understanding of intelligence, whether rights should be tethered to pain or to cognition, animal rights, and what the AutoML community can do to help with this line of work.

If you're interested in this line of work, here is a recent paper by José and his team: https://ryanburnell.com/wp-content/uploads/Burnell-et-al-2022-Not-a-Number.pdf

His book, The Measure of All Minds, can be found here: https://www.amazon.com/Measure-All-Minds-Evaluating-Intelligence/dp/1107153018

Today I’m talking with Wout Schellaert about assessor models. Wout is a PhD student at the Polytechnic University of Valencia..

We’ll be covering a lot of different topics, such as the distributional hypothesis in machine learning, evaluation criteria, the reductive nature of current evaluation methods, task systems, the desiderata of assessor models, how to build assessor models, when to use them, what happens when our models become increasingly complex, how assessor models can help with AI explainability, whether they can help against adversarial scenarios, the challenges of scoring, bias in human evaluators, how assessor models relate to AI alignment, and other topics.

If you'd like to learn more about assessor models, visit the following pages.

José Hernández-Orallo (Wout's advisor): https://scholar.google.com/citations?user=n9AWbcAAAAAJ&hl=en

Training on the Test Set: Mapping the System-Problem Space in AI paper: https://www.aaai.org/AAAI22Papers/SMT-00432-Hernandez-OralloJ.pdf

Today we’re talking to Karthi Ramamurthy about a novel approach to similarity learning explainability.

Karthi is a research staff member in IBM Research at the Watson Research Center.

He studies the relationship between humans, machines, data and the societal implications of machine learning.

He was involved in the initial development of the open source AI Fairness 360 toolkit, where he’s still an active contributor.

His papers won various best paper awards like the 2015 IEEE International Conference on Data Science and Advanced Analytics.

He is an associate editor of Digital Signal Processing and a member of the IEEE and he holds a PhD in electrical engineering from Arizona State University.

We will be discussing a recent paper that he published about explainability methods for similarity learners. So we’ll go into detail about what similarity learning, or metric learning, is.

You can find the paper here: https://arxiv.org/pdf/2202.01153v1.pdf

To learn more about metric learning, please see this survey paper: https://people.bu.edu/bkulis/pubs/ftml_metric_learning.pdf.

Today I’m speaking with Vasco Lopes, about the state of Neural Architecture Search, NAS, and about a new method that he published that takes a very creative look at how to do NAS.

We’ll be discussing the motivation behind NAS, the current state of its deployment, the biggest use-cases today, the three components that make up NAS, the drawbacks to the current NAS paradigm, search spaces and how to design them, search strategies and how to choose them, graph representations of neural architectures, evaluation strategies and zero-cost approximations, bias in search space design, risks, the future of hand-designed architectures, and other topics.

Vasco is a PhD student at NOVA School of Science and Technology in Portugal and a co-founder of a Computer Vision startup called DeepNeuronic.

Find his paper, Towards Less Constrained Macro-Neural Architecture Search, here: https://arxiv.org/pdf/2203.05508.pdf